From 6eddde06a4f25d55d538b5d15628dcc2b6882147 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Mon, 13 Jul 2026 18:37:57 +0200 Subject: [PATCH 01/24] CUDA: refactor MMQ kernel configuration (#24127) * CUDA: refactor MMQ kernel configuration * fix Blackwell config * remove legacy code --- ggml/src/ggml-cuda/mmq-config-ampere.cuh | 366 ++ ggml/src/ggml-cuda/mmq-config-blackwell.cuh | 37 + ggml/src/ggml-cuda/mmq-config-cdna.cuh | 177 + ggml/src/ggml-cuda/mmq-config-pascal.cuh | 261 ++ ggml/src/ggml-cuda/mmq-config-rdna2.cuh | 261 ++ ggml/src/ggml-cuda/mmq-config-rdna4.cuh | 282 ++ ggml/src/ggml-cuda/mmq-load-tiles.cuh | 1671 +++++++ ggml/src/ggml-cuda/mmq-vec-dot.cuh | 1251 ++++++ ggml/src/ggml-cuda/mmq.cu | 55 +- ggml/src/ggml-cuda/mmq.cuh | 4300 ++++--------------- 10 files changed, 5121 insertions(+), 3540 deletions(-) create mode 100644 ggml/src/ggml-cuda/mmq-config-ampere.cuh create mode 100644 ggml/src/ggml-cuda/mmq-config-blackwell.cuh create mode 100644 ggml/src/ggml-cuda/mmq-config-cdna.cuh create mode 100644 ggml/src/ggml-cuda/mmq-config-pascal.cuh create mode 100644 ggml/src/ggml-cuda/mmq-config-rdna2.cuh create mode 100644 ggml/src/ggml-cuda/mmq-config-rdna4.cuh create mode 100644 ggml/src/ggml-cuda/mmq-load-tiles.cuh create mode 100644 ggml/src/ggml-cuda/mmq-vec-dot.cuh diff --git a/ggml/src/ggml-cuda/mmq-config-ampere.cuh b/ggml/src/ggml-cuda/mmq-config-ampere.cuh new file mode 100644 index 000000000..0037bac3d --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-ampere.cuh @@ -0,0 +1,366 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_ampere(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + + return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true); +} diff --git a/ggml/src/ggml-cuda/mmq-config-blackwell.cuh b/ggml/src/ggml-cuda/mmq-config-blackwell.cuh new file mode 100644 index 000000000..9fbe32b69 --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-blackwell.cuh @@ -0,0 +1,37 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_blackwell(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + + return ggml_cuda_mmq_get_config_ampere(type, J, fallback); +} diff --git a/ggml/src/ggml-cuda/mmq-config-cdna.cuh b/ggml/src/ggml-cuda/mmq-config-cdna.cuh new file mode 100644 index 000000000..46ec6aa9d --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-cdna.cuh @@ -0,0 +1,177 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_cdna(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + + return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true); +} diff --git a/ggml/src/ggml-cuda/mmq-config-pascal.cuh b/ggml/src/ggml-cuda/mmq-config-pascal.cuh new file mode 100644 index 000000000..8f0faac88 --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-pascal.cuh @@ -0,0 +1,261 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_pascal(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + + return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true); +} diff --git a/ggml/src/ggml-cuda/mmq-config-rdna2.cuh b/ggml/src/ggml-cuda/mmq-config-rdna2.cuh new file mode 100644 index 000000000..de4db0a3d --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-rdna2.cuh @@ -0,0 +1,261 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna2(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + + return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true); +} diff --git a/ggml/src/ggml-cuda/mmq-config-rdna4.cuh b/ggml/src/ggml-cuda/mmq-config-rdna4.cuh new file mode 100644 index 000000000..6280e80ee --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-rdna4.cuh @@ -0,0 +1,282 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna4(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + + return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true); +} diff --git a/ggml/src/ggml-cuda/mmq-load-tiles.cuh b/ggml/src/ggml-cuda/mmq-load-tiles.cuh new file mode 100644 index 000000000..3978b1baa --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-load-tiles.cuh @@ -0,0 +1,1671 @@ +#pragma once + +#include "vecdotq.cuh" + +#include "mmq.cuh" + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q1_0( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int blocks_per_iter = MMQ_ITER_K / QK1_0; + constexpr int threads_per_row = blocks_per_iter * QI1_0; + constexpr int nrows = warp_size / threads_per_row; + constexpr int scale_entries_per_block = QK1_0 / QK8_1; + constexpr int scale_entries_per_row = blocks_per_iter * scale_entries_per_block; + + const int txi = threadIdx.x % threads_per_row; + const int kbx = txi / QI1_0; + const int kqsx = txi % QI1_0; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + kbx; + const int qs_offset = 4*kqsx; + const int qs0 = bxi->qs[qs_offset + 0] | (bxi->qs[qs_offset + 1] << 8) | + (bxi->qs[qs_offset + 2] << 16) | (bxi->qs[qs_offset + 3] << 24); + + int unpacked_bytes[8]; +#pragma unroll + for (int j = 0; j < 8; ++j) { + const int shift = j * 4; + const int bits4 = (qs0 >> shift) & 0x0F; + const int b0 = (bits4 & 0x01) ? 1 : -1; + const int b1 = (bits4 & 0x02) ? 1 : -1; + const int b2 = (bits4 & 0x04) ? 1 : -1; + const int b3 = (bits4 & 0x08) ? 1 : -1; + unpacked_bytes[j] = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24); + } + + const int dst_offset = kbx*(scale_entries_per_block*QI8_0) + kqsx*QI8_0; +#pragma unroll + for (int j = 0; j < 8; ++j) { +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + dst_offset + j] = unpacked_bytes[j]; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j] = unpacked_bytes[j]; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + } + + const int ksx = threadIdx.x % scale_entries_per_row; + const int scale_block = ksx / scale_entries_per_block; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps) { + int i = i0 + threadIdx.y; + + if (fallback) { + i = min(i, i_max); + } + + const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + scale_block; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + ksx] = bxi->d; +#else + x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + ksx] = bxi->d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_0( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_0); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI4_0; + const int kqsx = txi % QI4_0; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx; + const int qs0 = get_int_b2(bxi->qs, kqsx); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kbx*(2*QI4_0) + kqsx + 0] = __vsubss4((qs0 >> 0) & 0x0F0F0F0F, 0x08080808); + x_qs[i*sram_stride + kbx*(2*QI4_0) + kqsx + QI4_0] = __vsubss4((qs0 >> 4) & 0x0F0F0F0F, 0x08080808); +#else + x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_0; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kbxd] = bxi->d; +#else + x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + kbxd] = bxi->d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_1( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, I); + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_1); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI4_1; + const int kqsx = txi % QI4_1; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx; + const int qs0 = get_int_b4(bxi->qs, kqsx); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kbx*(2*QI4_1) + kqsx + 0] = (qs0 >> 0) & 0x0F0F0F0F; + x_qs[i*sram_stride + kbx*(2*QI4_1) + kqsx + QI4_1] = (qs0 >> 4) & 0x0F0F0F0F; +#else + x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_1; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_dm[i*sram_stride + kbxd] = bxi->dm; +#else + x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + kbxd] = bxi->dm; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_0( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_0); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI5_0; + const int kqsx = txi % QI5_0; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbx; + + const int ql = get_int_b2(bxi->qs, kqsx); + const int qh = get_int_b2(bxi->qh, 0) >> (4 * kqsx); + + int qs0 = (ql >> 0) & 0x0F0F0F0F; + qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 + qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 + qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 + qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 + qs0 = __vsubss4(qs0, 0x10101010); // subtract 16 + + int qs1 = (ql >> 4) & 0x0F0F0F0F; + qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 + qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 + qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 + qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 + qs1 = __vsubss4(qs1, 0x10101010); // subtract 16 + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kbx*(2*QI5_0) + kqsx + 0] = qs0; + x_qs[i*sram_stride + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + 0] = qs0; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_0; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kbxd] = bxi->d; +#else + x_df[i*(MMQ_TILE_NE_K/QI5_0) + i/QI5_0 + kbxd] = bxi->d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_1( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, I); + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_1); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI5_1; + const int kqsx = txi % QI5_1; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbx; + + const int ql = get_int_b4(bxi->qs, kqsx); + const int qh = get_int_b4(bxi->qh, 0) >> (4 * kqsx); + + int qs0 = (ql >> 0) & 0x0F0F0F0F; + qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 + qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 + qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 + qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 + + int qs1 = (ql >> 4) & 0x0F0F0F0F; + qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 + qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 + qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 + qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kbx*(2*QI5_1) + kqsx + 0] = qs0; + x_qs[i*sram_stride + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + 0] = qs0; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_1; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_dm[i*sram_stride + kbxd] = bxi->dm; +#else + x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + kbxd] = bxi->dm; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q8_0( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_tile + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + // MMQ_ITER_K / (4 * QR8_0) == 64 required. but NV has only 32 threads per warp + constexpr int threads_per_row = 32; + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI8_0; + const int kqsx = txi % QI8_0; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 0 + txi] = get_int_b2(bxi[0].qs, kqsx); + x_qs[i*sram_stride + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx); +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 0 + txi] = get_int_b2(bxi[0].qs, kqsx); + x_qs[i*(2*MMQ_TILE_NE_K + 1) + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = 2*MMQ_TILE_NE_K / QI8_0; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kbxd] = bxi->d; +#else + x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + kbxd] = bxi->d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +// --------------------------------------------------------------------------------------------- + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q2_K( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, I); + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR2_K); + constexpr int nrows = ggml_cuda_get_physical_warp_size() / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q2_K * bxi = (const block_q2_K *) x + kbx0 + i*stride; + + const int x_ql_0 = get_int_b2(bxi->qs, kqsx); + +#pragma unroll + for (int l = 0; l < QR2_K; ++l) { + const int k = (kqsx/8)*32 + l*8 + kqsx % 8; + + const int x_qs_k = (x_ql_0 >> (2*l)) & 0x03030303; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + k] = x_qs_k; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int sc_m = bxi->scales[kqsx]; +#ifdef FAST_FP16_AVAILABLE + const half2 x_dm_ik = __hmul2(bxi->dm, make_half2(sc_m & 0x0F, sc_m >> 4)); +#else + const float2 bxi_dmf = __half22float2(bxi->dm); + const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4)); +#endif // FAST_FP16_AVAILABLE + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_dm[i*sram_stride + kqsx] = x_dm_ik; +#else + x_dm[i*(MMQ_TILE_NE_K + 1) + kqsx] = x_dm_ik; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q3_K( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); + int * x_sc = (int *) (x_df + txs.dm); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR3_K); + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; + + const int x_ql_0 = get_int_b2(bxi->qs, kqsx); + const int x_qh_0 = get_int_b2(bxi->hmask, kqsx % (QI3_K/2)) >> (4 * (kqsx / (QI3_K/2))); + +#pragma unroll + for (int l = 0; l < QR3_K; ++l) { + const int k = (kqsx/8)*32 + l*8 + kqsx % 8; + + const int x_ql_k = (x_ql_0 >> (2*l)) & 0x03030303; + const int x_qh_k = ((x_qh_0 >> l) << 2) & 0x04040404; + + const int x_qs_k = __vsubss4(x_ql_k | x_qh_k, 0x04040404); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + k] = x_qs_k; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + } + + constexpr int rows_per_warp = warp_size / 4; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { + int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/4; + + if (fallback) { + i = min(i, i_max); + } + + const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; + + const int ksc = threadIdx.x % 4; + + const int ksc_low = ksc % (QI3_K/8); + const int shift_low = 4 * (ksc / (QI3_K/8)); + const int sc_low = (get_int_b2(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; + + const int ksc_high = QI3_K/8; + const int shift_high = 2 * ksc; + const int sc_high = ((get_int_b2(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; + + const int sc = __vsubss4(sc_low | sc_high, 0x20202020); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + const int8_t * sc8 = (const int8_t *) ≻ + const float d = bxi->d; + +#pragma unroll + for (int l = 0; l < int(sizeof(int)); ++l) { + x_df[i*sram_stride + sizeof(int)*ksc + l] = d*sc8[l]; + } +#else + x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = sc; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + +#if !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)) +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) { + int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; + + x_df[i] = bxi->d; + } +#endif // !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)) || defined(AMD_WMMA_AVAILABLE) +} + +static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, const int ksc) { + // scale arrangement after the following two lines: + // - ksc == 0: sc0, sc1, sc2, sc3 + // - ksc == 1: sc4, sc5, sc6, sc7 + // - ksc == 2: m0, m1, m2, m3 + // - ksc == 3: m4, m5, m6, m7 + return ((scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F) | // lower 4 bits + ((scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030); // upper 2 bits +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_K( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, I); + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + txs.qs); + int * x_sc = (int *) (x_dm + txs.dm); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_K); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; + const int qs0 = get_int_b4(bxi->qs, txi); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 16*(txi/8) + txi % 8 + 0] = (qs0 >> 0) & 0x0F0F0F0F; + x_qs[i*sram_stride + 16*(txi/8) + txi % 8 + 8] = (qs0 >> 4) & 0x0F0F0F0F; +#else + x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr int rows_per_warp = warp_size / 2; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + // Need if on AMD instead of % because warp_size == 64 + // This causes double work and throughput loss (MI300X) + // H100 loses about 100 t/s with 'if' condition over '%' + int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/2; + if (i < I) { +#else + int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % I; + { +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + if (fallback) { + i = min(i, i_max); + } + + const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; + + const int * scales = (const int *) bxi->scales; + const int ksc = threadIdx.x % 2; + + const int sc32 = unpack_scales_q45_K(scales, ksc + 0); + const int m32 = unpack_scales_q45_K(scales, ksc + 2); + + const uint8_t * sc8 = (const uint8_t *) &sc32; + const uint8_t * m8 = (const uint8_t *) &m32; + + const half2 dm = bxi->dm * make_half2(1.0f, -1.0f); + + #pragma unroll + for (int l = 0; l < sizeof(int); ++l) { + x_dm[i*sram_stride + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]); + } + } + } +#else +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) { + int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; + + x_dm[i] = bxi->dm; + } + constexpr int rows_per_warp = warp_size / 4; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { + int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / (QI4_K/8); + + const int * scales = (const int *) bxi->scales; + + const int ksc = threadIdx.x % (MMQ_TILE_NE_K/8); + const int scales8 = unpack_scales_q45_K(scales, ksc); + + x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8; + } +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_K( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, I); + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + txs.qs); + int * x_sc = (int *) (x_dm + txs.dm); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_K); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; + const int ky = QR5_K*txi; + + const int ql = get_int_b4(bxi->qs, txi); + const int ql0 = (ql >> 0) & 0x0F0F0F0F; + const int ql1 = (ql >> 4) & 0x0F0F0F0F; + + const int qh = get_int_b4(bxi->qh, txi % (QI5_K/4)); + const int qh0 = ((qh >> (2 * (txi / (QI5_K/4)) + 0)) << 4) & 0x10101010; + const int qh1 = ((qh >> (2 * (txi / (QI5_K/4)) + 1)) << 4) & 0x10101010; + + const int kq0 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + 0; + const int kq1 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + QI5_K/4; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kq0] = ql0 | qh0; + x_qs[i*sram_stride + kq1] = ql1 | qh1; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = ql0 | qh0; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = ql1 | qh1; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr int rows_per_warp = warp_size / 2; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { +#if defined(AMD_MFMA_AVAILABLE) + // Need if on AMD instead of % because warp_size == 64 + // This causes double work and throughput loss (MI300X) + // H100 loses about 100 t/s with 'if' condition over '%' + int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/2; + if (i < I) { +#else + int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % I; + { +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + if (fallback) { + i = min(i, i_max); + } + + const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; + + const int * scales = (const int *) bxi->scales; + const int ksc = threadIdx.x % 2; + + const int sc32 = unpack_scales_q45_K(scales, ksc + 0); + const int m32 = unpack_scales_q45_K(scales, ksc + 2); + + const uint8_t * sc8 = (const uint8_t *) &sc32; + const uint8_t * m8 = (const uint8_t *) &m32; + + const half2 dm = bxi->dm * make_half2(1.0f, -1.0f); + +#pragma unroll + for (int l = 0; l < int(sizeof(int)); ++l) { + x_dm[i*sram_stride + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]); + } + } + } +#else +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) { + int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; + + x_dm[i] = bxi->dm; + } + + constexpr int rows_per_warp = warp_size / 4; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { + int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; + + const int * scales = (const int *) bxi->scales; + + const int ksc = threadIdx.x % (MMQ_TILE_NE_K/8); + const int scales8 = unpack_scales_q45_K(scales, ksc); + + x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8; + } +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q6_K( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); + int * x_sc = (int *) (x_df + MMQ_TILE_NE_K/QI6_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); + int * x_sc = (int *) (x_df + txs.dm); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR6_K); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride; + + const int ql = get_int_b2(bxi->ql, txi); + const int ql0 = (ql >> 0) & 0x0F0F0F0F; + const int ql1 = (ql >> 4) & 0x0F0F0F0F; + + const int qh = get_int_b2(bxi->qh, (QI6_K/4) * (txi / (QI6_K/2)) + txi % (QI6_K/4)); + const int qh0 = ((qh >> ((txi & 0x08) >> 2)) << 4) & 0x30303030; + const int qh1 = (qh >> ((txi & 0x08) >> 2)) & 0x30303030; + + const int kq0 = 2*txi - txi % (QI6_K/2) + 0; + const int kq1 = 2*txi - txi % (QI6_K/2) + QI6_K/2; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kq0] = __vsubss4(ql0 | qh0, 0x20202020); + x_qs[i*sram_stride + kq1] = __vsubss4(ql1 | qh1, 0x20202020); +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020); + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) { + int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride] = bxi->d; +#else + x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K] = bxi->d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int rows_per_warp = warp_size / 4; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { + int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / 4; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_sc[i*sram_stride + threadIdx.x%4] = get_int_b2(bxi->scales, threadIdx.x % (MMQ_TILE_NE_K/8)); +#else + x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + threadIdx.x%(MMQ_TILE_NE_K/8)] = get_int_b2(bxi->scales, threadIdx.x%(QI6_K/8)); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +// --------------------------------------------------------------------------------------------- + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq1_s( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_ds = (half2 *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, I); + int * x_qs = (int *) x_tile; + half2 * x_ds = (half2 *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR1_S); + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq1_s * bxi = (const block_iq1_s *) x + kbx0 + i*stride; + + const int qs_packed = get_int_b2(bxi->qs, kqsx); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bxi->qh[kqsx]; + + #pragma unroll + for (int l = 0; l < QR1_S/2; ++l) { + const int grid = iq1s_grid_gpu[qs[l] | (((qh >> (3*l)) & 0x07) << 8)]; + + const int grid0 = (grid >> 0) & 0x0F0F0F0F; + const int grid1 = (grid >> 4) & 0x0F0F0F0F; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l+0)] = grid0; + x_qs[i*sram_stride + 8*kqsx + (2*l+1)] = grid1; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid0; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid1; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const float d1q = __half2float(bxi->d) * (((qh >> 11) & 0x0E) + 1); + const float delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_ds[i*sram_stride + kqsx] = make_half2(d1q, d1q*delta); +#else + x_ds[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = make_half2(d1q, d1q*delta); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_xxs( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XXS, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XXS)) / 2; + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq2_xxs * bxi = (const block_iq2_xxs *) x + kbx0 + i*stride; + + const int q2 = get_int_b2(bxi->qs, 2*kqsx+0); + const uint8_t * aux8 = (const uint8_t *) &q2; + const uint32_t aux32 = get_int_b2(bxi->qs, 2*kqsx+1); + +#pragma unroll + for (int l = 0; l < QR2_XXS; ++l) { + const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[l]]; + const uint32_t signs = unpack_ksigns(aux32 >> (7 * l)); + + const int signs0 = __vcmpne4(signs & 0x08040201, 0); + const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0); + + const int signs1 = __vcmpne4(signs & 0x80402010, 0); + const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid0; + x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid1; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid0; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid1; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int ls = aux32 >> 27 | 1; // (scale * 2 + 1) + const float d = bxi->d; +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4 +#else + x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4 +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_xs( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XS, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XS)) / 2; + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq2_xs * bxi = (const block_iq2_xs *) x + kbx0 + i*stride; + + const int2 q2_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); + const uint16_t * q2 = (const uint16_t *) &q2_packed; + + #pragma unroll + for (int l = 0; l < QR2_XS; ++l) { + const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l] & 0x1FF]; + const uint32_t signs = unpack_ksigns(q2[l] >> 9); + + const int signs0 = __vcmpne4(signs & 0x08040201, 0); + const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); + + const int signs1 = __vcmpne4(signs & 0x80402010, 0); + const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid_h; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int ls = bxi->scales[kqsx]; + const float d = bxi->d; +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; + x_df[i*sram_stride + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; +#else + x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; + x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_s( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_S, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_S)) / 2; + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq2_s * bxi = (const block_iq2_s *) x + kbx0 + i*stride; + + const int qs_packed = get_int_b2(bxi->qs, kqsx); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bxi->qh[kqsx]; + + const int signs_packed_32 = get_int_b2(bxi->qs, QK_K/32 + kqsx); + const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; + +#pragma unroll + for (int l = 0; l < QR2_S; ++l) { + const int * grid_pos = (const int *)(iq2s_grid + (qs[l] | ((qh << (8-2*l)) & 0x300))); + + const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000); + const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000); + + const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0); + const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid_h; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int ls = bxi->scales[kqsx]; + const float d = bxi->d; +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; + x_df[i*sram_stride + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; +#else + x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; + x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq3_xxs( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_XXS, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_XXS)) / 2; + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq3_xxs * bxi = (const block_iq3_xxs *) x + kbx0 + i*stride; + + const int2 q3_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); + const uint8_t * q3 = (const uint8_t *) &q3_packed; + const uint32_t aux32 = get_int_b2(bxi->qs, QK_K/16 + kqsx); + +#pragma unroll + for (int l = 0; l < QR3_XXS; ++l) { + const int2 grid_pos = make_int2(iq3xxs_grid[q3[2*l+0]], iq3xxs_grid[q3[2*l+1]]); + const uint32_t signs = unpack_ksigns(aux32 >> (7*l)); + + const int signs0 = __vcmpne4(signs & 0x08040201, 0); + const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); + + const int signs1 = __vcmpne4(signs & 0x80402010, 0); + const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid_h; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int ls = aux32 >> 28; + const float d = bxi->d; +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kqsx] = (ls*d + d/2)/2; +#else + x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/2; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq3_s( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_S)) / 2; + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq3_s * bxi = (const block_iq3_s *) x + kbx0 + i*stride; + + const int2 qs_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bxi->qh[kqsx]; + + const int signs_packed_32 = get_int_b2(bxi->signs, kqsx); + const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; + +#pragma unroll + for (int l = 0; l < QR3_S; ++l) { + const int2 grid_pos = make_int2( + iq3s_grid[qs[2*l+0] | ((qh << (8 - 2*l)) & 0x100)], + iq3s_grid[qs[2*l+1] | ((qh << (7 - 2*l)) & 0x100)]); + + const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000); + const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000); + + const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); + const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l+0)] = grid_l; + x_qs[i*sram_stride + 8*kqsx + (2*l+1)] = grid_h; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid_l; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid_h; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int ls = 1 + 2*((bxi->scales[kqsx/2] >> (((2*kqsx) << 1) & 0x04)) & 0x0F); + const float d = bxi->d; +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kqsx] = ls*d; +#else + x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = ls*d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq4_xs( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_XS); + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride; + + const int aux_q4 = get_int_b4(bxi->qs, kqsx); + const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl); + const int k0 = 8 * (kqsx / 4) + kqsx % 4; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + k0 + 0] = v.x; + x_qs[i*sram_stride + k0 + 4] = v.y; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 4] = v.y; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int rows_per_warp = warp_size / 8; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / (MMQ_TILE_NE_K/4); + + if (fallback) { + i = min(i, i_max); + } + + const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride; + + const float d = __half2float(bxi->d); + + const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F) + | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + threadIdx.x % 8] = d * (ls - 32); +#else + x_df[i*(MMQ_TILE_NE_K/4) + i/4 + threadIdx.x % 8] = d * (ls - 32); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq4_nl( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_NL); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI4_NL; + const int kqsx = txi % QI4_NL; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbx; + + const int aux_q4 = get_int_b2(bxi->qs, kqsx); + const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl); + const int k0 = kbx * (2 * QI4_NL) + kqsx; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + k0 + 0] = v.x; + x_qs[i*sram_stride + k0 + QI4_NL] = v.y; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI4_NL] = v.y; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_NL; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kbxd] = __half2float(bxi->d); +#else + x_df[i*(MMQ_TILE_NE_K/QI4_NL) + i/QI4_NL + kbxd] = __half2float(bxi->d); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +// --------------------------------------------------------------------------------------------- + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_mxfp4( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_MXFP4, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR_MXFP4); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI_MXFP4; + const int kqsx = txi % QI_MXFP4; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbx; + + const int aux_q4 = get_int_b1(bxi->qs, kqsx); + const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4); + const int k0 = kbx * (2 * QI_MXFP4) + kqsx; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + k0 + 0] = v.x; + x_qs[i*sram_stride + k0 + QI_MXFP4] = v.y; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI_MXFP4] = v.y; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI_MXFP4; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f; +#else + x_df[i*(MMQ_TILE_NE_K/QI_MXFP4) + i/QI_MXFP4 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_mxfp4_fp4( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + + int * x_qs = (int *) x_tile; + uint32_t * x_sc = (uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K); + + const int txi = threadIdx.x; + + constexpr int iter_k = ggml_cuda_mmq_get_K_vram(type, J, fallback); + + constexpr int threads_per_row = iter_k / QK_MXFP4; // each thread processes 1 block + constexpr int rows_per_warp = warp_size / threads_per_row; + const int kbx = txi % threads_per_row; + const int row_in_warp = txi / threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += rows_per_warp * nwarps) { + int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; + + if constexpr (fallback) { + i = min(i, i_max); + } + + const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i * stride + kbx; + + // quantize_mxfp4_mmq permutes nibbles to match the quantized format + const int k0 = kbx * 4; + memcpy(x_qs + i*sram_stride + k0, bxi->qs, 16); + + // Load E8M0 scales: pack 2 consecutive scales into one uint32 + if (kbx % 2 == 0) { + uint32_t e = bxi->e; + e |= ((bxi + 1)->e << 8); + x_sc[i*sram_stride + kbx / 2] = e; + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_nvfp4( + const char * __restrict__ x, int * __restrict__ x_tile, const int kb0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_NVFP4, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / QK_NVFP4; + constexpr int rows_per_warp = warp_size / threads_per_row; + const int kbx = threadIdx.x % threads_per_row; + const int row_in_warp = threadIdx.x / threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += rows_per_warp * nwarps) { + int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; + + if constexpr (fallback) { + i = min(i, i_max); + } + + const block_nvfp4 * bxi = (const block_nvfp4 *) x + kb0 + i * stride + kbx; + const uint32_t * __restrict__ src_qs = reinterpret_cast(bxi->qs); + const int kqs = 16 * kbx; + const int ksc = 4 * kbx; + +#pragma unroll + for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) { + const int2 q0 = get_int_from_table_16(src_qs[2 * sub + 0], kvalues_mxfp4); + const int2 q1 = get_int_from_table_16(src_qs[2 * sub + 1], kvalues_mxfp4); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kqs + 4 * sub + 0] = q0.x; + x_qs[i*sram_stride + kqs + 4 * sub + 1] = q1.x; + x_qs[i*sram_stride + kqs + 4 * sub + 2] = q0.y; + x_qs[i*sram_stride + kqs + 4 * sub + 3] = q1.y; + x_df[i*sram_stride + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]); +#else + x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 0] = q0.x; + x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 1] = q1.x; + x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 2] = q0.y; + x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 3] = q1.y; + x_df[i * (2 * MMQ_TILE_NE_K * 2 / QI_NVFP4) + i / (QK_NVFP4_SUB / QI_NVFP4) + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_nvfp4_nvfp4( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int iter_k = ggml_cuda_mmq_get_K_vram(type, J, fallback); + constexpr int threads_per_row = iter_k / QK_NVFP4; // each thread processes 1 block + constexpr int rows_per_warp = warp_size / threads_per_row; + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + + uint32_t * x_u32 = (uint32_t *) x_tile; + + const int txi = threadIdx.x; + const int kbx = txi % threads_per_row; + const int row_in_warp = txi / threads_per_row; + + const block_nvfp4 * bxi_base = (const block_nvfp4 *) x + kbx0 + kbx; + uint32_t * x_u32_scale = x_u32 + 64 + kbx; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += rows_per_warp * nwarps) { + int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; + + if constexpr (fallback) { + i = min(i, i_max); + } + + const block_nvfp4 * bxi = bxi_base + i * stride; + + const uint32_t * src_qs = reinterpret_cast(bxi->qs); + +#pragma unroll + for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) { + x_u32[i*sram_stride + 8*kbx + 2 * sub + 0] = src_qs[2 * sub + 0]; + x_u32[i*sram_stride + 8*kbx + 2 * sub + 1] = src_qs[2 * sub + 1]; + } + + x_u32_scale[i*sram_stride] = get_int_b4(bxi->d, 0); + } +} diff --git a/ggml/src/ggml-cuda/mmq-vec-dot.cuh b/ggml/src/ggml-cuda/mmq-vec-dot.cuh new file mode 100644 index 000000000..d57343386 --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-vec-dot.cuh @@ -0,0 +1,1251 @@ +#pragma once + +#include "vecdotq.cuh" +#include "mma.cuh" + +using namespace ggml_cuda_mma; + +#include "mmq.cuh" + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q4_0_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, I); + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_0*VDR_Q4_0_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2); + + int u[2*VDR_Q4_0_Q8_1_MMQ]; + + constexpr int max_cpy = ggml_cuda_get_max_cpy_bytes(); + constexpr int mcpy_int = max_cpy / sizeof(int); + static_assert(VDR_Q4_0_Q8_1_MMQ == 4, "bad VDR_Q4_0_Q8_1_MMQ"); + + int tmp0[4], tmp1[4]; + + #pragma unroll + for (int l0 = 0; l0 < 4 / mcpy_int; ++l0) { + ggml_cuda_memcpy_1(tmp0 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + l0 * mcpy_int] ); + ggml_cuda_memcpy_1(tmp1 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + QI4_0 + l0 * mcpy_int]); + } + + u[0]=tmp0[0]; u[2]=tmp0[1]; u[4]=tmp0[2]; u[6]=tmp0[3]; + u[1]=tmp1[0]; u[3]=tmp1[1]; u[5]=tmp1[2]; u[7]=tmp1[3]; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q4_0_q8_1_impl + (&x_qs[i*(MMQ_TILE_NE_K + 1) + k0/QR4_0], u, + x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + k0/(QR4_0*QI4_0)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q4_1_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, I); + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_1*VDR_Q4_1_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2); + + int u[2*VDR_Q4_1_Q8_1_MMQ]; + + constexpr int max_cpy = ggml_cuda_get_max_cpy_bytes(); + constexpr int mcpy_int = max_cpy / sizeof(int); + static_assert(VDR_Q4_0_Q8_1_MMQ == 4, "bad VDR_Q4_0_Q8_1_MMQ"); + + int tmp0[4], tmp1[4]; + + #pragma unroll + for (int l0 = 0; l0 < 4 / mcpy_int; ++l0) { + ggml_cuda_memcpy_1(tmp0 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + l0 * mcpy_int] ); + ggml_cuda_memcpy_1(tmp1 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + QI4_1 + l0 * mcpy_int]); + } + + u[0]=tmp0[0]; u[2]=tmp0[1]; u[4]=tmp0[2]; u[6]=tmp0[3]; + u[1]=tmp1[0]; u[3]=tmp1[1]; u[5]=tmp1[2]; u[7]=tmp1[3]; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q4_1_q8_1_impl + (&x_qs[i*(MMQ_TILE_NE_K + 1) + k0/QR4_1], u, + x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + k0/(QR4_1*QI4_1)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, I); + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += VDR_Q8_0_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q8_0_q8_1_impl + (&x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k0 % MMQ_TILE_NE_K], + x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + k0/QI8_0], y_df[j*MMQ_TILE_Y_K + (k0/QI8_1) % (MMQ_TILE_NE_K/QI8_1)]); + } + } + } +} + +template +static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr data_layout input_layout = get_input_data_layout(); + typedef tile<16, 8, int, input_layout> tile_A; + typedef tile<16, 8, int, input_layout> tile_B; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + 2*MMQ_TILE_NE_K; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + const half2 * y_ds = (const half2 *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + tile_B B; + load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + float dB; + const int j = j0 + tile_C::get_j(0); + if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) { + dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + } else { + dB = __low2float(y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_A::I + tile_C::get_i(l); + const float dA = x_df[i*sram_stride + k0/QI8_0]; + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l]*dA*dB; + } + } + } + } +#else + typedef tile<16, 8, int> tile_A; + typedef tile< 8, 8, int> tile_B; + typedef tile<16, 8, int> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + 2*MMQ_TILE_NE_K; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + const half2 * y_ds = (const half2 *) y; + + tile_A A[ntx][MMQ_TILE_NE_K/QI8_0]; + float dA[ntx][tile_C::ne/2][MMQ_TILE_NE_K/QI8_0]; + + const int i0 = (threadIdx.y/ntx)*rows_per_warp; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { + const int k0 = k00 + k01; + + load_ldmatrix(A[n][k01/QI8_0], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_A::I + tile_C::get_i(2*l); + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { + const int k0 = k00 + k01; + + dA[n][l][k01/QI8_0] = x_df[i*sram_stride + k0/QI8_0]; + } + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { + tile_B B; + float dB[tile_C::ne/2]; + + load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) { + dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + } else { + dB[l] = __low2float(y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n][k01/QI8_0], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l]*dA[n][l/2][k01/QI8_0]*dB[l%2]; + } + } + } + } +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) +} + + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_1_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, I); + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += VDR_Q8_0_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q8_1_q8_1_impl + (&x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], + x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + k0/QI8_1], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr data_layout input_layout = get_input_data_layout(); + typedef tile<16, 8, int, input_layout> tile_A; + typedef tile<16, 8, int, input_layout> tile_B; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + 2*MMQ_TILE_NE_K; + const int * y_qs = (const int *) y + 4; + const half2 * y_dm = (const half2 *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + tile_B B; + load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float2 dsB = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]); + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_A::I + tile_C::get_i(l); + float2 dmA = __half22float2(x_dm[i*sram_stride + k0/QI8_1]); + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA.x*dsB.x*C.x[l]; + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA.y*dsB.y; + } + } + } + } +#else + typedef tile<16, 8, int> tile_A; + typedef tile< 8, 8, int> tile_B; + typedef tile<16, 8, int> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + 2*MMQ_TILE_NE_K; + const int * y_qs = (const int *) y + 4; + const half2 * y_dm = (const half2 *) y; + + tile_A A[ntx][MMQ_TILE_NE_K/QI8_1]; + float2 dmA[ntx][tile_C::ne/2][MMQ_TILE_NE_K/QI8_1]; + + const int i0 = (threadIdx.y/ntx)*rows_per_warp; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + const int k0 = k00 + k01; + + load_ldmatrix(A[n][k01/QI8_1], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_A::I + tile_C::get_i(2*l); + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + const int k0 = k00 + k01; + + dmA[n][l][k01/QI8_1] = __half22float2(x_dm[i*sram_stride + k0/QI8_1]); + } + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + tile_B B; + float2 dsB[tile_C::ne/2]; + + load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + dsB[l] = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n][k01/QI8_1], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA[n][l/2][k01/QI8_1].x*dsB[l%2].x*C.x[l]; + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA[n][l/2][k01/QI8_1].y*dsB[l%2].y; + } + } + } + } +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) +} + +// Used for NVFP4, Q3_K, IQ2_S, and IQ2_XS +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, I); + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q8_0_16_q8_1_impl( + &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], + &y_qs[j*MMQ_TILE_Y_K + k01], + &x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + k0/(QI8_0/2)], + y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +// Used for Q3_K, IQ2_S, and IQ2_XS: +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr data_layout input_layout = get_input_data_layout(); + typedef tile<16, 4, int, input_layout> tile_A; + typedef tile<16, 4, int, input_layout> tile_B; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + tile_B B; + load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(l); + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * x_df[i*sram_stride + k0/4] * dB; + } + } + } + } +#elif defined(TURING_MMA_AVAILABLE) + + typedef tile<16, 4, int> tile_A; + typedef tile<16, 8, int> tile_A_8; + typedef tile< 8, 4, int> tile_B; + typedef tile<16, 8, int> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + + const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); + + tile_A A[ntx][8]; + float dA[ntx][tile_C::ne/2][8]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { + const int k0 = k00 + k01; + + load_ldmatrix(((tile_A_8 *) A[n])[k01/8], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + dA[n][l][k01/4] = x_df[i*sram_stride + k0/4]; + } + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) { + tile_B B[2]; + float dB[tile_C::ne/2]; + + // Here load_generic is faster than load_ldmatrix. + load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); + load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + (k01 + tile_B::J), MMQ_TILE_Y_K); + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C[2]; + mma(C[0], A[n][k01/4 + 0], B[0]); + mma(C[1], A[n][k01/4 + 1], B[1]); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dB[l%2]*(C[0].x[l]*dA[n][l/2][k01/4 + 0] + C[1].x[l]*dA[n][l/2][k01/4 + 1]); + } + } + } + } +#else + GGML_UNUSED_VARS(x, y, sum, k00); + NO_DEVICE_CODE; +#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q2_K_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, I); + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + + float2 y_df[J/nwarps]; +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + + y_df[j0/nwarps] = __half22float2(y_ds[j*MMQ_TILE_Y_K]); + } + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K/2; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + constexpr int ns = 2; + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q2_K_q8_1_impl_mmq( + &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], + &x_dm[i*(MMQ_TILE_NE_K + 1) + k0/4], k01 < MMQ_TILE_NE_K/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, + &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); + } + } + } + + // Some compilers fail to unroll the loop over k01 if there is a conditional statement for ns in the inner loop. + // As a workaround 2 separate loops are used instead. +#pragma unroll + for (int k01 = MMQ_TILE_NE_K/2; k01 < MMQ_TILE_NE_K; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + constexpr int ns = 1; + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q2_K_q8_1_impl_mmq( + &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], + &x_dm[i*(MMQ_TILE_NE_K + 1) + k0/4], k01 < MMQ_TILE_NE_K/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, + &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q2_K_q8_1_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr data_layout input_layout = get_input_data_layout(); + typedef tile<16, 4, int, input_layout> tile_A; + typedef tile<16, 4, int, input_layout> tile_B; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + tile_B B; + load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y; + const float sB = (k01 >= MMQ_TILE_NE_K * 3/4) ? 0 + : (((k01/4)%2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).y + : __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).x); + + tile_C Cm; + if (k01 >= MMQ_TILE_NE_K * 3/4) { + tile_A A1; +#pragma unroll + for (int l = 0; l < tile_A::ne; ++l) { + A1.x[l] = 0x01010101; + } + mma(Cm, A1, B); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C Cd; + mma(Cd, A[n], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(l); + const float2 dm = __half22float2(x_dm[i*sram_stride + k0/4]); + float tmp = Cd.x[l]*dm.x; + if (k01 >= MMQ_TILE_NE_K * 3/4) { + tmp -= Cm.x[l]*dm.y; + } + sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*dB; + sum[(j0/tile_C::J + n)*tile_C::ne + l] -= dm.y*sB; + } + } + } + } +#elif defined(TURING_MMA_AVAILABLE) + + typedef tile<16, 4, int> tile_A; + typedef tile<16, 8, int> tile_A_8; + typedef tile< 8, 4, int> tile_B; + typedef tile<16, 8, int> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + + const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); + + tile_A A[ntx][8]; + float dA[ntx][tile_C::ne/2][8]; + float mA[ntx][tile_C::ne/2][8]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + const int k0 = k00 + k01; + + load_ldmatrix(((tile_A_8 *) A[n])[k01/QI8_1], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1/2) { + const int k0 = k00 + k01; + + const float2 dm = __half22float2(x_dm[i*sram_stride + k0/(QI8_1/2)]); + + dA[n][l][k01/(QI8_1/2)] = dm.x; + mA[n][l][k01/(QI8_1/2)] = dm.y; + } + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + float2 dB[tile_C::ne/2]; + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + dB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K]); + } + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + tile_B B[2]; + + // Here load_generic is faster than load_ldmatrix. + load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); + load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + (k01 + tile_B::J), MMQ_TILE_Y_K); + + tile_C Cm[2]; + if (k01 >= MMQ_TILE_NE_K * 3/4) { + tile_A A1; + A1.x[0] = 0x01010101; + A1.x[1] = 0x01010101; + mma(Cm[0], A1, B[0]); + mma(Cm[1], A1, B[1]); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C Cd[2]; + + mma(Cd[0], A[n][k01/4 + 0], B[0]); + mma(Cd[1], A[n][k01/4 + 1], B[1]); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + float tmp = Cd[0].x[l]*dA[n][l/2][k01/4 + 0] + Cd[1].x[l]*dA[n][l/2][k01/4 + 1]; + if (k01 >= MMQ_TILE_NE_K * 3/4) { + tmp -= Cm[0].x[l]*mA[n][l/2][k01/4 + 0] + Cm[1].x[l]*mA[n][l/2][k01/4 + 1]; + } + sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*(k01 < MMQ_TILE_NE_K/2 ? dB[l%2].x : dB[l%2].y); + } + } + } + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K * 3/4; k01 += QI8_1) { + float2 sB[tile_C::ne/2]; + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + sB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] -= mA[n][l/2][k01/4 + 0]*sB[l%2].x; + sum[(j0/tile_C::J + n)*tile_C::ne + l] -= mA[n][l/2][k01/4 + 1]*sB[l%2].y; + } + } + } + } +#else + GGML_UNUSED_VARS(x, y, sum, k00); + NO_DEVICE_CODE; +#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q3_K_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, I); + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + txs.qs; + const int * x_sc = (const int *) x_df + txs.dm; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + const int8_t * scales = ((const int8_t *) (x_sc + i*(MMQ_TILE_NE_K/8) + i/8)) + k0/4; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q3_K_q8_1_impl_mmq( + &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], scales, + x_df[i], y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q4_K_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, I); + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + txs.qs; + const int * x_sc = (const int *) x_dm + txs.dm; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_K*VDR_Q4_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + const uint8_t * sc = (const uint8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k0/32] + 2*(k01/16); + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q4_K_q8_1_impl_mmq( + &x_qs[i*(MMQ_TILE_NE_K + 1) + k0/2], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8, + x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q5_K_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, I); + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + txs.qs; + const int * x_sc = (const int *) x_dm + txs.dm; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR5_K*VDR_Q5_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + const uint8_t * sc = ((const uint8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k00/32]) + 2*(k01/16); + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q5_K_q8_1_impl_mmq( + &x_qs[i*(QR5_K*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8, + x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q6_K_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, I); + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + txs.qs; + const int * x_sc = (const int *) x_df + txs.dm; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR6_K*VDR_Q6_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + const int8_t * sc = ((const int8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k0/16]); + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q6_K_q8_1_impl_mmq( + &x_qs[i*(QR6_K*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc, + x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K], &y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q6_K_q8_1_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr data_layout input_layout = get_input_data_layout(); + typedef tile<16, 4, int, input_layout> tile_A; + typedef tile<16, 4, int, input_layout> tile_B; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; + const int * x_sc = (const int *) x_df + MMQ_TILE_NE_K/QI6_K; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + tile_B B; + load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(l); + const int8_t * sc = (const int8_t *) (x_sc + i*sram_stride + k00/16); + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * sc[k01/4] * x_df[i*sram_stride] * dB; + } + } + } + } +#elif defined(TURING_MMA_AVAILABLE) + + typedef tile<16, 4, int> tile_A; + typedef tile< 8, 4, int> tile_B; + typedef tile<16, 8, int> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; + const int * x_sc = (const int *) x_df + MMQ_TILE_NE_K/QI6_K; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + + const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); + + tile_A A[ntx][8]; + int scA[ntx][tile_C::ne/2][8]; + float dA[ntx][tile_C::ne/2]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { + const int k0 = k00 + k01; + + load_ldmatrix(A[n][k01/4 + 0], x_qs + (i0 + n*tile_A::I)*sram_stride + (k0 + 0), sram_stride); + load_ldmatrix(A[n][k01/4 + 1], x_qs + (i0 + n*tile_A::I)*sram_stride + (k0 + tile_A::J), sram_stride); + } + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 16) { + const int k0 = k00 + k01; + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); + + const int sc_packed = x_sc[i*sram_stride + k0/16]; + const int8_t * sc = (const int8_t *) &sc_packed; + +#pragma unroll + for (int ksc = 0; ksc < sizeof(int); ++ksc) { + scA[n][l][k01/4 + ksc] = sc[ksc]; + } + } + } + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); + + dA[n][l] = x_df[i*sram_stride]; + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + float tmp[ntx][tile_C::ne] = {{0.0f}}; + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { + tile_B B[2]; + float dB[tile_C::ne/2]; + + // Here load_generic is faster than load_ldmatrix. + load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + 0 + k01, MMQ_TILE_Y_K); + load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + tile_B::J + k01, MMQ_TILE_Y_K); + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C[2]; + mma(C[0], A[n][k01/4 + 0], B[0]); + mma(C[1], A[n][k01/4 + 1], B[1]); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + tmp[n][l] += (C[0].x[l]*scA[n][l/2][k01/4 + 0] + C[1].x[l]*scA[n][l/2][k01/4 + 1])*dB[l%2]; + } + } + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp[n][l]*dA[n][l/2]; + } + } + } +#else + GGML_UNUSED_VARS(x, y, sum, k00); + NO_DEVICE_CODE; +#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE +} + +// --------------------------------------------------------------------------------------------- + +// Shared MMA kernel for MXFP4 and NVFP4 on Blackwell. +// Both quantizations encode values as e2m1 (FP4) and produce one uint32 scale per +// m16n8k64 MMA call; only the PTX kind (scale_vec::2X ue8m0 vs scale_vec::4X ue4m3) +// and the per-type stride constant differ. +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_fp4_fp4_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + + typedef tile<16, 8, int> tile_A; + typedef tile<8, 8, int> tile_B; + typedef tile<16, 8, float> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp / tile_C::I; + constexpr int nfrags = MMQ_TILE_NE_K / tile_A::J; + + y += (threadIdx.y % ntx) * (tile_C::J * MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const uint32_t * x_sc = (const uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K); + const int * y_qs = (const int *) y + 4; + const uint32_t * y_sc = (const uint32_t *) y; + + // 2 threads per quad supply the packed scale register to the block_scale MMA, + // see https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-block-scaling + const int tidx_A = threadIdx.x / 4 + (threadIdx.x % 2) * 8; + const int tidx_B = threadIdx.x / 4; + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + tile_A A[ntx][nfrags]; + uint32_t scaleA[ntx][nfrags]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int frag = 0; frag < nfrags; ++frag) { + const int k0 = k00 + frag * tile_A::J; + load_ldmatrix(A[n][frag], x_qs + (i0 + n * tile_A::I) * sram_stride + k0, sram_stride); + scaleA[n][frag] = x_sc[(i0 + n * tile_A::I + tidx_A) * sram_stride + k0 / tile_A::J]; + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx * tile_C::J) { + tile_B B[nfrags]; + uint32_t scaleB[nfrags]; + +#pragma unroll + for (int frag = 0; frag < nfrags; ++frag) { + const int k0 = frag * tile_B::J; + load_generic(B[frag], y_qs + j0 * MMQ_TILE_Y_K + k0, MMQ_TILE_Y_K); + scaleB[frag] = y_sc[(j0 + tidx_B) * MMQ_TILE_Y_K + frag]; + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int frag = 0; frag < nfrags; ++frag) { + tile_C C = {}; + mma_block_scaled_fp4(C, A[n][frag], B[frag], scaleA[n][frag], scaleB[frag]); +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0 / tile_C::J + n) * tile_C::ne + l] += C.x[l]; + } + } + } + } +} + diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu index 6b3b0d064..bf9f5d526 100644 --- a/ggml/src/ggml-cuda/mmq.cu +++ b/ggml/src/ggml-cuda/mmq.cu @@ -3,6 +3,8 @@ #include "quantize.cuh" #include "mmid.cuh" +#include + static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { switch (args.type_x) { case GGML_TYPE_Q1_0: @@ -118,15 +120,14 @@ void ggml_cuda_mul_mat_q( const int64_t s03 = src0->nb[3] / ts_src0; const int64_t s3 = dst->nb[3] / ts_dst; - const bool use_stream_k = (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) - || GGML_CUDA_CC_IS_CDNA(cc); + const bool fallback = ne01 % 128 != 0; // TODO: tighter pool buffer size vs q8 path const bool use_native_fp4 = blackwell_mma_available(cc) && (src0->type == GGML_TYPE_MXFP4 || src0->type == GGML_TYPE_NVFP4); if (!ids) { const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 + - get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq); + ggml_cuda_mmq_get_J_max(src0->type, fallback, cc, ne11) * sizeof(block_q8_1_mmq); ggml_cuda_pool_alloc src1_q8_1(ctx.pool(), nbytes_src1_q8_1); { @@ -156,7 +157,7 @@ void ggml_cuda_mul_mat_q( ne00, ne01, ne1, s01, ne11, s1, ne02, ne12, s02, s12, s2, ne03, ne13, s03, s13, s3, - use_stream_k, ne1}; + ne1}; ggml_cuda_mul_mat_q_switch_type(ctx, args, stream); return; } @@ -184,7 +185,7 @@ void ggml_cuda_mul_mat_q( } const size_t nbytes_src1_q8_1 = ne12*n_expert_used*ne10_padded * sizeof(block_q8_1)/QK8_1 + - get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq); + ggml_cuda_mmq_get_J_max(src0->type, fallback, cc, ne11) * sizeof(block_q8_1_mmq); ggml_cuda_pool_alloc src1_q8_1(ctx.pool(), nbytes_src1_q8_1); const int64_t ne11_flat = ne12*n_expert_used; @@ -217,53 +218,11 @@ void ggml_cuda_mul_mat_q( ne00, ne01, ne_get_rows, s01, ne_get_rows, s1, ne02, ne02, s02, s12, s2, ne03, ne13, s03, s13, s3, - use_stream_k, ne12}; + ne12}; ggml_cuda_mul_mat_q_switch_type(ctx, args, stream); } -void ggml_cuda_op_mul_mat_q( - ggml_backend_cuda_context & ctx, - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, - const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, - const int64_t src1_padded_row_size, cudaStream_t stream) { - - const int64_t ne00 = src0->ne[0]; - - const int64_t ne10 = src1->ne[0]; - const int64_t ne11 = src1->ne[1]; - GGML_ASSERT(ne10 % QK8_1 == 0); - - const int64_t ne0 = dst->ne[0]; - - const int64_t row_diff = row_high - row_low; - const int64_t stride01 = ne00 / ggml_blck_size(src0->type); - - const int id = ggml_cuda_get_device(); - const int cc = ggml_cuda_info().devices[id].cc; - - // the main device has a larger memory buffer to hold the results from all GPUs - // nrows_dst == nrows of the matrix that the kernel writes into - const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff; - - // The stream-k decomposition is only faster for recent NVIDIA GPUs. - // Also its fixup needs to allocate a temporary buffer in the memory pool. - // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer. - const bool use_stream_k = ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) - || GGML_CUDA_CC_IS_CDNA(cc)) - && src1_ncols == ne11; - const mmq_args args = { - src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i, - ne00, row_diff, src1_ncols, stride01, ne11, nrows_dst, - 1, 1, 0, 0, 0, - 1, 1, 0, 0, 0, - use_stream_k, src1_ncols}; - - ggml_cuda_mul_mat_q_switch_type(ctx, args, stream); - - GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_padded_row_size); -} - bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t n_experts) { #ifdef GGML_CUDA_FORCE_CUBLAS return false; diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index edf546d8f..607e433bf 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -1,22 +1,18 @@ #pragma once #include "common.cuh" -#include "vecdotq.cuh" -#include "mma.cuh" #include #include -using namespace ggml_cuda_mma; - #define MMQ_DP4A_MAX_BATCH_SIZE 64 // Max. batch size to use for dp4a MMQ kernels when FP16 tensor cores are available. #define MMQ_ITER_K 256 #define MMQ_ITER_K_FP4 512 #define MMQ_NWARPS 8 -typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int kbx0, const int i_max, const int stride); -typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00); -typedef void (*mmq_write_back_t)(const float * __restrict__ sum, const int32_t * __restrict__ get_rows_to_sorted, +typedef void (*ggml_cuda_mmq_load_tiles_t)(const char * __restrict__ x, int * x_tile, const int kbx0, const int i_max, const int stride); +typedef void (*ggml_cuda_mmq_vec_dot_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00); +typedef void (*ggml_cuda_mmq_write_back_t)(const float * __restrict__ sum, const int32_t * __restrict__ get_rows_to_sorted, float * __restrict__ dst, const int stride, const int i_max, const int j_max); enum mmq_q8_1_ds_layout { @@ -106,69 +102,6 @@ struct tile_x_sizes { int sc; }; -static int get_mmq_x_max_host(const int cc) { - return (turing_mma_available(cc) || amd_wmma_available(cc)) ? 128 : - GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ? -#ifdef GGML_CUDA_FORCE_MMQ - 128 : 64; -#else - MMQ_DP4A_MAX_BATCH_SIZE : 64; -#endif // GGML_CUDA_FORCE_MMQ -} - -static constexpr __device__ int get_mmq_x_max_device() { -#if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - return 128; -#else // defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - -#if defined(GGML_USE_HIP) - return 64; -#else // defined(GGML_USE_HIP) - -#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA -#ifdef GGML_CUDA_FORCE_MMQ - return 128; -#else // GGML_CUDA_FORCE_MMQ - return MMQ_DP4A_MAX_BATCH_SIZE; -#endif // GGML_CUDA_FORCE_MMQ -#else // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA - return 64; -#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA - -#endif // defined(GGML_USE_HIP) -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -} - -static int get_mmq_y_host(const int cc) { - return GGML_CUDA_CC_IS_AMD(cc) ? (GGML_CUDA_CC_IS_RDNA1(cc) ? 64 : 128) : - ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ? 128 : 64); -} - -static constexpr __device__ int get_iter_k([[maybe_unused]] const ggml_type type) { -#if defined(BLACKWELL_MMA_AVAILABLE) -if (type == GGML_TYPE_NVFP4 || type == GGML_TYPE_MXFP4) { - return MMQ_ITER_K_FP4; -} -#endif // defined(BLACKWELL_MMA_AVAILABLE) - return MMQ_ITER_K; -} - -static constexpr __device__ int get_mmq_y_device() { -#if defined(GGML_USE_HIP) -#if defined(RDNA1) - return 64; -#else - return 128; -#endif // defined RDNA1 -#else -#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA - return 128; -#else - return 64; -#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA -#endif // defined(GGML_USE_HIP) -} - // Decouple shared memory tile sizes from WARP_SIZE to allow for different warp sizes. // The K dimension of the tiles has either, // 1*MMQ_TILE_NE_K==32 (always for TILE_Y_K) or 2*MMQ_TILE_NE_K==64 (typically for TILE_X_K), @@ -178,18 +111,263 @@ static constexpr __device__ int get_mmq_y_device() { // in terms of 32 bit elements that means K % 2 == 1 for dp4a or K % 8 == 4 for mma. #define MMQ_TILE_NE_K 32 -#define MMQ_DP4A_TXS_Q4_0 tile_x_sizes{mmq_y*MMQ_TILE_NE_K + mmq_y, mmq_y*MMQ_TILE_NE_K/QI4_0 + mmq_y/QI4_0, 0} -#define MMQ_DP4A_TXS_Q4_1 tile_x_sizes{mmq_y*MMQ_TILE_NE_K + mmq_y, mmq_y*MMQ_TILE_NE_K/QI4_1 + mmq_y/QI4_1, 0} -#define MMQ_DP4A_TXS_Q8_0 tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K*2/QI8_0 + mmq_y/(QI8_0/2), 0} -#define MMQ_DP4A_TXS_Q8_0_16 tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K*4/QI8_0 + mmq_y/(QI8_0/4), 0} -#define MMQ_DP4A_TXS_Q8_1 tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K*2/QI8_1 + mmq_y/(QI8_1/2), 0} -#define MMQ_DP4A_TXS_Q2_K tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K + mmq_y, 0} -#define MMQ_DP4A_TXS_Q3_K tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y, mmq_y*MMQ_TILE_NE_K/8 + mmq_y/8} -#define MMQ_DP4A_TXS_Q4_K tile_x_sizes{mmq_y*MMQ_TILE_NE_K + mmq_y, mmq_y*MMQ_TILE_NE_K/QI4_K, mmq_y*MMQ_TILE_NE_K/8 + mmq_y/8} -#define MMQ_DP4A_TXS_Q5_K tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K/QI5_K + mmq_y/QI5_K, mmq_y*MMQ_TILE_NE_K/8 + mmq_y/8} -#define MMQ_DP4A_TXS_Q6_K tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K/QI6_K + mmq_y/QI6_K, mmq_y*MMQ_TILE_NE_K/8 + mmq_y/8} +// block_q8_1_mmq has (128 8-bit ints == 32 32-bit ints + 4 32-bit scales) +#define MMQ_TILE_Y_K (MMQ_TILE_NE_K + MMQ_TILE_NE_K / QI8_1) +#define MMQ_TILE_Y_FP4_K MMQ_TILE_Y_K -static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int mmq_y) { +enum ggml_cuda_mmq_sram_layout { + GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, + GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, + GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, + GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, + GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, + GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, // MXFP4 and NVFP4 on Blackwell. + GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, // Generic NVFP4 +}; + +static constexpr __host__ __device__ int ggml_cuda_mmq_get_sram_stride(ggml_cuda_mmq_sram_layout sram_layout) { + switch (sram_layout) { + case GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0: + return 2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0 + 4; + case GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1: + return 2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_1 + 4; + case GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K: + return 2*MMQ_TILE_NE_K + MMQ_TILE_NE_K + 4; + case GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K: + return 2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2 + 4; + case GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K: + return 2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/QI6_K + MMQ_TILE_NE_K/8 + 7; + case GGML_CUDA_MMQ_SRAM_LAYOUT_FP4: + return 2*MMQ_TILE_NE_K + 8 + 4; + case GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4: + return 2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2 + 4; + default: + return -1; + } +} + +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_FP4) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4) % 8 == 4, "Wrong padding."); + +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_FP4) == ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1), "Wrong tile size for MXFP4"); + +// Config options for the MMQ kernel. +// Should not affect results, only speed/register pressure/shared memory use. +struct ggml_cuda_mmq_config { + ggml_type type; // src0->type + int nthreads; // Number of threads per CUDA block. + int occupancy; // Targeted occupancy for the MMA kernel. + int I; // SRAM tile width in src0->ne[1]/dst->ne[0] direction. + int J; // SRAM tile width in src1->ne[1]/dst->ne[1] direction. + ggml_cuda_mmq_sram_layout sram_layout; // SRAM tile length in src0->ne[0]/src1->ne[0] direction (physical 32 bit elements). + int K_vram; // VRAM tile length in src0->ne[0]/src1->ne[0] direction (logical elements). + bool stream_k; // Whether or not to use stream-k decomposition. + bool fallback; // Whether a fallback for out-of-bounds check in src0->ne[1] direction is needed. + + constexpr __host__ __device__ ggml_cuda_mmq_config( + ggml_type type, int nthreads, int occupancy, int I, int J, ggml_cuda_mmq_sram_layout sram_layout, int K_vram, bool stream_k, bool fallback) : + type(type), nthreads(nthreads), occupancy(occupancy), I(I), J(J), sram_layout(sram_layout), K_vram(K_vram), stream_k(stream_k), fallback(fallback) {} + + constexpr __device__ int rows_per_warp() const { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + return 16; +#else + return J >= 48 && J % 16 == 0 ? 32 : 16; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + // TODO transition all combinations of GPUs and quantizations to the MMA data layout. + __host__ int use_mma_data_layout(const int cc) const { + if (amd_mfma_available(cc) || amd_wmma_available(cc) || turing_mma_available(cc)) { + return true; + } + return false; + } + + constexpr __device__ bool use_mma_data_layout() const { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) + return true; +#else + return false; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) + } + +}; + +#define CASE(type_, nthreads_, occupancy_, I_, J_, sram_layout_, K_vram_, stream_k_, fallback_) \ + if (type == (type_) && J == (J_) && fallback == (fallback_)) { \ + static_assert((nthreads_) % 32 == 0 && (nthreads_) <= 512, "bad nthreads"); \ + static_assert( (occupancy_) <= 8, "bad occupancy"); \ + static_assert((I_) % 32 == 0, "bad I"); \ + static_assert((J_) % 8 == 0, "bad J"); \ + static_assert((K_vram_) % 256 == 0, "bad K_vram"); \ + return ggml_cuda_mmq_config((type_), (nthreads_), (occupancy_), (I_), (J_), (sram_layout_), (K_vram_), (stream_k_), (fallback_)); \ + } \ + +#include "mmq-config-pascal.cuh" +#include "mmq-config-ampere.cuh" +#include "mmq-config-blackwell.cuh" + +#include "mmq-config-cdna.cuh" +#include "mmq-config-rdna2.cuh" +#include "mmq-config-rdna4.cuh" + +#undef CASE + +static __host__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(const ggml_type type, const int J, const bool fallback, const int cc) { + if (GGML_CUDA_CC_IS_AMD(cc)) { + if (GGML_CUDA_CC_IS_CDNA(cc)) { + return ggml_cuda_mmq_get_config_cdna(type, J, fallback); + } + if (amd_wmma_available(cc)) { + return ggml_cuda_mmq_get_config_rdna4(type, J, fallback); + } + return ggml_cuda_mmq_get_config_rdna2(type, J, fallback); + } + if (blackwell_mma_available(cc)) { + return ggml_cuda_mmq_get_config_blackwell(type, J, fallback); + } + if (ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) { + return ggml_cuda_mmq_get_config_ampere(type, J, fallback); + } + return ggml_cuda_mmq_get_config_pascal(type, J, fallback); +} + +static constexpr __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(ggml_type type, int J, bool fallback) { +#ifdef GGML_USE_HIP +#ifdef CDNA + return ggml_cuda_mmq_get_config_cdna(type, J, fallback); +#elif defined(AMD_WMMA_AVAILABLE) + return ggml_cuda_mmq_get_config_rdna4(type, J, fallback); +#else + return ggml_cuda_mmq_get_config_rdna2(type, J, fallback); +#endif // CDNA +#else +#ifdef BLACKWELL_MMA_AVAILABLE + return ggml_cuda_mmq_get_config_blackwell(type, J, fallback); +#elif __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA + return ggml_cuda_mmq_get_config_ampere(type, J, fallback); +#else + return ggml_cuda_mmq_get_config_pascal(type, J, fallback); +#endif // BLACKWELL_MMA_AVAILABLE +#endif // GGML_USE_HIP + GGML_UNUSED_VARS(type, J, fallback); +} + +static __host__ int ggml_cuda_mmq_get_type(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).type; +} + +static constexpr __device__ int ggml_cuda_mmq_get_type(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).type; +} + +static __host__ int ggml_cuda_mmq_get_nthreads(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).nthreads; +} + +static constexpr __device__ int ggml_cuda_mmq_get_nthreads(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).nthreads; +} + +static __host__ int ggml_cuda_mmq_get_occupancy(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).occupancy; +} + +static constexpr __device__ int ggml_cuda_mmq_get_occupancy(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).occupancy; +} + +static __host__ int ggml_cuda_mmq_get_I(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).I; +} + +static constexpr __device__ int ggml_cuda_mmq_get_I(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).I; +} + +static __host__ int ggml_cuda_mmq_get_J(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).J; +} + +static constexpr __device__ int ggml_cuda_mmq_get_J(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).J; +} + +static __host__ ggml_cuda_mmq_sram_layout ggml_cuda_mmq_get_sram_layout(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).sram_layout; +} + +static constexpr __device__ ggml_cuda_mmq_sram_layout ggml_cuda_mmq_get_sram_layout(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).sram_layout; +} + +static __host__ int ggml_cuda_mmq_get_K_vram(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).K_vram; +} + +static constexpr __device__ int ggml_cuda_mmq_get_K_vram(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).K_vram; +} + +static __host__ bool ggml_cuda_mmq_get_stream_k(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).stream_k; +} + +static constexpr __device__ bool ggml_cuda_mmq_get_stream_k(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).stream_k; +} + +static __host__ int ggml_cuda_mmq_get_fallback(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).fallback; +} + +static constexpr __device__ int ggml_cuda_mmq_get_fallback(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).fallback; +} + +// --------------------------------------------------------------------------------------------- + +static __host__ int ggml_cuda_mmq_get_sram_stride(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_sram_stride(ggml_cuda_mmq_get_sram_layout(type, J, fallback, cc)); +} + +static constexpr __device__ int ggml_cuda_mmq_get_sram_stride(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_sram_stride(ggml_cuda_mmq_get_sram_layout(type, J, fallback)); +} + +static __host__ int ggml_cuda_mmq_get_J_max(const ggml_type type, const bool fallback, const int cc, const int64_t ne11) { + int ret = std::min(ne11, int64_t(512)); + ret -= ret % 8; + for (;ret > 0; ret -= 8) { + if (ggml_cuda_mmq_get_config(type, ret, fallback, cc).type != GGML_TYPE_COUNT) { + return ret; + } + } + return ret; +} + +static constexpr __device__ int ggml_cuda_mmq_get_rows_per_warp(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).rows_per_warp(); +} + +#define MMQ_DP4A_TXS_Q4_0 tile_x_sizes{I*MMQ_TILE_NE_K + I, I*MMQ_TILE_NE_K/QI4_0 + I/QI4_0, 0} +#define MMQ_DP4A_TXS_Q4_1 tile_x_sizes{I*MMQ_TILE_NE_K + I, I*MMQ_TILE_NE_K/QI4_1 + I/QI4_1, 0} +#define MMQ_DP4A_TXS_Q8_0 tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K*2/QI8_0 + I/(QI8_0/2), 0} +#define MMQ_DP4A_TXS_Q8_0_16 tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K*4/QI8_0 + I/(QI8_0/4), 0} +#define MMQ_DP4A_TXS_Q8_1 tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K*2/QI8_1 + I/(QI8_1/2), 0} +#define MMQ_DP4A_TXS_Q2_K tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K + I, 0} +#define MMQ_DP4A_TXS_Q3_K tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I, I*MMQ_TILE_NE_K/8 + I/8} +#define MMQ_DP4A_TXS_Q4_K tile_x_sizes{I*MMQ_TILE_NE_K + I, I*MMQ_TILE_NE_K/QI4_K, I*MMQ_TILE_NE_K/8 + I/8} +#define MMQ_DP4A_TXS_Q5_K tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K/QI5_K + I/QI5_K, I*MMQ_TILE_NE_K/8 + I/8} +#define MMQ_DP4A_TXS_Q6_K tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K/QI6_K + I/QI6_K, I*MMQ_TILE_NE_K/8 + I/8} + +static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int I) { switch (type) { case GGML_TYPE_Q1_0: return MMQ_DP4A_TXS_Q8_0; case GGML_TYPE_Q4_0: return MMQ_DP4A_TXS_Q4_0; @@ -216,2982 +394,29 @@ static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml } } -#define MMQ_MMA_TILE_X_K_Q8_0 (2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0 + 4) -#define MMQ_MMA_TILE_X_K_FP4 (2*MMQ_TILE_NE_K + 8 + 4) // MXFP4 and NVFP4 Blackwell -#define MMQ_MMA_TILE_X_K_NVFP4 (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2 + 4) // NVFP4 Generic -#define MMQ_MMA_TILE_X_K_Q8_1 (2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0 + 4) -#define MMQ_MMA_TILE_X_K_Q2_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K + 4) -#define MMQ_MMA_TILE_X_K_Q3_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2 + 4) -#define MMQ_MMA_TILE_X_K_Q6_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/QI6_K + MMQ_TILE_NE_K/8 + 7) - -static_assert(MMQ_MMA_TILE_X_K_Q8_0 % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_Q8_1 % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_Q2_K % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_Q3_K % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_Q6_K % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_FP4 % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_FP4 == MMQ_MMA_TILE_X_K_Q8_1, "Wrong tile size for MXFP4"); -static_assert(MMQ_MMA_TILE_X_K_NVFP4 % 8 == 4, "Wrong padding."); - - -static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) { - switch (type) { - case GGML_TYPE_Q1_0: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_Q4_0: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_Q4_1: return MMQ_MMA_TILE_X_K_Q8_1; - case GGML_TYPE_Q5_0: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_Q5_1: return MMQ_MMA_TILE_X_K_Q8_1; - case GGML_TYPE_Q8_0: return MMQ_MMA_TILE_X_K_Q8_0; - // tile sizes are the same for Q8_1 and FP4 for blackwell - case GGML_TYPE_MXFP4: return MMQ_MMA_TILE_X_K_Q8_1; -#if defined(BLACKWELL_MMA_AVAILABLE) - case GGML_TYPE_NVFP4: return MMQ_MMA_TILE_X_K_FP4; -#else - case GGML_TYPE_NVFP4: return MMQ_MMA_TILE_X_K_NVFP4; -#endif // defined(BLACKWELL_MMA_AVAILABLE) - case GGML_TYPE_Q2_K: return MMQ_MMA_TILE_X_K_Q2_K; - case GGML_TYPE_Q3_K: return MMQ_MMA_TILE_X_K_Q3_K; - case GGML_TYPE_Q4_K: return MMQ_MMA_TILE_X_K_Q8_1; - case GGML_TYPE_Q5_K: return MMQ_MMA_TILE_X_K_Q8_1; - case GGML_TYPE_Q6_K: return MMQ_MMA_TILE_X_K_Q6_K; - case GGML_TYPE_IQ2_XXS: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_IQ2_XS: return MMQ_MMA_TILE_X_K_Q3_K; - case GGML_TYPE_IQ2_S: return MMQ_MMA_TILE_X_K_Q3_K; - case GGML_TYPE_IQ3_XXS: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_IQ3_S: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_IQ1_S: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_IQ4_XS: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_IQ4_NL: return MMQ_MMA_TILE_X_K_Q8_0; - default: return 0; +// FIXME temporary until all combinations of data types and GPUs can use the MMA data layout +static __host__ int ggml_cuda_mmq_get_nbytes_shared_x(const ggml_cuda_mmq_config & config, const int cc) { + if (config.use_mma_data_layout(cc)) { + return config.I * ggml_cuda_mmq_get_sram_stride(config.sram_layout) * 4; } -} - -// block_q8_1_mmq has (128 8-bit ints == 32 32-bit ints + 4 32-bit scales) -#define MMQ_TILE_Y_K (MMQ_TILE_NE_K + MMQ_TILE_NE_K / QI8_1) -#define MMQ_TILE_Y_FP4_K MMQ_TILE_Y_K - -static int mmq_get_granularity_host(const int mmq_x, const int cc) { - if (amd_mfma_available(cc) || amd_wmma_available(cc)) { - return mmq_x >= 128 ? 32 : 16; - } else if (turing_mma_available(cc) && mmq_x >= 48) { - return 16; - } else { - return 8; - } -} - -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) { - return mmq_x >= 128 ? 32 : 16; -} -#elif defined(TURING_MMA_AVAILABLE) -static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) { - return mmq_x >= 48 ? 16 : 8; -} -#else -static constexpr __device__ int mmq_get_granularity_device(const int /*mmq_x*/) { - return 8; -} -#endif // AMD_MFMA_AVAILABLE - -#if defined(GGML_USE_HIP) -static int mmq_get_nwarps_host(const int cc, const int warp_size) { - return amd_mfma_available(cc) ? 8 : 256/warp_size; -} -#else -static int mmq_get_nwarps_host(const int /*cc*/, const int warp_size) { - return 256/warp_size; -} -#endif // (GGML_USE_HIP) - -static constexpr __device__ int mmq_get_nwarps_device() { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - return 8; -#else - return 256/ggml_cuda_get_physical_warp_size(); -#endif // AMD_MFMA_AVAILABLE + const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(config.type, config.I); + return (txs.qs + txs.dm + txs.sc) * 4; } // ------------------------------------------------------------ -template static __device__ __forceinline__ void load_tiles_q1_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); +#include "mmq-load-tiles.cuh" +#include "mmq-vec-dot.cuh" -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int blocks_per_iter = MMQ_ITER_K / QK1_0; - constexpr int threads_per_row = blocks_per_iter * QI1_0; - constexpr int nrows = warp_size / threads_per_row; - constexpr int scale_entries_per_block = QK1_0 / QK8_1; - constexpr int scale_entries_per_row = blocks_per_iter * scale_entries_per_block; - - const int txi = threadIdx.x % threads_per_row; - const int kbx = txi / QI1_0; - const int kqsx = txi % QI1_0; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + kbx; - const int qs_offset = 4*kqsx; - const int qs0 = bxi->qs[qs_offset + 0] | (bxi->qs[qs_offset + 1] << 8) | - (bxi->qs[qs_offset + 2] << 16) | (bxi->qs[qs_offset + 3] << 24); - - int unpacked_bytes[8]; -#pragma unroll - for (int j = 0; j < 8; ++j) { - const int shift = j * 4; - const int bits4 = (qs0 >> shift) & 0x0F; - const int b0 = (bits4 & 0x01) ? 1 : -1; - const int b1 = (bits4 & 0x02) ? 1 : -1; - const int b2 = (bits4 & 0x04) ? 1 : -1; - const int b3 = (bits4 & 0x08) ? 1 : -1; - unpacked_bytes[j] = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24); - } - - const int dst_offset = kbx*(scale_entries_per_block*QI8_0) + kqsx*QI8_0; -#pragma unroll - for (int j = 0; j < 8; ++j) { -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + dst_offset + j] = unpacked_bytes[j]; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j] = unpacked_bytes[j]; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - } - - const int ksx = threadIdx.x % scale_entries_per_row; - const int scale_block = ksx / scale_entries_per_block; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + threadIdx.y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + scale_block; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + ksx] = bxi->d; -#else - x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + ksx] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_q4_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_0); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI4_0; - const int kqsx = txi % QI4_0; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx; - const int qs0 = get_int_b2(bxi->qs, kqsx); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + 0] = __vsubss4((qs0 >> 0) & 0x0F0F0F0F, 0x08080808); - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + QI4_0] = __vsubss4((qs0 >> 4) & 0x0F0F0F0F, 0x08080808); -#else - x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_0; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = bxi->d; -#else - x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + kbxd] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y); - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_0*VDR_Q4_0_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2); - - int u[2*VDR_Q4_0_Q8_1_MMQ]; - - constexpr int max_cpy = ggml_cuda_get_max_cpy_bytes(); - constexpr int mcpy_int = max_cpy / sizeof(int); - static_assert(VDR_Q4_0_Q8_1_MMQ == 4, "bad VDR_Q4_0_Q8_1_MMQ"); - - int tmp0[4], tmp1[4]; - - #pragma unroll - for (int l0 = 0; l0 < 4 / mcpy_int; ++l0) { - ggml_cuda_memcpy_1(tmp0 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + l0 * mcpy_int] ); - ggml_cuda_memcpy_1(tmp1 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + QI4_0 + l0 * mcpy_int]); - } - - u[0]=tmp0[0]; u[2]=tmp0[1]; u[4]=tmp0[2]; u[6]=tmp0[3]; - u[1]=tmp1[0]; u[3]=tmp1[1]; u[5]=tmp1[2]; u[7]=tmp1[3]; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q4_0_q8_1_impl - (&x_qs[i*(MMQ_TILE_NE_K + 1) + k0/QR4_0], u, - x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + k0/(QR4_0*QI4_0)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template static __device__ __forceinline__ void load_tiles_q4_1( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_1); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI4_1; - const int kqsx = txi % QI4_1; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx; - const int qs0 = get_int_b4(bxi->qs, kqsx); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + 0] = (qs0 >> 0) & 0x0F0F0F0F; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + QI4_1] = (qs0 >> 4) & 0x0F0F0F0F; -#else - x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_1; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + kbxd] = bxi->dm; -#else - x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + kbxd] = bxi->dm; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y); - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_1*VDR_Q4_1_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2); - - int u[2*VDR_Q4_1_Q8_1_MMQ]; - - constexpr int max_cpy = ggml_cuda_get_max_cpy_bytes(); - constexpr int mcpy_int = max_cpy / sizeof(int); - static_assert(VDR_Q4_0_Q8_1_MMQ == 4, "bad VDR_Q4_0_Q8_1_MMQ"); - - int tmp0[4], tmp1[4]; - - #pragma unroll - for (int l0 = 0; l0 < 4 / mcpy_int; ++l0) { - ggml_cuda_memcpy_1(tmp0 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + l0 * mcpy_int] ); - ggml_cuda_memcpy_1(tmp1 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + QI4_1 + l0 * mcpy_int]); - } - - u[0]=tmp0[0]; u[2]=tmp0[1]; u[4]=tmp0[2]; u[6]=tmp0[3]; - u[1]=tmp1[0]; u[3]=tmp1[1]; u[5]=tmp1[2]; u[7]=tmp1[3]; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q4_1_q8_1_impl - (&x_qs[i*(MMQ_TILE_NE_K + 1) + k0/QR4_1], u, - x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + k0/(QR4_1*QI4_1)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template static __device__ __forceinline__ void load_tiles_q5_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_0); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI5_0; - const int kqsx = txi % QI5_0; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbx; - - const int ql = get_int_b2(bxi->qs, kqsx); - const int qh = get_int_b2(bxi->qh, 0) >> (4 * kqsx); - - int qs0 = (ql >> 0) & 0x0F0F0F0F; - qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 - qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 - qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 - qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 - qs0 = __vsubss4(qs0, 0x10101010); // subtract 16 - - int qs1 = (ql >> 4) & 0x0F0F0F0F; - qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 - qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 - qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 - qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 - qs1 = __vsubss4(qs1, 0x10101010); // subtract 16 - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + 0] = qs0; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + 0] = qs0; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_0; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = bxi->d; -#else - x_df[i*(MMQ_TILE_NE_K/QI5_0) + i/QI5_0 + kbxd] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_q5_1( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_1); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI5_1; - const int kqsx = txi % QI5_1; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbx; - - const int ql = get_int_b4(bxi->qs, kqsx); - const int qh = get_int_b4(bxi->qh, 0) >> (4 * kqsx); - - int qs0 = (ql >> 0) & 0x0F0F0F0F; - qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 - qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 - qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 - qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 - - int qs1 = (ql >> 4) & 0x0F0F0F0F; - qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 - qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 - qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 - qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + 0] = qs0; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + 0] = qs0; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_1; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + kbxd] = bxi->dm; -#else - x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + kbxd] = bxi->dm; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_q8_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_tile + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - // MMQ_ITER_K / (4 * QR8_0) == 64 required. but NV has only 32 threads per warp - constexpr int threads_per_row = 32; - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI8_0; - const int kqsx = txi % QI8_0; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 0 + txi] = get_int_b2(bxi[0].qs, kqsx); - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx); -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 0 + txi] = get_int_b2(bxi[0].qs, kqsx); - x_qs[i*(2*MMQ_TILE_NE_K + 1) + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = 2*MMQ_TILE_NE_K / QI8_0; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = bxi->d; -#else - x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + kbxd] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_mxfp4( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_MXFP4, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR_MXFP4); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI_MXFP4; - const int kqsx = txi % QI_MXFP4; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbx; - - const int aux_q4 = get_int_b1(bxi->qs, kqsx); - const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4); - const int k0 = kbx * (2 * QI_MXFP4) + kqsx; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + 0] = v.x; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + QI_MXFP4] = v.y; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI_MXFP4] = v.y; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI_MXFP4; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_1 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f; -#else - x_df[i*(MMQ_TILE_NE_K/QI_MXFP4) + i/QI_MXFP4 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void load_tiles_mxfp4_fp4(const char * __restrict__ x, - int * __restrict__ x_tile, - const int kbx0, - const int i_max, - const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - int * x_qs = (int *) x_tile; - uint32_t * x_sc = (uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K); - - const int txi = threadIdx.x; - - constexpr int iter_k = get_iter_k(GGML_TYPE_MXFP4); - - constexpr int threads_per_row = iter_k / QK_MXFP4; // each thread processes 1 block - constexpr int rows_per_warp = warp_size / threads_per_row; - const int kbx = txi % threads_per_row; - const int row_in_warp = txi / threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += rows_per_warp * nwarps) { - int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; - - if constexpr (need_check) { - i = min(i, i_max); - } - - const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i * stride + kbx; - - // quantize_mxfp4_mmq permutes nibbles to match the quantized format - const int k0 = kbx * 4; - memcpy(x_qs + i * MMQ_MMA_TILE_X_K_FP4 + k0, bxi->qs, 16); - - // Load E8M0 scales: pack 2 consecutive scales into one uint32 - if (kbx % 2 == 0) { - uint32_t e = bxi->e; - e |= ((bxi + 1)->e << 8); - x_sc[i * MMQ_MMA_TILE_X_K_FP4 + kbx / 2] = e; - } - } -} - -#ifdef BLACKWELL_MMA_AVAILABLE -template -static __device__ __forceinline__ void load_tiles_nvfp4_nvfp4(const char * __restrict__ x, - int * __restrict__ x_tile, - const int kbx0, - const int i_max, - const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - constexpr int iter_k = get_iter_k(GGML_TYPE_NVFP4); - constexpr int threads_per_row = iter_k / QK_NVFP4; // each thread processes 1 block - constexpr int rows_per_warp = warp_size / threads_per_row; - - uint32_t * x_u32 = (uint32_t *) x_tile; - - const int txi = threadIdx.x; - const int kbx = txi % threads_per_row; - const int row_in_warp = txi / threads_per_row; - - const block_nvfp4 * bxi_base = (const block_nvfp4 *) x + kbx0 + kbx; - uint32_t * x_u32_scale = x_u32 + 64 + kbx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += rows_per_warp * nwarps) { - int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; - - if constexpr (need_check) { - i = min(i, i_max); - } - - const block_nvfp4 * bxi = bxi_base + i * stride; - const int row_base = i * MMQ_MMA_TILE_X_K_FP4; - const int q_base = row_base + 8 * kbx; - - const uint32_t * src_qs = reinterpret_cast(bxi->qs); - -#pragma unroll - for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) { - x_u32[q_base + 2 * sub + 0] = src_qs[2 * sub + 0]; - x_u32[q_base + 2 * sub + 1] = src_qs[2 * sub + 1]; - } - - x_u32_scale[row_base] = get_int_b4(bxi->d, 0); - } -} - -// Shared MMA kernel for MXFP4 and NVFP4 on Blackwell. -// Both quantizations encode values as e2m1 (FP4) and produce one uint32 scale per -// m16n8k64 MMA call; only the PTX kind (scale_vec::2X ue8m0 vs scale_vec::4X ue4m3) -// and the per-type stride constant differ. -template -static __device__ __forceinline__ void vec_dot_fp4_fp4_mma(const int * __restrict__ x, - const int * __restrict__ y, - float * __restrict__ sum, - const int k00) { - static_assert(type == GGML_TYPE_MXFP4 || type == GGML_TYPE_NVFP4, - "vec_dot_fp4_fp4_mma: type must be MXFP4 or NVFP4"); - - typedef tile<16, 8, int> tile_A; - typedef tile<8, 8, int> tile_B; - typedef tile<16, 8, float> tile_C; - - constexpr int stride = MMQ_MMA_TILE_X_K_FP4; - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp / tile_C::I; - constexpr int nfrags = MMQ_TILE_NE_K / tile_A::J; - - y += (threadIdx.y % ntx) * (tile_C::J * MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const uint32_t * x_sc = (const uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K); - const int * y_qs = (const int *) y + 4; - const uint32_t * y_sc = (const uint32_t *) y; - - // 2 threads per quad supply the packed scale register to the block_scale MMA, - // see https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-block-scaling - const int tidx_A = threadIdx.x / 4 + (threadIdx.x % 2) * 8; - const int tidx_B = threadIdx.x / 4; - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - tile_A A[ntx][nfrags]; - uint32_t scaleA[ntx][nfrags]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int frag = 0; frag < nfrags; ++frag) { - const int k0 = k00 + frag * tile_A::J; - load_ldmatrix(A[n][frag], x_qs + (i0 + n * tile_A::I) * stride + k0, stride); - scaleA[n][frag] = x_sc[(i0 + n * tile_A::I + tidx_A) * stride + k0 / tile_A::J]; - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx * tile_C::J) { - tile_B B[nfrags]; - uint32_t scaleB[nfrags]; - -#pragma unroll - for (int frag = 0; frag < nfrags; ++frag) { - const int k0 = frag * tile_B::J; - load_generic(B[frag], y_qs + j0 * MMQ_TILE_Y_K + k0, MMQ_TILE_Y_K); - scaleB[frag] = y_sc[(j0 + tidx_B) * MMQ_TILE_Y_K + frag]; - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int frag = 0; frag < nfrags; ++frag) { - tile_C C = {}; - mma_block_scaled_fp4(C, A[n][frag], B[frag], scaleA[n][frag], scaleB[frag]); -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0 / tile_C::J + n) * tile_C::ne + l] += C.x[l]; - } - } - } - } -} -#endif // BLACKWELL_MMA_AVAILABLE - - -template -static __device__ __forceinline__ void load_tiles_nvfp4(const char * __restrict__ x, - int * __restrict__ x_tile, - const int kb0, - const int i_max, - const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_NVFP4, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / QK_NVFP4; - constexpr int rows_per_warp = warp_size / threads_per_row; - const int kbx = threadIdx.x % threads_per_row; - const int row_in_warp = threadIdx.x / threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += rows_per_warp * nwarps) { - int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; - - if constexpr (need_check) { - i = min(i, i_max); - } - - const block_nvfp4 * bxi = (const block_nvfp4 *) x + kb0 + i * stride + kbx; - const uint32_t * __restrict__ src_qs = reinterpret_cast(bxi->qs); - const int kqs = 16 * kbx; - const int ksc = 4 * kbx; - -#pragma unroll - for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) { - const int2 q0 = get_int_from_table_16(src_qs[2 * sub + 0], kvalues_mxfp4); - const int2 q1 = get_int_from_table_16(src_qs[2 * sub + 1], kvalues_mxfp4); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 0] = q0.x; - x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 1] = q1.x; - x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 2] = q0.y; - x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 3] = q1.y; - x_df[i * MMQ_MMA_TILE_X_K_NVFP4 + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]); -#else - x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 0] = q0.x; - x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 1] = q1.x; - x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 2] = q0.y; - x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 3] = q1.y; - x_df[i * (2 * MMQ_TILE_NE_K * 2 / QI_NVFP4) + i / (QK_NVFP4_SUB / QI_NVFP4) + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - } -} - -template -static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y); - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += VDR_Q8_0_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q8_0_q8_1_impl - (&x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k0 % MMQ_TILE_NE_K], - x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + k0/QI8_0], y_df[j*MMQ_TILE_Y_K + (k0/QI8_1) % (MMQ_TILE_NE_K/QI8_1)]); - } - } - } -} - -template -static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr data_layout input_layout = get_input_data_layout(); - typedef tile<16, 8, int, input_layout> tile_A; - typedef tile<16, 8, int, input_layout> tile_B; - typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + 2*MMQ_TILE_NE_K; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - const half2 * y_ds = (const half2 *) y; - - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { - const int k0 = k00 + k01; - - tile_A A[ntx]; -#pragma unroll - for (int n = 0; n < ntx; ++n) { - load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0); - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - tile_B B; - load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); - - float dB; - const int j = j0 + tile_C::get_j(0); - if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) { - dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - } else { - dB = __low2float(y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - const int i = i0 + n*tile_A::I + tile_C::get_i(l); - const float dA = x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + k0/QI8_0]; - sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l]*dA*dB; - } - } - } - } -#else - typedef tile<16, 8, int> tile_A; - typedef tile< 8, 8, int> tile_B; - typedef tile<16, 8, int> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + 2*MMQ_TILE_NE_K; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - const half2 * y_ds = (const half2 *) y; - - tile_A A[ntx][MMQ_TILE_NE_K/QI8_0]; - float dA[ntx][tile_C::ne/2][MMQ_TILE_NE_K/QI8_0]; - - const int i0 = (threadIdx.y/ntx)*rows_per_warp; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { - const int k0 = k00 + k01; - - load_ldmatrix(A[n][k01/QI8_0], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0); - } - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_A::I + tile_C::get_i(2*l); - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { - const int k0 = k00 + k01; - - dA[n][l][k01/QI8_0] = x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + k0/QI8_0]; - } - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { - tile_B B; - float dB[tile_C::ne/2]; - - load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) { - dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - } else { - dB[l] = __low2float(y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n][k01/QI8_0], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l]*dA[n][l/2][k01/QI8_0]*dB[l%2]; - } - } - } - } -#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -} - - -template -static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y); - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += VDR_Q8_0_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q8_1_q8_1_impl - (&x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], - x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + k0/QI8_1], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template -static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr data_layout input_layout = get_input_data_layout(); - typedef tile<16, 8, int, input_layout> tile_A; - typedef tile<16, 8, int, input_layout> tile_B; - typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + 2*MMQ_TILE_NE_K; - const int * y_qs = (const int *) y + 4; - const half2 * y_dm = (const half2 *) y; - - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - const int k0 = k00 + k01; - - tile_A A[ntx]; -#pragma unroll - for (int n = 0; n < ntx; ++n) { - load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1); - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - tile_B B; - load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); - - const int j = j0 + tile_C::get_j(0); - const float2 dsB = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]); - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - const int i = i0 + n*tile_A::I + tile_C::get_i(l); - float2 dmA = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + k0/QI8_1]); - sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA.x*dsB.x*C.x[l]; - sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA.y*dsB.y; - } - } - } - } -#else - typedef tile<16, 8, int> tile_A; - typedef tile< 8, 8, int> tile_B; - typedef tile<16, 8, int> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + 2*MMQ_TILE_NE_K; - const int * y_qs = (const int *) y + 4; - const half2 * y_dm = (const half2 *) y; - - tile_A A[ntx][MMQ_TILE_NE_K/QI8_1]; - float2 dmA[ntx][tile_C::ne/2][MMQ_TILE_NE_K/QI8_1]; - - const int i0 = (threadIdx.y/ntx)*rows_per_warp; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - const int k0 = k00 + k01; - - load_ldmatrix(A[n][k01/QI8_1], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1); - } - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_A::I + tile_C::get_i(2*l); - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - const int k0 = k00 + k01; - - dmA[n][l][k01/QI8_1] = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + k0/QI8_1]); - } - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - tile_B B; - float2 dsB[tile_C::ne/2]; - - load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - dsB[l] = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n][k01/QI8_1], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA[n][l/2][k01/QI8_1].x*dsB[l%2].x*C.x[l]; - sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA[n][l/2][k01/QI8_1].y*dsB[l%2].y; - } - } - } - } -#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -} - -// Used for NVFP4, Q3_K, IQ2_S, and IQ2_XS -template -static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16; - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q8_0_16_q8_1_impl( - &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], - &y_qs[j*MMQ_TILE_Y_K + k01], - &x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + k0/(QI8_0/2)], - y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -// Used for Q3_K, IQ2_S, and IQ2_XS: -template -static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr data_layout input_layout = get_input_data_layout(); - typedef tile<16, 4, int, input_layout> tile_A; - typedef tile<16, 4, int, input_layout> tile_B; - typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { - const int k0 = k00 + k01; - - tile_A A[ntx]; -#pragma unroll - for (int n = 0; n < ntx; ++n) { - load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K); - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - tile_B B; - load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); - - const int j = j0 + tile_C::get_j(0); - const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(l); - sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4] * dB; - } - } - } - } -#elif defined(TURING_MMA_AVAILABLE) - - typedef tile<16, 4, int> tile_A; - typedef tile<16, 8, int> tile_A_8; - typedef tile< 8, 4, int> tile_B; - typedef tile<16, 8, int> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - - const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); - - tile_A A[ntx][8]; - float dA[ntx][tile_C::ne/2][8]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { - const int k0 = k00 + k01; - - load_ldmatrix(((tile_A_8 *) A[n])[k01/8], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K); - } - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { - const int k0 = k00 + k01; - - dA[n][l][k01/4] = x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4]; - } - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) { - tile_B B[2]; - float dB[tile_C::ne/2]; - - // Here load_generic is faster than load_ldmatrix. - load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); - load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + (k01 + tile_B::J), MMQ_TILE_Y_K); - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C[2]; - mma(C[0], A[n][k01/4 + 0], B[0]); - mma(C[1], A[n][k01/4 + 1], B[1]); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0/tile_C::J + n)*tile_C::ne + l] += dB[l%2]*(C[0].x[l]*dA[n][l/2][k01/4 + 0] + C[1].x[l]*dA[n][l/2][k01/4 + 1]); - } - } - } - } -#else - GGML_UNUSED_VARS(x, y, sum, k00); - NO_DEVICE_CODE; -#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE -} - -template static __device__ __forceinline__ void load_tiles_q2_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR2_K); - constexpr int nrows = ggml_cuda_get_physical_warp_size() / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q2_K * bxi = (const block_q2_K *) x + kbx0 + i*stride; - - const int x_ql_0 = get_int_b2(bxi->qs, kqsx); - -#pragma unroll - for (int l = 0; l < QR2_K; ++l) { - const int k = (kqsx/8)*32 + l*8 + kqsx % 8; - - const int x_qs_k = (x_ql_0 >> (2*l)) & 0x03030303; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q2_K + k] = x_qs_k; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int sc_m = bxi->scales[kqsx]; -#ifdef FAST_FP16_AVAILABLE - const half2 x_dm_ik = __hmul2(bxi->dm, make_half2(sc_m & 0x0F, sc_m >> 4)); -#else - const float2 bxi_dmf = __half22float2(bxi->dm); - const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4)); -#endif // FAST_FP16_AVAILABLE - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + kqsx] = x_dm_ik; -#else - x_dm[i*(MMQ_TILE_NE_K + 1) + kqsx] = x_dm_ik; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y); - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - - float2 y_df[mmq_x/nwarps]; -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - - y_df[j0/nwarps] = __half22float2(y_ds[j*MMQ_TILE_Y_K]); - } - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K/2; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - constexpr int ns = 2; - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q2_K_q8_1_impl_mmq( - &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], - &x_dm[i*(MMQ_TILE_NE_K + 1) + k0/4], k01 < MMQ_TILE_NE_K/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, - &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); - } - } - } - - // Some compilers fail to unroll the loop over k01 if there is a conditional statement for ns in the inner loop. - // As a workaround 2 separate loops are used instead. -#pragma unroll - for (int k01 = MMQ_TILE_NE_K/2; k01 < MMQ_TILE_NE_K; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - constexpr int ns = 1; - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q2_K_q8_1_impl_mmq( - &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], - &x_dm[i*(MMQ_TILE_NE_K + 1) + k0/4], k01 < MMQ_TILE_NE_K/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, - &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); - } - } - } -} - -template -static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr data_layout input_layout = get_input_data_layout(); - typedef tile<16, 4, int, input_layout> tile_A; - typedef tile<16, 4, int, input_layout> tile_B; - typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { - const int k0 = k00 + k01; - - tile_A A[ntx]; -#pragma unroll - for (int n = 0; n < ntx; ++n) { - load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K); - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - tile_B B; - load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); - - const int j = j0 + tile_C::get_j(0); - const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y; - const float sB = (k01 >= MMQ_TILE_NE_K * 3/4) ? 0 - : (((k01/4)%2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).y - : __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).x); - - tile_C Cm; - if (k01 >= MMQ_TILE_NE_K * 3/4) { - tile_A A1; -#pragma unroll - for (int l = 0; l < tile_A::ne; ++l) { - A1.x[l] = 0x01010101; - } - mma(Cm, A1, B); - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C Cd; - mma(Cd, A[n], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(l); - const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/4]); - float tmp = Cd.x[l]*dm.x; - if (k01 >= MMQ_TILE_NE_K * 3/4) { - tmp -= Cm.x[l]*dm.y; - } - sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*dB; - sum[(j0/tile_C::J + n)*tile_C::ne + l] -= dm.y*sB; - } - } - } - } -#elif defined(TURING_MMA_AVAILABLE) - - typedef tile<16, 4, int> tile_A; - typedef tile<16, 8, int> tile_A_8; - typedef tile< 8, 4, int> tile_B; - typedef tile<16, 8, int> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - - const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); - - tile_A A[ntx][8]; - float dA[ntx][tile_C::ne/2][8]; - float mA[ntx][tile_C::ne/2][8]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - const int k0 = k00 + k01; - - load_ldmatrix(((tile_A_8 *) A[n])[k01/QI8_1], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K); - } - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1/2) { - const int k0 = k00 + k01; - - const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/(QI8_1/2)]); - - dA[n][l][k01/(QI8_1/2)] = dm.x; - mA[n][l][k01/(QI8_1/2)] = dm.y; - } - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - float2 dB[tile_C::ne/2]; - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - dB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K]); - } - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - tile_B B[2]; - - // Here load_generic is faster than load_ldmatrix. - load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); - load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + (k01 + tile_B::J), MMQ_TILE_Y_K); - - tile_C Cm[2]; - if (k01 >= MMQ_TILE_NE_K * 3/4) { - tile_A A1; - A1.x[0] = 0x01010101; - A1.x[1] = 0x01010101; - mma(Cm[0], A1, B[0]); - mma(Cm[1], A1, B[1]); - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C Cd[2]; - - mma(Cd[0], A[n][k01/4 + 0], B[0]); - mma(Cd[1], A[n][k01/4 + 1], B[1]); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - float tmp = Cd[0].x[l]*dA[n][l/2][k01/4 + 0] + Cd[1].x[l]*dA[n][l/2][k01/4 + 1]; - if (k01 >= MMQ_TILE_NE_K * 3/4) { - tmp -= Cm[0].x[l]*mA[n][l/2][k01/4 + 0] + Cm[1].x[l]*mA[n][l/2][k01/4 + 1]; - } - sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*(k01 < MMQ_TILE_NE_K/2 ? dB[l%2].x : dB[l%2].y); - } - } - } - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K * 3/4; k01 += QI8_1) { - float2 sB[tile_C::ne/2]; - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - sB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0/tile_C::J + n)*tile_C::ne + l] -= mA[n][l/2][k01/4 + 0]*sB[l%2].x; - sum[(j0/tile_C::J + n)*tile_C::ne + l] -= mA[n][l/2][k01/4 + 1]*sB[l%2].y; - } - } - } - } -#else - GGML_UNUSED_VARS(x, y, sum, k00); - NO_DEVICE_CODE; -#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE -} - -template static __device__ __forceinline__ void load_tiles_q3_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); - int * x_sc = (int *) (x_df + txs.dm); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR3_K); - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; - - const int x_ql_0 = get_int_b2(bxi->qs, kqsx); - const int x_qh_0 = get_int_b2(bxi->hmask, kqsx % (QI3_K/2)) >> (4 * (kqsx / (QI3_K/2))); - -#pragma unroll - for (int l = 0; l < QR3_K; ++l) { - const int k = (kqsx/8)*32 + l*8 + kqsx % 8; - - const int x_ql_k = (x_ql_0 >> (2*l)) & 0x03030303; - const int x_qh_k = ((x_qh_0 >> l) << 2) & 0x04040404; - - const int x_qs_k = __vsubss4(x_ql_k | x_qh_k, 0x04040404); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k] = x_qs_k; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - } - - constexpr int rows_per_warp = warp_size / 4; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { - int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/4; - - if (need_check) { - i = min(i, i_max); - } - - const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; - - const int ksc = threadIdx.x % 4; - - const int ksc_low = ksc % (QI3_K/8); - const int shift_low = 4 * (ksc / (QI3_K/8)); - const int sc_low = (get_int_b2(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; - - const int ksc_high = QI3_K/8; - const int shift_high = 2 * ksc; - const int sc_high = ((get_int_b2(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; - - const int sc = __vsubss4(sc_low | sc_high, 0x20202020); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - const int8_t * sc8 = (const int8_t *) ≻ - const float d = bxi->d; - -#pragma unroll - for (int l = 0; l < int(sizeof(int)); ++l) { - x_df[i*MMQ_MMA_TILE_X_K_Q3_K + sizeof(int)*ksc + l] = d*sc8[l]; - } -#else - x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = sc; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - -#if !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)) -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) { - int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; - - x_df[i] = bxi->d; - } -#endif // !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)) || defined(AMD_WMMA_AVAILABLE) -} - -template -static __device__ __forceinline__ void vec_dot_q3_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y); - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + txs.qs; - const int * x_sc = (const int *) x_df + txs.dm; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - const int8_t * scales = ((const int8_t *) (x_sc + i*(MMQ_TILE_NE_K/8) + i/8)) + k0/4; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q3_K_q8_1_impl_mmq( - &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], scales, - x_df[i], y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, const int ksc) { - // scale arrangement after the following two lines: - // - ksc == 0: sc0, sc1, sc2, sc3 - // - ksc == 1: sc4, sc5, sc6, sc7 - // - ksc == 2: m0, m1, m2, m3 - // - ksc == 3: m4, m5, m6, m7 - return ((scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F) | // lower 4 bits - ((scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030); // upper 2 bits -} - -template static __device__ __forceinline__ void load_tiles_q4_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + txs.qs); - int * x_sc = (int *) (x_dm + txs.dm); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_K); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; - const int qs0 = get_int_b4(bxi->qs, txi); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(txi/8) + txi % 8 + 0] = (qs0 >> 0) & 0x0F0F0F0F; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(txi/8) + txi % 8 + 8] = (qs0 >> 4) & 0x0F0F0F0F; -#else - x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr int rows_per_warp = warp_size / 2; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - // Need if on AMD instead of % because warp_size == 64 - // This causes double work and throughput loss (MI300X) - // H100 loses about 100 t/s with 'if' condition over '%' - int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/2; - if (i < mmq_y) { -#else - int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % mmq_y; - { -#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; - - const int * scales = (const int *) bxi->scales; - const int ksc = threadIdx.x % 2; - - const int sc32 = unpack_scales_q45_K(scales, ksc + 0); - const int m32 = unpack_scales_q45_K(scales, ksc + 2); - - const uint8_t * sc8 = (const uint8_t *) &sc32; - const uint8_t * m8 = (const uint8_t *) &m32; - - const half2 dm = bxi->dm * make_half2(1.0f, -1.0f); - - #pragma unroll - for (int l = 0; l < sizeof(int); ++l) { - x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]); - } - } - } -#else -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) { - int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; - - x_dm[i] = bxi->dm; - } - constexpr int rows_per_warp = warp_size / 4; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { - int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / (QI4_K/8); - - const int * scales = (const int *) bxi->scales; - - const int ksc = threadIdx.x % (MMQ_TILE_NE_K/8); - const int scales8 = unpack_scales_q45_K(scales, ksc); - - x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8; - } -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -} - -template -static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y); - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + txs.qs; - const int * x_sc = (const int *) x_dm + txs.dm; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_K*VDR_Q4_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - const uint8_t * sc = (const uint8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k0/32] + 2*(k01/16); - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q4_K_q8_1_impl_mmq( - &x_qs[i*(MMQ_TILE_NE_K + 1) + k0/2], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8, - x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template static __device__ __forceinline__ void load_tiles_q5_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + txs.qs); - int * x_sc = (int *) (x_dm + txs.dm); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_K); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; - const int ky = QR5_K*txi; - - const int ql = get_int_b4(bxi->qs, txi); - const int ql0 = (ql >> 0) & 0x0F0F0F0F; - const int ql1 = (ql >> 4) & 0x0F0F0F0F; - - const int qh = get_int_b4(bxi->qh, txi % (QI5_K/4)); - const int qh0 = ((qh >> (2 * (txi / (QI5_K/4)) + 0)) << 4) & 0x10101010; - const int qh1 = ((qh >> (2 * (txi / (QI5_K/4)) + 1)) << 4) & 0x10101010; - - const int kq0 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + 0; - const int kq1 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + QI5_K/4; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq0] = ql0 | qh0; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq1] = ql1 | qh1; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = ql0 | qh0; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = ql1 | qh1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr int rows_per_warp = warp_size / 2; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { -#if defined(AMD_MFMA_AVAILABLE) - // Need if on AMD instead of % because warp_size == 64 - // This causes double work and throughput loss (MI300X) - // H100 loses about 100 t/s with 'if' condition over '%' - int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/2; - if (i < mmq_y) { -#else - int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % mmq_y; - { -#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; - - const int * scales = (const int *) bxi->scales; - const int ksc = threadIdx.x % 2; - - const int sc32 = unpack_scales_q45_K(scales, ksc + 0); - const int m32 = unpack_scales_q45_K(scales, ksc + 2); - - const uint8_t * sc8 = (const uint8_t *) &sc32; - const uint8_t * m8 = (const uint8_t *) &m32; - - const half2 dm = bxi->dm * make_half2(1.0f, -1.0f); - -#pragma unroll - for (int l = 0; l < int(sizeof(int)); ++l) { - x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]); - } - } - } -#else -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) { - int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; - - x_dm[i] = bxi->dm; - } - - constexpr int rows_per_warp = warp_size / 4; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { - int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; - - const int * scales = (const int *) bxi->scales; - - const int ksc = threadIdx.x % (MMQ_TILE_NE_K/8); - const int scales8 = unpack_scales_q45_K(scales, ksc); - - x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8; - } -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -} - -template -static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y); - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + txs.qs; - const int * x_sc = (const int *) x_dm + txs.dm; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR5_K*VDR_Q5_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - const uint8_t * sc = ((const uint8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k00/32]) + 2*(k01/16); - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q5_K_q8_1_impl_mmq( - &x_qs[i*(QR5_K*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8, - x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template static __device__ __forceinline__ void load_tiles_q6_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); - int * x_sc = (int *) (x_df + MMQ_TILE_NE_K/QI6_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); - int * x_sc = (int *) (x_df + txs.dm); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR6_K); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride; - - const int ql = get_int_b2(bxi->ql, txi); - const int ql0 = (ql >> 0) & 0x0F0F0F0F; - const int ql1 = (ql >> 4) & 0x0F0F0F0F; - - const int qh = get_int_b2(bxi->qh, (QI6_K/4) * (txi / (QI6_K/2)) + txi % (QI6_K/4)); - const int qh0 = ((qh >> ((txi & 0x08) >> 2)) << 4) & 0x30303030; - const int qh1 = (qh >> ((txi & 0x08) >> 2)) & 0x30303030; - - const int kq0 = 2*txi - txi % (QI6_K/2) + 0; - const int kq1 = 2*txi - txi % (QI6_K/2) + QI6_K/2; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq0] = __vsubss4(ql0 | qh0, 0x20202020); - x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq1] = __vsubss4(ql1 | qh1, 0x20202020); -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020); - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) { - int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q6_K] = bxi->d; -#else - x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int rows_per_warp = warp_size / 4; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { - int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / 4; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + threadIdx.x%4] = get_int_b2(bxi->scales, threadIdx.x % (MMQ_TILE_NE_K/8)); -#else - x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + threadIdx.x%(MMQ_TILE_NE_K/8)] = get_int_b2(bxi->scales, threadIdx.x%(QI6_K/8)); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y); - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + txs.qs; - const int * x_sc = (const int *) x_df + txs.dm; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR6_K*VDR_Q6_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - const int8_t * sc = ((const int8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k0/16]); - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q6_K_q8_1_impl_mmq( - &x_qs[i*(QR6_K*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc, - x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K], &y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template -static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr data_layout input_layout = get_input_data_layout(); - typedef tile<16, 4, int, input_layout> tile_A; - typedef tile<16, 4, int, input_layout> tile_B; - typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; - const int * x_sc = (const int *) x_df + MMQ_TILE_NE_K/QI6_K; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { - const int k0 = k00 + k01; - - tile_A A[ntx]; -#pragma unroll - for (int n = 0; n < ntx; ++n) { - load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K); - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - tile_B B; - load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); - - const int j = j0 + tile_C::get_j(0); - const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(l); - const int8_t * sc = (const int8_t *) (x_sc + i*MMQ_MMA_TILE_X_K_Q6_K + k00/16); - sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * sc[k01/4] * x_df[i*MMQ_MMA_TILE_X_K_Q6_K] * dB; - } - } - } - } -#elif defined(TURING_MMA_AVAILABLE) - - typedef tile<16, 4, int> tile_A; - typedef tile< 8, 4, int> tile_B; - typedef tile<16, 8, int> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; - const int * x_sc = (const int *) x_df + MMQ_TILE_NE_K/QI6_K; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - - const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); - - tile_A A[ntx][8]; - int scA[ntx][tile_C::ne/2][8]; - float dA[ntx][tile_C::ne/2]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { - const int k0 = k00 + k01; - - load_ldmatrix(A[n][k01/4 + 0], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + 0), MMQ_MMA_TILE_X_K_Q6_K); - load_ldmatrix(A[n][k01/4 + 1], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + tile_A::J), MMQ_MMA_TILE_X_K_Q6_K); - } - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 16) { - const int k0 = k00 + k01; - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); - - const int sc_packed = x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + k0/16]; - const int8_t * sc = (const int8_t *) &sc_packed; - -#pragma unroll - for (int ksc = 0; ksc < sizeof(int); ++ksc) { - scA[n][l][k01/4 + ksc] = sc[ksc]; - } - } - } - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); - - dA[n][l] = x_df[i*MMQ_MMA_TILE_X_K_Q6_K]; - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - float tmp[ntx][tile_C::ne] = {{0.0f}}; - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { - tile_B B[2]; - float dB[tile_C::ne/2]; - - // Here load_generic is faster than load_ldmatrix. - load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + 0 + k01, MMQ_TILE_Y_K); - load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + tile_B::J + k01, MMQ_TILE_Y_K); - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C[2]; - mma(C[0], A[n][k01/4 + 0], B[0]); - mma(C[1], A[n][k01/4 + 1], B[1]); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - tmp[n][l] += (C[0].x[l]*scA[n][l/2][k01/4 + 0] + C[1].x[l]*scA[n][l/2][k01/4 + 1])*dB[l%2]; - } - } - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp[n][l]*dA[n][l/2]; - } - } - } -#else - GGML_UNUSED_VARS(x, y, sum, k00); - NO_DEVICE_CODE; -#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE -} - -template static __device__ __forceinline__ void load_tiles_iq4_nl( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_NL); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI4_NL; - const int kqsx = txi % QI4_NL; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbx; - - const int aux_q4 = get_int_b2(bxi->qs, kqsx); - const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl); - const int k0 = kbx * (2 * QI4_NL) + kqsx; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + QI4_NL] = v.y; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI4_NL] = v.y; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_NL; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = __half2float(bxi->d); -#else - x_df[i*(MMQ_TILE_NE_K/QI4_NL) + i/QI4_NL + kbxd] = __half2float(bxi->d); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq2_xxs( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XXS, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XXS)) / 2; - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq2_xxs * bxi = (const block_iq2_xxs *) x + kbx0 + i*stride; - - const int q2 = get_int_b2(bxi->qs, 2*kqsx+0); - const uint8_t * aux8 = (const uint8_t *) &q2; - const uint32_t aux32 = get_int_b2(bxi->qs, 2*kqsx+1); - -#pragma unroll - for (int l = 0; l < QR2_XXS; ++l) { - const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[l]]; - const uint32_t signs = unpack_ksigns(aux32 >> (7 * l)); - - const int signs0 = __vcmpne4(signs & 0x08040201, 0); - const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0); - - const int signs1 = __vcmpne4(signs & 0x80402010, 0); - const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid0; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid1; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid0; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int ls = aux32 >> 27 | 1; // (scale * 2 + 1) - const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4 -#else - x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4 -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq2_xs( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16; - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XS)) / 2; - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq2_xs * bxi = (const block_iq2_xs *) x + kbx0 + i*stride; - - const int2 q2_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); - const uint16_t * q2 = (const uint16_t *) &q2_packed; - - #pragma unroll - for (int l = 0; l < QR2_XS; ++l) { - const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l] & 0x1FF]; - const uint32_t signs = unpack_ksigns(q2[l] >> 9); - - const int signs0 = __vcmpne4(signs & 0x08040201, 0); - const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); - - const int signs1 = __vcmpne4(signs & 0x80402010, 0); - const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int ls = bxi->scales[kqsx]; - const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; - x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; -#else - x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; - x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq2_s( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_S, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_S)) / 2; - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq2_s * bxi = (const block_iq2_s *) x + kbx0 + i*stride; - - const int qs_packed = get_int_b2(bxi->qs, kqsx); - const uint8_t * qs = (const uint8_t *) &qs_packed; - - const int qh = bxi->qh[kqsx]; - - const int signs_packed_32 = get_int_b2(bxi->qs, QK_K/32 + kqsx); - const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; - -#pragma unroll - for (int l = 0; l < QR2_S; ++l) { - const int * grid_pos = (const int *)(iq2s_grid + (qs[l] | ((qh << (8-2*l)) & 0x300))); - - const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000); - const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000); - - const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int ls = bxi->scales[kqsx]; - const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; - x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; -#else - x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; - x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq3_xxs( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_XXS, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_XXS)) / 2; - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq3_xxs * bxi = (const block_iq3_xxs *) x + kbx0 + i*stride; - - const int2 q3_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); - const uint8_t * q3 = (const uint8_t *) &q3_packed; - const uint32_t aux32 = get_int_b2(bxi->qs, QK_K/16 + kqsx); - -#pragma unroll - for (int l = 0; l < QR3_XXS; ++l) { - const int2 grid_pos = make_int2(iq3xxs_grid[q3[2*l+0]], iq3xxs_grid[q3[2*l+1]]); - const uint32_t signs = unpack_ksigns(aux32 >> (7*l)); - - const int signs0 = __vcmpne4(signs & 0x08040201, 0); - const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); - - const int signs1 = __vcmpne4(signs & 0x80402010, 0); - const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid_h; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int ls = aux32 >> 28; - const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = (ls*d + d/2)/2; -#else - x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/2; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq3_s( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_S)) / 2; - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq3_s * bxi = (const block_iq3_s *) x + kbx0 + i*stride; - - const int2 qs_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); - const uint8_t * qs = (const uint8_t *) &qs_packed; - - const int qh = bxi->qh[kqsx]; - - const int signs_packed_32 = get_int_b2(bxi->signs, kqsx); - const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; - -#pragma unroll - for (int l = 0; l < QR3_S; ++l) { - const int2 grid_pos = make_int2( - iq3s_grid[qs[2*l+0] | ((qh << (8 - 2*l)) & 0x100)], - iq3s_grid[qs[2*l+1] | ((qh << (7 - 2*l)) & 0x100)]); - - const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000); - const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000); - - const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); - const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+0)] = grid_l; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+1)] = grid_h; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid_l; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int ls = 1 + 2*((bxi->scales[kqsx/2] >> (((2*kqsx) << 1) & 0x04)) & 0x0F); - const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = ls*d; -#else - x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = ls*d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq1_s( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_ds = (half2 *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_ds = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR1_S); - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq1_s * bxi = (const block_iq1_s *) x + kbx0 + i*stride; - - const int qs_packed = get_int_b2(bxi->qs, kqsx); - const uint8_t * qs = (const uint8_t *) &qs_packed; - - const int qh = bxi->qh[kqsx]; - - #pragma unroll - for (int l = 0; l < QR1_S/2; ++l) { - const int grid = iq1s_grid_gpu[qs[l] | (((qh >> (3*l)) & 0x07) << 8)]; - - const int grid0 = (grid >> 0) & 0x0F0F0F0F; - const int grid1 = (grid >> 4) & 0x0F0F0F0F; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+0)] = grid0; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+1)] = grid1; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid0; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const float d1q = __half2float(bxi->d) * (((qh >> 11) & 0x0E) + 1); - const float delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_ds[i*MMQ_MMA_TILE_X_K_Q8_1 + kqsx] = make_half2(d1q, d1q*delta); -#else - x_ds[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = make_half2(d1q, d1q*delta); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq4_xs( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_XS); - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride; - - const int aux_q4 = get_int_b4(bxi->qs, kqsx); - const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl); - const int k0 = 8 * (kqsx / 4) + kqsx % 4; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 4] = v.y; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 4] = v.y; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int rows_per_warp = warp_size / 8; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / (MMQ_TILE_NE_K/4); - - if (need_check) { - i = min(i, i_max); - } - - const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride; - - const float d = __half2float(bxi->d); - - const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F) - | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + threadIdx.x % 8] = d * (ls - 32); -#else - x_df[i*(MMQ_TILE_NE_K/4) + i/4 + threadIdx.x % 8] = d * (ls - 32); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void mmq_write_back_dp4a( +template static __device__ __forceinline__ void ggml_cuda_mmq_write_back_dp4a( const float * __restrict__ sum, const int32_t * __restrict__ ids_dst, float * __restrict__ dst, const int stride, const int i_max, const int j_max) { - constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { + for (int j0 = 0; j0 < J; j0 += nwarps) { const int j = j0 + threadIdx.y; if (j > j_max) { @@ -3199,45 +424,38 @@ static __device__ __forceinline__ void mmq_write_back_dp4a( } #pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { + for (int i0 = 0; i0 < I; i0 += warp_size) { const int i = i0 + threadIdx.x; - if (need_check && i > i_max) { + if (fallback && i > i_max) { continue; } - dst[ids_dst[j]*stride + i] = sum[(j0/nwarps) * (mmq_y/warp_size) + i0/warp_size]; + dst[ids_dst[j]*stride + i] = sum[(j0/nwarps) * (I/warp_size) + i0/warp_size]; } } } -template -static __device__ __forceinline__ void mmq_write_back_mma( - const float * __restrict__ sum, const int * __restrict__ ids_dst, float * __restrict__ dst, - const int stride, const int i_max, const int j_max) { - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int nwarps = mmq_get_nwarps_device(); - +template +static __device__ __forceinline__ void ggml_cuda_mmq_write_back_mma( + const float * __restrict__ sum, const int * __restrict__ ids_dst, float * __restrict__ dst, + const int stride, const int i_max, const int j_max) { #if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr int tileC_IJ = mmq_get_granularity_device(0); - typedef tile tile_C; - constexpr int rows_per_warp = granularity; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; #else - typedef tile<16, 8, int> tile_C; - constexpr int rows_per_warp = 2 * granularity; -#endif // defined(AMD_MFMA_AVAILABLE) - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + typedef tile<16, 8, int> tile_C; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. const int i0 = (threadIdx.y / ntx) * (ntx*tile_C::I); -#if defined(TURING_MMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - static_assert(nwarps*tile_C::I == mmq_y, "nwarps*tile_C::I != mmq_y"); -#else - GGML_UNUSED(nwarps); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { #pragma unroll for (int n = 0; n < ntx; ++n) { #pragma unroll @@ -3250,7 +468,7 @@ static __device__ __forceinline__ void mmq_write_back_mma( const int i = i0 + n*tile_C::I + tile_C::get_i(l); - if (need_check && i > i_max) { + if (fallback && i > i_max) { continue; } @@ -3262,188 +480,339 @@ static __device__ __forceinline__ void mmq_write_back_mma( // ------------------------------------------------------------------------------------------------------------------------------------- -template -struct mmq_type_traits; +// TODO remove this struct and use ggml_cuda_mmq_sram_layout instead. +struct ggml_cuda_mmq_util_funcs { + int vdr; + ggml_cuda_mmq_load_tiles_t load_tiles; + ggml_cuda_mmq_vec_dot_t vec_dot; + ggml_cuda_mmq_write_back_t write_back; -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q1_0_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q1_0; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; + constexpr __host__ __device__ ggml_cuda_mmq_util_funcs( + int vdr, ggml_cuda_mmq_load_tiles_t load_tiles, ggml_cuda_mmq_vec_dot_t vec_dot, ggml_cuda_mmq_write_back_t write_back) : + vdr(vdr), load_tiles(load_tiles), vec_dot(vec_dot), write_back(write_back) {} }; -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q4_0_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q4_0; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q4_0_q8_1_dp4a; -}; +template +static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_funcs() { + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q4_1_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q4_1; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_1_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q4_1_q8_1_dp4a; -}; + if (!ggml_cuda_mmq_get_config(type, J, fallback).use_mma_data_layout()) { + switch (type) { + case GGML_TYPE_Q1_0: + return ggml_cuda_mmq_util_funcs( + VDR_Q1_0_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q1_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q4_0: + return ggml_cuda_mmq_util_funcs( + VDR_Q4_0_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q4_0, + ggml_cuda_mmq_vec_dot_q4_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q4_1: + return ggml_cuda_mmq_util_funcs( + VDR_Q4_1_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q4_1, + ggml_cuda_mmq_vec_dot_q4_1_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q5_0: + return ggml_cuda_mmq_util_funcs( + VDR_Q5_0_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q5_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q5_1: + return ggml_cuda_mmq_util_funcs( + VDR_Q5_1_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q5_1, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q8_0: + return ggml_cuda_mmq_util_funcs( + VDR_Q8_0_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q8_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_Q2_K: + return ggml_cuda_mmq_util_funcs( + VDR_Q2_K_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q2_K, + ggml_cuda_mmq_vec_dot_q2_K_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q3_K: + return ggml_cuda_mmq_util_funcs( + VDR_Q3_K_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q3_K, + ggml_cuda_mmq_vec_dot_q3_K_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q4_K: + return ggml_cuda_mmq_util_funcs( + VDR_Q4_K_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q4_K, + ggml_cuda_mmq_vec_dot_q4_K_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q5_K: + return ggml_cuda_mmq_util_funcs( + VDR_Q5_K_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q5_K, + ggml_cuda_mmq_vec_dot_q5_K_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q6_K: + return ggml_cuda_mmq_util_funcs( + VDR_Q6_K_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q6_K, + ggml_cuda_mmq_vec_dot_q6_K_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_IQ1_S: + return ggml_cuda_mmq_util_funcs( + VDR_IQ1_S_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq1_s, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ2_XXS: + return ggml_cuda_mmq_util_funcs( + VDR_IQ2_XXS_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq2_xxs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ2_XS: + return ggml_cuda_mmq_util_funcs( + VDR_IQ2_XS_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq2_xs, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ2_S: + return ggml_cuda_mmq_util_funcs( + VDR_IQ2_S_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq2_s, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ3_XXS: + return ggml_cuda_mmq_util_funcs( + VDR_IQ3_XXS_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq3_xxs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ3_S: + return ggml_cuda_mmq_util_funcs( + VDR_IQ3_S_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq3_s, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ4_XS: + return ggml_cuda_mmq_util_funcs( + VDR_IQ4_XS_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq4_xs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ4_NL: + return ggml_cuda_mmq_util_funcs( + VDR_IQ4_NL_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq4_nl, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_MXFP4: + return ggml_cuda_mmq_util_funcs( + VDR_MXFP4_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_mxfp4, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_NVFP4: + return ggml_cuda_mmq_util_funcs( + VDR_NVFP4_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_nvfp4, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + default: + return ggml_cuda_mmq_util_funcs(1, nullptr, nullptr, nullptr); + } + } -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q5_0_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q5_0; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; +// --------------------------------------------------------------------------------------------- -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q5_1_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q5_1; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_1_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_1_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q8_0_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q8_0; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_MXFP4_Q8_1_MMQ; #ifdef BLACKWELL_MMA_AVAILABLE - static constexpr load_tiles_mmq_t load_tiles = load_tiles_mxfp4_fp4; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_fp4_fp4_mma; -#else - static constexpr load_tiles_mmq_t load_tiles = load_tiles_mxfp4; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; + switch (type) { + case GGML_TYPE_MXFP4: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_mxfp4_fp4, + ggml_cuda_mmq_vec_dot_fp4_fp4_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_NVFP4: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_nvfp4_nvfp4, + ggml_cuda_mmq_vec_dot_fp4_fp4_mma, + ggml_cuda_mmq_write_back_mma); + default: + break; + } #endif // BLACKWELL_MMA_AVAILABLE - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; -template -struct mmq_type_traits { - static constexpr int vdr = VDR_NVFP4_Q8_1_MMQ; -#ifdef BLACKWELL_MMA_AVAILABLE - static constexpr load_tiles_mmq_t load_tiles = load_tiles_nvfp4_nvfp4; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_fp4_fp4_mma; -#else - static constexpr load_tiles_mmq_t load_tiles = load_tiles_nvfp4; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_16_q8_1_mma; -#endif // BLACKWELL_MMA_AVAILABLE - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a; -}; +// --------------------------------------------------------------------------------------------- -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q2_K_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q2_K; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q2_K_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q2_K_q8_1_dp4a; -}; + switch (type) { + case GGML_TYPE_Q1_0: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q1_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q4_0: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q4_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q4_1: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q4_1, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q5_0: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q5_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q5_1: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q5_1, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q8_0: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q8_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_Q2_K: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q2_K, + ggml_cuda_mmq_vec_dot_q2_K_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q3_K: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q3_K, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q4_K: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q4_K, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q5_K: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q5_K, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q6_K: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q6_K, + ggml_cuda_mmq_vec_dot_q6_K_q8_1_mma, + ggml_cuda_mmq_write_back_mma); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_IQ1_S: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq1_s, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ2_XXS: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq2_xxs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ2_XS: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq2_xs, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ2_S: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq2_s, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ3_XXS: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq3_xxs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ3_S: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq3_s, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ4_XS: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq4_xs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ4_NL: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq4_nl, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_MXFP4: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_mxfp4, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_NVFP4: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_nvfp4, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + default: + return ggml_cuda_mmq_util_funcs(1, nullptr, nullptr, nullptr); + } +} -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q3_K_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q3_K; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_16_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q3_K_q8_1_dp4a; -}; +template +static constexpr __device__ int ggml_cuda_mmq_get_vdr() { + return ggml_cuda_mmq_get_util_funcs().vdr; +} -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q4_K_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q4_K; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_1_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q4_K_q8_1_dp4a; -}; +template +static constexpr __device__ ggml_cuda_mmq_load_tiles_t ggml_cuda_mmq_get_load_tiles() { + return ggml_cuda_mmq_get_util_funcs().load_tiles; +} -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q5_K_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q5_K; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_1_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q5_K_q8_1_dp4a; -}; +template +static constexpr __device__ ggml_cuda_mmq_vec_dot_t ggml_cuda_mmq_get_vec_dot() { + return ggml_cuda_mmq_get_util_funcs().vec_dot; +} -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q6_K_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q6_K; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q6_K_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q6_K_q8_1_dp4a; -}; +template +static constexpr __device__ ggml_cuda_mmq_write_back_t ggml_cuda_mmq_get_write_back() { + return ggml_cuda_mmq_get_util_funcs().write_back; +} -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ2_XXS_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq2_xxs; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; +// --------------------------------------------------------------------------------------------- -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ2_XS_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq2_xs; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_16_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ2_S_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq2_s; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_16_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ3_XXS_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq3_xxs; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ3_S_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq3_s; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ1_S_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq1_s; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_1_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_1_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ4_NL_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq4_nl; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ4_XS_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq4_xs; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; - -template +template static __device__ __forceinline__ void mul_mat_q_process_tile( const char * __restrict__ x, const int offset_x, const int * __restrict__ y, const int * __restrict__ ids_dst, float * __restrict__ dst, float * __restrict__ tmp_fixup, @@ -3451,22 +820,16 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( const int tile_x_max_i, const int tile_y_max_j, const int kb0_start, const int kb0_stop) { constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - constexpr int nwarps = mmq_get_nwarps_device(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; constexpr int qk = ggml_cuda_type_traits::qk; - constexpr int mmq_y = get_mmq_y_device(); - constexpr load_tiles_mmq_t load_tiles = mmq_type_traits::load_tiles; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr ggml_cuda_mmq_load_tiles_t load_tiles = ggml_cuda_mmq_get_load_tiles(); + constexpr ggml_cuda_mmq_vec_dot_t vec_dot = ggml_cuda_mmq_get_vec_dot(); + constexpr ggml_cuda_mmq_write_back_t write_back = ggml_cuda_mmq_get_write_back(); extern __shared__ int data_mul_mat_q[]; - int * tile_y = data_mul_mat_q + mmq_x; - int * tile_x = tile_y + GGML_PAD(mmq_x*MMQ_TILE_Y_K, nwarps*warp_size); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr vec_dot_mmq_t vec_dot = mmq_type_traits::vec_dot_mma; - constexpr mmq_write_back_t write_back = mmq_write_back_mma; -#else - constexpr vec_dot_mmq_t vec_dot = mmq_type_traits::vec_dot_dp4a; - constexpr mmq_write_back_t write_back = mmq_write_back_dp4a; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * tile_y = data_mul_mat_q + J; + int * tile_x = tile_y + GGML_PAD(J*MMQ_TILE_Y_K, nwarps*warp_size); #if defined(BLACKWELL_MMA_AVAILABLE) // FP4 tile stores 8 blocks @@ -3475,10 +838,10 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( constexpr int ne_block = 4 * QK8_1; #endif // defined(BLACKWELL_MMA_AVAILABLE) - constexpr int ITER_K = get_iter_k(type); + constexpr int ITER_K = ggml_cuda_mmq_get_K_vram(type, J, fallback); constexpr int blocks_per_iter = ITER_K / qk; - float sum[mmq_x*mmq_y / (nwarps*warp_size)] = {0.0f}; + float sum[J*I / (nwarps*warp_size)] = {0.0f}; constexpr int sz = sizeof(block_q8_1_mmq) / sizeof(int); @@ -3487,7 +850,7 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( { const int * by0 = y + ncols_y * (kb0 * qk / ne_block) * sz; #pragma unroll - for (int l0 = 0; l0 < mmq_x * MMQ_TILE_Y_K; l0 += nwarps * warp_size) { + for (int l0 = 0; l0 < J * MMQ_TILE_Y_K; l0 += nwarps * warp_size) { int l = l0 + threadIdx.y*warp_size + threadIdx.x; tile_y[l] = by0[l]; @@ -3503,7 +866,7 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( { const int * by0 = y + ncols_y * ((kb0 * qk / ne_block) * sz + sz); #pragma unroll - for (int l0 = 0; l0 < mmq_x * MMQ_TILE_Y_K; l0 += nwarps * warp_size) { + for (int l0 = 0; l0 < J * MMQ_TILE_Y_K; l0 += nwarps * warp_size) { int l = l0 + threadIdx.y*warp_size + threadIdx.x; tile_y[l] = by0[l]; @@ -3518,7 +881,7 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( } if (fixup) { - write_back(sum, ids_dst, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x); + write_back(sum, ids_dst, tmp_fixup + blockIdx.x*(J*I), I, I, J); } else { write_back(sum, ids_dst, dst, stride_col_dst, tile_x_max_i, tile_y_max_j); } @@ -3527,18 +890,8 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( // The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598 -template -#if defined(GGML_USE_HIP) -#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) - __launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device(), 2) -#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) -#else -#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA - __launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device(), 1) -#else - __launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device(), 2) -#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA -#endif // defined(GGML_USE_HIP) +template +__launch_bounds__(ggml_cuda_mmq_get_nthreads(type, J, fallback), ggml_cuda_mmq_get_occupancy(type, J, fallback)) static __global__ void mul_mat_q( const char * __restrict__ x, const int * __restrict__ y, const int32_t * __restrict__ ids_dst, const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup, @@ -3548,28 +901,27 @@ static __global__ void mul_mat_q( const uint3 ntx) { // Skip unused template specializations for faster compilation: - if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) { + if (ggml_cuda_mmq_get_config(type, J, fallback).type == GGML_TYPE_COUNT) { NO_DEVICE_CODE; return; } - constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int qk = ggml_cuda_type_traits::qk; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); - constexpr int qk = ggml_cuda_type_traits::qk; - constexpr int mmq_y = get_mmq_y_device(); - - const uint32_t nty = (nrows_x + mmq_y - 1) / mmq_y; // Number of tiles y + const uint32_t nty = (nrows_x + I - 1) / I; // Number of tiles y // Initialize the ids for writing back data with just the index. // For regular matrix multiplications this is never changed. // For MoE the correct indices are loaded from ids_dst. extern __shared__ int ids_dst_shared[]; // Stored at beginning of shared memory. #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps*warp_size) { + for (int j0 = 0; j0 < J; j0 += nwarps*warp_size) { const int j = j0 + threadIdx.y*warp_size + threadIdx.x; - if (j0 + nwarps*warp_size > mmq_x && j >= mmq_x) { + if (j0 + nwarps*warp_size > J && j >= J) { break; } @@ -3577,9 +929,7 @@ static __global__ void mul_mat_q( } __syncthreads(); - // On non-CDNA AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead: -#if (defined(GGML_USE_HIP) && !defined(CDNA)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA - { + if constexpr (!ggml_cuda_mmq_get_stream_k(type, J, fallback)) { const uint2 tmp2 = fast_div_modulo(blockIdx.z, nchannels_y); const int wt = tmp2.x; const int zt = tmp2.y; @@ -3591,7 +941,7 @@ static __global__ void mul_mat_q( int col_high = ncols_dst; int col_diff = ncols_dst; int offset_y = wt*stride_sample_y + zt*stride_channel_y; - int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst; + int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*J*stride_col_dst; if (ids_dst) { col_low = expert_bounds[zt + 0]; @@ -3601,41 +951,40 @@ static __global__ void mul_mat_q( offset_y = 0; offset_dst = 0; - if (jt*mmq_x >= col_diff) { + if (jt*J >= col_diff) { return; } // __syncthreads(); // There is no previous tile that could cause a race condition. #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps*warp_size) { + for (int j0 = 0; j0 < J; j0 += nwarps*warp_size) { const int j = j0 + threadIdx.y*warp_size + threadIdx.x; - if (j0 + nwarps*warp_size > mmq_x && j >= mmq_x) { + if (j0 + nwarps*warp_size > J && j >= J) { break; } - ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j]; + ids_dst_shared[j] = ids_dst[col_low + jt*J + j]; } __syncthreads(); } - offset_y += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int)); - offset_dst += it*mmq_y; + offset_y += (col_low + jt*J)*(sizeof(block_q8_1_mmq)/sizeof(int)); + offset_dst += it*I; - const int tile_x_max_i = nrows_x - it*mmq_y - 1; - const int tile_y_max_j = col_diff - jt*mmq_x - 1; + const int tile_x_max_i = nrows_x - it*I - 1; + const int tile_y_max_j = col_diff - jt*J - 1; - const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x; + const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*I*stride_row_x; constexpr bool fixup = false; - mul_mat_q_process_tile + mul_mat_q_process_tile (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst, tile_x_max_i, tile_y_max_j, 0, blocks_per_ne00.z); return; } -#endif // (defined(GGML_USE_HIP) && !defined(CDNA4) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA - constexpr int ITER_K = get_iter_k(type); + constexpr int ITER_K = ggml_cuda_mmq_get_K_vram(type, J, fallback); constexpr int blocks_per_iter = ITER_K / qk; // kbc == k block continuous, current index in continuous ijk space. @@ -3665,7 +1014,7 @@ static __global__ void mul_mat_q( int col_high = ncols_dst; int col_diff = ncols_dst; int offset_y = wt*stride_sample_y + zt*stride_channel_y; - int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst; + int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*J*stride_col_dst; if (ids_dst) { col_low = expert_bounds[zt + 0]; @@ -3675,7 +1024,7 @@ static __global__ void mul_mat_q( offset_y = 0; offset_dst = 0; - if (jt*mmq_x >= col_diff) { + if (jt*J >= col_diff) { kbc += blocks_per_ne00.z; kbc -= fastmodulo(kbc, blocks_per_ne00); @@ -3687,28 +1036,28 @@ static __global__ void mul_mat_q( __syncthreads(); #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps*warp_size) { + for (int j0 = 0; j0 < J; j0 += nwarps*warp_size) { const int j = j0 + threadIdx.y*warp_size + threadIdx.x; - if (j0 + nwarps*warp_size > mmq_x && j >= mmq_x) { + if (j0 + nwarps*warp_size > J && j >= J) { break; } - ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j]; + ids_dst_shared[j] = ids_dst[col_low + jt*J + j]; } __syncthreads(); } - offset_y += (col_low + jt * mmq_x) * (sizeof(block_q8_1_mmq) / sizeof(int)); - offset_dst += it*mmq_y; + offset_y += (col_low + jt * J) * (sizeof(block_q8_1_mmq) / sizeof(int)); + offset_dst += it*I; - const int tile_x_max_i = nrows_x - it*mmq_y - 1; - const int tile_y_max_j = col_diff - jt*mmq_x - 1; + const int tile_x_max_i = nrows_x - it*I - 1; + const int tile_y_max_j = col_diff - jt*J - 1; - const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x; + const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*I*stride_row_x; constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer. - mul_mat_q_process_tile + mul_mat_q_process_tile (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst, tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop); @@ -3739,7 +1088,7 @@ static __global__ void mul_mat_q( int col_high = ncols_dst; int col_diff = ncols_dst; int offset_y = wt*stride_sample_y + zt*stride_channel_y; - int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst; + int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*J*stride_col_dst; if (ids_dst) { col_low = expert_bounds[zt + 0]; @@ -3749,17 +1098,17 @@ static __global__ void mul_mat_q( offset_y = 0; offset_dst = 0; - if (jt*mmq_x >= col_diff) { + if (jt*J >= col_diff) { return; } // The memory layout for the fixup buffer is always contiguous, therefore reset ids: __syncthreads(); #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps*warp_size) { + for (int j0 = 0; j0 < J; j0 += nwarps*warp_size) { const int j = j0 + threadIdx.y*warp_size + threadIdx.x; - if (j0 + nwarps*warp_size > mmq_x && j >= mmq_x) { + if (j0 + nwarps*warp_size > J && j >= J) { break; } @@ -3768,39 +1117,38 @@ static __global__ void mul_mat_q( __syncthreads(); } - offset_y += (col_low + jt * mmq_x) * (sizeof(block_q8_1_mmq) / sizeof(int)); - offset_dst += it*mmq_y; + offset_y += (col_low + jt * J) * (sizeof(block_q8_1_mmq) / sizeof(int)); + offset_dst += it*I; - const int tile_x_max_i = nrows_x - it*mmq_y - 1; - const int tile_y_max_j = col_diff - jt*mmq_x - 1; + const int tile_x_max_i = nrows_x - it*I - 1; + const int tile_y_max_j = col_diff - jt*J - 1; - const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x; + const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*I*stride_row_x; constexpr bool fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks. - mul_mat_q_process_tile + mul_mat_q_process_tile (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst, tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop); } -template -__launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device()/2, 1) +template +__launch_bounds__(ggml_cuda_mmq_get_nthreads(type, J, fallback)/2, 1) static __global__ void mul_mat_q_stream_k_fixup( const int32_t * __restrict__ ids_dst, const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_last_tile, const uint3 blocks_per_ne00, const int nrows_x, const int ncols_dst, const int stride_col_dst, const uint3 nchannels_y, const int stride_channel_dst, const uint3 nsamples_y, const int stride_sample_dst, const uint3 ntx) { - constexpr int mmq_y = get_mmq_y_device(); + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = (ggml_cuda_mmq_get_nthreads(type, J, fallback) / 2) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); constexpr int qk = ggml_cuda_type_traits::qk; - constexpr int ITER_K = get_iter_k(type); + constexpr int ITER_K = ggml_cuda_mmq_get_K_vram(type, J, fallback); constexpr int blocks_per_iter = ITER_K / qk; - constexpr int nwarps = mmq_get_nwarps_device()/2; - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - float sum[mmq_x / nwarps] = {0.0f}; + float sum[J / nwarps] = {0.0f}; const int i = blockIdx.y*warp_size + threadIdx.x; - const int nty = (nrows_x + mmq_y - 1) / mmq_y; + const int nty = (nrows_x + I - 1) / I; const int bidx0 = blockIdx.x; @@ -3838,10 +1186,10 @@ static __global__ void mul_mat_q_stream_k_fixup( #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { + for (int j0 = 0; j0 < J; j0 += nwarps) { const int j = j0 + threadIdx.y; - sum[j0/nwarps] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i]; + sum[j0/nwarps] += tmp_last_tile[bidx*(J*I) + j*I + i]; } // If this block started in a previous tile we are done and don't need to combine additional partial results. @@ -3868,17 +1216,17 @@ static __global__ void mul_mat_q_stream_k_fixup( const int it = tmp2.x; if (!ids_dst) { - const int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst + it*mmq_y; + const int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*J*stride_col_dst + it*I; dst += offset_dst; - const int i_max = nrows_x - it*mmq_y - 1; - const int j_max = ncols_dst - jt*mmq_x - 1; - if (need_check && i > i_max) { + const int i_max = nrows_x - it*I - 1; + const int j_max = ncols_dst - jt*J - 1; + if (fallback && i > i_max) { return; } #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { + for (int j0 = 0; j0 < J; j0 += nwarps) { const int j = j0 + threadIdx.y; if (j > j_max) { @@ -3890,27 +1238,27 @@ static __global__ void mul_mat_q_stream_k_fixup( return; } - __shared__ int ids_dst_shared[mmq_x]; + __shared__ int ids_dst_shared[J]; const int col_low = expert_bounds[zt + 0]; const int col_high = expert_bounds[zt + 1]; const int col_diff = col_high - col_low; - for (int j = threadIdx.y*warp_size + threadIdx.x; j < mmq_x; j += nwarps*warp_size) { - ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j]; + for (int j = threadIdx.y*warp_size + threadIdx.x; j < J; j += nwarps*warp_size) { + ids_dst_shared[j] = ids_dst[col_low + jt*J + j]; } __syncthreads(); - const int offset_dst = it*mmq_y; + const int offset_dst = it*I; dst += offset_dst; - const int i_max = nrows_x - it*mmq_y - 1; - const int j_max = col_diff - jt*mmq_x - 1; - if (need_check && i > i_max) { + const int i_max = nrows_x - it*I - 1; + const int j_max = col_diff - jt*J - 1; + if (fallback && i > i_max) { return; } #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { + for (int j0 = 0; j0 < J; j0 += nwarps) { const int j = j0 + threadIdx.y; if (j > j_max) { @@ -3926,37 +1274,35 @@ struct mmq_args { int64_t ncols_x; int64_t nrows_x; int64_t ncols_dst; int64_t stride_row_x; int64_t ncols_y; int64_t nrows_dst; int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst; int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst; - bool use_stream_k; int64_t ncols_max; + int64_t ncols_max; }; -template -static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int cc, const int warp_size, const int nwarps) { - const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y); - const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type); - const size_t nbs_ids = mmq_x*sizeof(int); - const size_t nbs_x = (turing_mma_available(cc) || amd_mfma_available(cc) || amd_wmma_available(cc)) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int); - const size_t nbs_y = mmq_x * (sizeof(block_q8_1_mmq)); - return nbs_ids + nbs_x + GGML_PAD(nbs_y, nwarps*warp_size*sizeof(int)); +static size_t mmq_get_nbytes_shared(const ggml_cuda_mmq_config & config, const int cc) { + const size_t nbs_ids = config.J*sizeof(int); + const size_t nbs_x = ggml_cuda_mmq_get_nbytes_shared_x(config, cc); + const size_t nbs_y = config.J * (sizeof(block_q8_1_mmq)); + return nbs_ids + nbs_x + GGML_PAD(nbs_y, config.nthreads*sizeof(int)); } -template +template static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { const int id = ggml_cuda_get_device(); const int cc = ggml_cuda_info().devices[id].cc; const int nsm = ggml_cuda_info().devices[id].nsm; const int warp_size = ggml_cuda_info().devices[id].warp_size; - const int nwarps = mmq_get_nwarps_host(cc, warp_size); - const int mmq_y = get_mmq_y_host(cc); + + const ggml_cuda_mmq_config config = ggml_cuda_mmq_get_config(type, J, fallback, cc); + GGML_ASSERT(config.nthreads % warp_size == 0); + const int nwarps = config.nthreads / warp_size; + const int nbytes_shared = mmq_get_nbytes_shared(config, cc); const dim3 block_dims(warp_size, nwarps, 1); - const int nbytes_shared = mmq_get_nbytes_shared(mmq_x, mmq_y, cc, warp_size, nwarps); + CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q), nbytes_shared); + CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q), nbytes_shared); - CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q), nbytes_shared); - CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q), nbytes_shared); - - const int nty = (args.nrows_x + mmq_y - 1) / mmq_y; - const int ntx = (args.ncols_max + mmq_x - 1) / mmq_x; + const int nty = (args.nrows_x + config.I - 1) / config.I; + const int ntx = (args.ncols_max + config.J - 1) / config.J; const int ntzw = args.nchannels_y * args.nsamples_y; const dim3 block_nums_xy_tiling(nty, ntx, ntzw); @@ -3972,24 +1318,13 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a const uint3 channel_ratio_fd = init_fastdiv_values(channel_ratio); const uint3 sample_ratio_fd = init_fastdiv_values(sample_ratio); - if (!args.use_stream_k) { - if (args.nrows_x % mmq_y == 0) { - constexpr bool need_check = false; - mul_mat_q<<>> - (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr, - blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, - channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, - sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, - ntx_fd); - } else { - constexpr bool need_check = true; - mul_mat_q<<>> - (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr, - blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, - channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, - sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, - ntx_fd); - } + if (!ggml_cuda_mmq_get_stream_k(type, J, fallback, cc)) { + mul_mat_q<<>> + (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr, + blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, + channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, + sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, + ntx_fd); return; } @@ -4007,133 +1342,121 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a ggml_cuda_pool & pool = ctx.pool(id); ggml_cuda_pool_alloc tmp_fixup(pool); if (fixup_needed) { - tmp_fixup.alloc(block_nums_stream_k.x * mmq_x*mmq_y); + tmp_fixup.alloc(block_nums_stream_k.x * config.J*config.I); } - const dim3 block_nums_fixup(block_nums_stream_k.x, mmq_y/warp_size, 1); + const dim3 block_nums_fixup(block_nums_stream_k.x, config.I/warp_size, 1); const dim3 block_dims_fixup(block_dims.x, block_dims.y/2, block_dims.z); - if (args.nrows_x % mmq_y == 0) { - constexpr bool need_check = false; - mul_mat_q<<>> - (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, - blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, - channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, - sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, - ntx_fd); + mul_mat_q<<>> + (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, + blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, + channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, + sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, + ntx_fd); - if (!fixup_needed) { - return; + if (!fixup_needed) { + return; + } + + CUDA_CHECK(cudaGetLastError()); + mul_mat_q_stream_k_fixup<<>> + (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, + args.nrows_dst, nchannels_y_fd, args.stride_channel_dst, nsamples_y_fd, args.stride_sample_dst, + ntx_fd); +} + +template +void mul_mat_q_switch_J(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { + const int id = ggml_cuda_get_device(); + const int cc = ggml_cuda_info().devices[id].cc; + const size_t smpbo = ggml_cuda_info().devices[id].smpbo; + + int J_best = 0; + int ntiles_J_best = INT_MAX; + + for (int J = 8; J <= 128 && ntiles_J_best > 1; J += 8) { + const ggml_cuda_mmq_config config = ggml_cuda_mmq_get_config(type, J, fallback, cc); + if (config.type == GGML_TYPE_COUNT) { + continue; } - CUDA_CHECK(cudaGetLastError()); - mul_mat_q_stream_k_fixup<<>> - (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, - args.nrows_dst, nchannels_y_fd, args.stride_channel_dst, nsamples_y_fd, args.stride_sample_dst, - ntx_fd); - } else { - constexpr bool need_check = true; - mul_mat_q<<>> - (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, - blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, - channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, - sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, - ntx_fd); - - if (!fixup_needed) { - return; + if (mmq_get_nbytes_shared(config, cc) > smpbo) { + continue; } - CUDA_CHECK(cudaGetLastError()); - mul_mat_q_stream_k_fixup<<>> - (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, - args.nrows_dst, nchannels_y_fd, args.stride_channel_dst, nsamples_y_fd, args.stride_sample_dst, - ntx_fd); + const int ntiles_x = (args.ncols_max + config.J - 1) / config.J; + + if (ntiles_x < ntiles_J_best) { + J_best = J; + ntiles_J_best = ntiles_x; + } + } + + switch (J_best) { + case 8: + launch_mul_mat_q(ctx, args, stream); + break; + case 16: + launch_mul_mat_q(ctx, args, stream); + break; + case 24: + launch_mul_mat_q(ctx, args, stream); + break; + case 32: + launch_mul_mat_q(ctx, args, stream); + break; + case 40: + launch_mul_mat_q(ctx, args, stream); + break; + case 48: + launch_mul_mat_q(ctx, args, stream); + break; + case 56: + launch_mul_mat_q(ctx, args, stream); + break; + case 64: + launch_mul_mat_q(ctx, args, stream); + break; + case 72: + launch_mul_mat_q(ctx, args, stream); + break; + case 80: + launch_mul_mat_q(ctx, args, stream); + break; + case 88: + launch_mul_mat_q(ctx, args, stream); + break; + case 96: + launch_mul_mat_q(ctx, args, stream); + break; + case 104: + launch_mul_mat_q(ctx, args, stream); + break; + case 112: + launch_mul_mat_q(ctx, args, stream); + break; + case 120: + launch_mul_mat_q(ctx, args, stream); + break; + case 128: + launch_mul_mat_q(ctx, args, stream); + break; + default: + fprintf(stderr, "J_best=%d\n", J_best); + GGML_ABORT("fatal error"); + break; } } template void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { - const int id = ggml_cuda_get_device(); - const int cc = ggml_cuda_info().devices[id].cc; - const size_t smpbo = ggml_cuda_info().devices[id].smpbo; - const int warp_size = ggml_cuda_info().devices[id].warp_size; - const int nwarps = mmq_get_nwarps_host(cc, warp_size); - - const int mmq_x_max = get_mmq_x_max_host(cc); - const int mmq_y = get_mmq_y_host(cc); - - int mmq_x_best = 0; - int ntiles_x_best = INT_MAX; - - for (int mmq_x = 8; mmq_x <= mmq_x_max && ntiles_x_best > 1; mmq_x += 8) { - const int granularity = mmq_get_granularity_host(mmq_x, cc); - - if (mmq_x % granularity != 0 || mmq_get_nbytes_shared(mmq_x, mmq_y, cc, warp_size, nwarps) > smpbo) { - continue; - } - - const int ntiles_x = (args.ncols_max + mmq_x - 1) / mmq_x; - - if (ntiles_x < ntiles_x_best) { - mmq_x_best = mmq_x; - ntiles_x_best = ntiles_x; - } - } - - switch (mmq_x_best) { - case 8: - launch_mul_mat_q(ctx, args, stream); - break; - case 16: - launch_mul_mat_q(ctx, args, stream); - break; - case 24: - launch_mul_mat_q(ctx, args, stream); - break; - case 32: - launch_mul_mat_q(ctx, args, stream); - break; - case 40: - launch_mul_mat_q(ctx, args, stream); - break; - case 48: - launch_mul_mat_q(ctx, args, stream); - break; - case 56: - launch_mul_mat_q(ctx, args, stream); - break; - case 64: - launch_mul_mat_q(ctx, args, stream); - break; - case 72: - launch_mul_mat_q(ctx, args, stream); - break; - case 80: - launch_mul_mat_q(ctx, args, stream); - break; - case 88: - launch_mul_mat_q(ctx, args, stream); - break; - case 96: - launch_mul_mat_q(ctx, args, stream); - break; - case 104: - launch_mul_mat_q(ctx, args, stream); - break; - case 112: - launch_mul_mat_q(ctx, args, stream); - break; - case 120: - launch_mul_mat_q(ctx, args, stream); - break; - case 128: - launch_mul_mat_q(ctx, args, stream); - break; - default: - fprintf(stderr, "mmq_x_best=%d\n", mmq_x_best); - GGML_ABORT("fatal error"); - break; + if (args.nrows_x % 128 == 0) { + constexpr bool fallback = false; + mul_mat_q_switch_J(ctx, args, stream); + } else { + constexpr bool fallback = true; + mul_mat_q_switch_J(ctx, args, stream); } } @@ -4166,11 +1489,4 @@ extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS); void ggml_cuda_mul_mat_q( ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst); -void ggml_cuda_op_mul_mat_q( - ggml_backend_cuda_context & ctx, - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, - const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, - const int64_t src1_padded_row_size, cudaStream_t stream); - bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t n_experts); - From 2969d6d15d67a08e7b83f26164b15350c79c5248 Mon Sep 17 00:00:00 2001 From: Satinder Grewal Date: Tue, 14 Jul 2026 10:31:04 +1200 Subject: [PATCH 02/24] model: add Hy3 (hy_v3) support with MTP speculative decoding (#25395) * model: add Hy3 (hy_v3) architecture support Adds Tencent Hunyuan 3 (HF architecture HYV3ForCausalLM, GGUF arch hy_v3): a MoE decoder stack with per-head Q/K RMSNorm, a sigmoid router with expert selection bias, an always-active ungated shared expert, and leading dense block(s) (first_k_dense_replace). The base implementation is ported from charlie12345's fork (https://github.com/charlie12345/ROCmFPX, src/models/hyv3.cpp), adapted to current mainline APIs (hparams.n_layer(), build_qkv, build_moe_ffn with fused gate_up + scale tensors, output_s). Note: blk.N.exp_probs_b is stored without a .bias suffix for compatibility with existing hy_v3 GGUFs produced by that fork. Co-Authored-By: charlie12345 Co-authored-by: Piotr Wilkin Assisted-by: Claude Fable 5 --- common/chat-auto-parser-generator.cpp | 4 + common/chat-auto-parser.h | 7 +- common/chat-diff-analyzer.cpp | 30 ++ common/jinja/value.cpp | 39 +++ conversion/__init__.py | 1 + conversion/hunyuan.py | 103 +++++++ gguf-py/gguf/constants.py | 33 +++ models/templates/tencent-Hy3.jinja | 222 +++++++++++++++ src/llama-arch.cpp | 1 + src/llama-arch.h | 1 + src/llama-model.cpp | 5 +- src/models/hy-v3.cpp | 390 ++++++++++++++++++++++++++ src/models/models.h | 16 ++ tests/test-chat-auto-parser.cpp | 3 + tests/test-jinja.cpp | 30 ++ tests/test-llama-archs.cpp | 1 + 16 files changed, 882 insertions(+), 4 deletions(-) create mode 100644 models/templates/tencent-Hy3.jinja create mode 100644 src/models/hy-v3.cpp diff --git a/common/chat-auto-parser-generator.cpp b/common/chat-auto-parser-generator.cpp index ddf81fc4d..3f91364c5 100644 --- a/common/chat-auto-parser-generator.cpp +++ b/common/chat-auto-parser-generator.cpp @@ -262,6 +262,10 @@ common_peg_parser analyze_tools::build_func_parser(common_chat_peg_builder & p, bool matched_atomic = false; common_peg_parser func_parser = p.eps(); + if (!function.args_separator.empty()) { + open = open + p.space() + p.literal(function.args_separator); + } + if (!function.name_suffix.empty()) { func_parser = open + call_id_section + p.space() + args; matched_atomic = true; diff --git a/common/chat-auto-parser.h b/common/chat-auto-parser.h index 9e8113f24..d47b09dcf 100644 --- a/common/chat-auto-parser.h +++ b/common/chat-auto-parser.h @@ -192,9 +192,10 @@ struct tool_format_analysis { }; struct tool_function_analysis { - std::string name_prefix; // e.g., "", "\"", ":0" - std::string close; // e.g., "", "" (for tag-based) + std::string name_prefix; // e.g., "", "\"", ":0" + std::string args_separator; // e.g., "" (marker between function name and arguments) + std::string close; // e.g., "", "" (for tag-based) }; struct tool_arguments_analysis { diff --git a/common/chat-diff-analyzer.cpp b/common/chat-diff-analyzer.cpp index 79598368c..127278dfb 100644 --- a/common/chat-diff-analyzer.cpp +++ b/common/chat-diff-analyzer.cpp @@ -259,6 +259,7 @@ void autoparser::analyze_template(const common_chat_template & tmpl) { LOG_DBG("per_call_end: '%s'\n", tools.format.per_call_end.c_str()); LOG_DBG("func_name_prefix: '%s'\n", tools.function.name_prefix.c_str()); LOG_DBG("func_name_suffix: '%s'\n", tools.function.name_suffix.c_str()); + LOG_DBG("func_args_separator: '%s'\n", tools.function.args_separator.c_str()); LOG_DBG("func_close: '%s'\n", tools.function.close.c_str()); LOG_DBG("call_id_prefix: '%s'\n", tools.call_id.prefix.c_str()); LOG_DBG("call_id_suffix: '%s'\n", tools.call_id.suffix.c_str()); @@ -302,6 +303,7 @@ void autoparser::collect_preserved_tokens() { add_token(tools.format.per_call_end); add_token(tools.function.name_prefix); add_token(tools.function.name_suffix); + add_token(tools.function.args_separator); add_token(tools.function.close); add_token(tools.arguments.start); add_token(tools.arguments.end); @@ -1051,6 +1053,23 @@ void analyze_tools::check_per_call_markers() { format.section_start.clear(); format.section_end.clear(); } + + if (!format.per_call_end.empty()) { + auto count_occurrences = [](const std::string & haystack, const std::string & needle) { + size_t count = 0; + for (size_t pos = haystack.find(needle); pos != std::string::npos; + pos = haystack.find(needle, pos + needle.size())) { + count++; + } + return count; + }; + size_t calls_one = count_occurrences(one_vs_two->output_A, format.per_call_end); + size_t calls_two = count_occurrences(one_vs_two->output_B, format.per_call_end); + if (calls_one > 0 && calls_one == calls_two) { + format.section_end = format.per_call_end; + format.per_call_end.clear(); + } + } } void analyze_tools::extract_function_markers() { @@ -1132,6 +1151,17 @@ void analyze_tools::extract_function_markers() { auto suf_result = suffix_parser.parse_and_extract(diff.suffix); if (suf_result.result.success()) { function.name_suffix += suf_result.tags["ext"]; + + auto arg_start = [&](common_peg_parser_builder &p) { + return p.marker() + p.space() + p.choice({ p.literal(ARG_FIRST), p.literal(ARG_SECOND) }); + }; + auto sep_parser = build_tagged_peg_parser([&](common_peg_parser_builder &p) { + return p.tag("sep", p.zero_or_more(p.negate(arg_start(p)) + p.any())) + arg_start(p); + }); + auto sep_result = sep_parser.parse_and_extract(diff.suffix.substr(suf_result.tags["ext"].size())); + if (sep_result.result.success()) { + function.args_separator = trim_whitespace(sep_result.tags["sep"]); + } } } diff --git a/common/jinja/value.cpp b/common/jinja/value.cpp index 5055ae9ac..870596d61 100644 --- a/common/jinja/value.cpp +++ b/common/jinja/value.cpp @@ -750,11 +750,50 @@ const func_builtins & value_string_t::get_builtins() const { res->val_str.mark_input_based_on(args.get_pos(0)->val_str); return res; }}, + {"format", [](const func_args & args) -> value { + value val_input = args.get_pos(0); + if (!is_val(val_input)) { + throw raised_exception("format() first argument must be a string"); + } + const jinja::string & fmt = val_input->as_string(); + const bool fmt_is_input = fmt.all_parts_are_input(); + + const std::string str = fmt.str(); + jinja::string result; + std::string literal; + auto flush_literal = [&]() { + if (!literal.empty()) { + result.parts.push_back({fmt_is_input, literal}); + literal.clear(); + } + }; + + size_t arg_idx = 1; // positional args follow the format string + for (size_t i = 0; i < str.size(); ++i) { + if (str[i] != '{') { + literal += str[i]; + continue; + } + if (i + 1 >= str.size() || str[i + 1] != '}') { + throw not_implemented_exception("format() only supports simple '{}' placeholders"); + } + ++i; + flush_literal(); + const jinja::string arg_str = args.get_pos(arg_idx++)->as_string(); + result.parts.insert(result.parts.end(), arg_str.parts.begin(), arg_str.parts.end()); + } + flush_literal(); + return mk_val(result); + }}, {"int", [](const func_args & args) -> value { value val_input = args.get_pos(0); value val_default = args.get_kwarg_or_pos("default", 1); value val_base = args.get_kwarg_or_pos("base", 2); const int base = val_base->is_undefined() ? 10 : val_base->as_int(); + if (base != 0 && (base < 2 || base > 36)) { + // an out-of-range base makes std::stoi fail fast on the MSVC CRT instead of throwing + throw raised_exception("int() base must be 0 or between 2 and 36"); + } if (is_val(val_input) == false) { throw raised_exception("int() first argument must be a string"); } diff --git a/conversion/__init__.py b/conversion/__init__.py index 02ea63852..02102fac8 100644 --- a/conversion/__init__.py +++ b/conversion/__init__.py @@ -106,6 +106,7 @@ TEXT_MODEL_MAP: dict[str, str] = { "HunYuanDenseV1ForCausalLM": "hunyuan", "HunYuanMoEV1ForCausalLM": "hunyuan", "HunYuanVLForConditionalGeneration": "hunyuan", + "HYV3ForCausalLM": "hunyuan", "IQuestCoderForCausalLM": "llama", "InternLM2ForCausalLM": "internlm", "InternLM3ForCausalLM": "internlm", diff --git a/conversion/hunyuan.py b/conversion/hunyuan.py index 537f023aa..4d2545f8b 100644 --- a/conversion/hunyuan.py +++ b/conversion/hunyuan.py @@ -1,6 +1,7 @@ from __future__ import annotations import json +import re from pathlib import Path from typing import Callable, Iterable, TYPE_CHECKING @@ -355,3 +356,105 @@ class HunyuanVLTextModel(HunYuanModel): self.gguf_writer.add_context_length(ctx_len) self.gguf_writer.add_rope_dimension_sections(list(self.rope_parameters["xdrope_section"])) + + +@ModelBase.register("HYV3ForCausalLM") +class HYV3Model(TextModel): + model_arch = gguf.MODEL_ARCH.HY_V3 + + # Trunk layer count, stashed before indexing so the classmethod + # filter_tensors can identify the appended MTP block(s) (mirrors + # Step35Model). + _n_main_layers: int | None = None + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # NextN/MTP layers are appended past num_hidden_layers; extend the + # tensor map so the MTP block's tensors resolve to blk..* names. + n_nextn = int(self.hparams.get("num_nextn_predict_layers", 0)) + if n_nextn > 0 and not self.no_mtp: + self.block_count += n_nextn + self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) + + def index_tensors(self, remote_hf_model_id: str | None = None): + type(self)._n_main_layers = self.hparams["num_hidden_layers"] + return super().index_tensors(remote_hf_model_id=remote_hf_model_id) + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + super().set_gguf_parameters() + self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"]) + self.gguf_writer.add_expert_shared_feed_forward_length( + self.hparams["moe_intermediate_size"] * self.hparams.get("num_shared_experts", 1) + ) + self.gguf_writer.add_expert_weights_norm(self.hparams.get("route_norm", True)) + self.gguf_writer.add_expert_weights_scale(float(self.hparams.get("router_scaling_factor", 1.0))) + # sigmoid router with expert selection bias + self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID) + + n_nextn = int(self.hparams.get("num_nextn_predict_layers", 0)) + if n_nextn > 0 and not self.no_mtp: + self.gguf_writer.add_nextn_predict_layers(n_nextn) + + @classmethod + def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None: + if (titem := super().filter_tensors(item)) is None: + return None + name, gen = titem + + # HY V3 appends the MTP block(s) past num_hidden_layers. + assert cls._n_main_layers is not None + is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers + + # --no-mtp: drop the appended MTP block(s) entirely. + if is_mtp and cls.no_mtp: + return None + # --mtp: keep ONLY MTP-block tensors plus the shared embeddings/norm/ + # lm_head (so the resulting GGUF carries just the draft head). + if cls.mtp_only and not is_mtp and name not in ( + "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight", + ): + return None + + # The MTP block's trailing final_layernorm (applied after the decoder + # block, before the shared LM head) maps to nextn.shared_head_norm. + if is_mtp: + name = name.replace(".final_layernorm.", ".shared_head.norm.") + + return name, gen + + _experts: list[dict[str, Tensor]] | None = None + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + # merge the per-expert tensors into stacked 3d tensors + if name.startswith("model.layers.") and ".mlp.experts." in name: + n_experts = self.find_hparam(["num_local_experts", "num_experts"]) + assert bid is not None + + if self._experts is None: + self._experts = [{} for _ in range(self.block_count)] + + self._experts[bid][name] = data_torch + + if len(self._experts[bid]) >= n_experts * 3: + for w_name in ("down_proj", "gate_proj", "up_proj"): + datas: list[Tensor] = [] + for xid in range(n_experts): + ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight" + datas.append(self._experts[bid][ename]) + del self._experts[bid][ename] + + merged = torch.stack(datas, dim=0) + yield from super().modify_tensors(merged, f"model.layers.{bid}.mlp.experts.{w_name}.weight", bid) + return + + yield from super().modify_tensors(data_torch, name, bid) + + def prepare_tensors(self): + super().prepare_tensors() + if self._experts is not None: + experts = [k for d in self._experts for k in d.keys()] + if experts: + raise ValueError(f"Unprocessed experts: {experts}") diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 869e436ac..63ac2ed1f 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -512,6 +512,7 @@ class MODEL_ARCH(IntEnum): HUNYUAN_MOE = auto() HUNYUAN_DENSE = auto() HUNYUAN_VL = auto() + HY_V3 = auto() SMOLLM3 = auto() GPT_OSS = auto() LFM2 = auto() @@ -1093,6 +1094,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.HUNYUAN_MOE: "hunyuan-moe", MODEL_ARCH.HUNYUAN_DENSE: "hunyuan-dense", MODEL_ARCH.HUNYUAN_VL: "hunyuan_vl", + MODEL_ARCH.HY_V3: "hy_v3", MODEL_ARCH.SMOLLM3: "smollm3", MODEL_ARCH.GPT_OSS: "gpt-oss", MODEL_ARCH.LFM2: "lfm2", @@ -3936,6 +3938,37 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.HY_V3: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_EXP_PROBS_B, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + # NextN/MTP tensors (draft head) + MODEL_TENSOR.NEXTN_EH_PROJ, + MODEL_TENSOR.NEXTN_EMBED_TOKENS, + MODEL_TENSOR.NEXTN_ENORM, + MODEL_TENSOR.NEXTN_HNORM, + MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD, + MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM, + ], MODEL_ARCH.SMOLLM3: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/models/templates/tencent-Hy3.jinja b/models/templates/tencent-Hy3.jinja new file mode 100644 index 000000000..7591102ca --- /dev/null +++ b/models/templates/tencent-Hy3.jinja @@ -0,0 +1,222 @@ +{#- ------------- special token variables ------------- -#} +{%- set HYTK = ':opensource' %} +{%- set eos_token = '<｜hy_eos{}｜>'.format(HYTK) %} +{%- set bos_token = '<｜hy_begin_of_sentence{}｜>'.format(HYTK) %} +{%- set pad_token = '<｜hy_pad{}｜>'.format(HYTK) %} +{%- set user_token = '<｜hy_User{}｜>'.format(HYTK) %} +{%- set assistant_token = '<｜hy_Assistant{}｜>'.format(HYTK) %} +{%- set think_begin_token = ''.format(HYTK) %} +{%- set think_end_token = ''.format(HYTK) %} +{%- set toolcalls_begin_token = ''.format(HYTK) %} +{%- set toolcalls_end_token = ''.format(HYTK) %} +{%- set toolcall_begin_token = ''.format(HYTK) %} +{%- set toolcall_end_token = ''.format(HYTK) %} +{%- set toolsep_token = ''.format(HYTK) %} +{%- set argkey_begin_token = ''.format(HYTK) %} +{%- set argkey_end_token = ''.format(HYTK) %} +{%- set argvalue_begin_token = ''.format(HYTK) %} +{%- set argvalue_end_token = ''.format(HYTK) %} +{%- set toolresponses_begin_token = ''.format(HYTK) %} +{%- set toolresponses_end_token = ''.format(HYTK) %} +{%- set toolresponse_begin_token = ''.format(HYTK) %} +{%- set toolresponse_end_token = ''.format(HYTK) %} +{%- set reasoning_mode_token = '<｜reasoning_mode{}｜>'.format(HYTK) %} + +{#- ------------- hyperparameters variables ------------- -#} +{%- if not add_generation_prompt is defined %} + {%- set add_generation_prompt = false %} +{%- endif %} +{%- if not preserved_thinking is defined %} + {%- if not tools %} + {%- set preserved_thinking = false %} + {%- else %} + {%- set preserved_thinking = true %} + {%- endif %} +{%- endif %} +{%- if not is_training is defined %} + {%- set is_training = false %} +{%- endif %} + +{%- if not reasoning_effort is defined %} + {%- set reasoning_effort = 'no_think' %} +{%- elif reasoning_effort not in ['high', 'low', 'no_think'] %} + {%- if reasoning_effort is none %} + {{- raise_exception('reasoning_effort error : None, should be no_think/low/high') }} + {%- else %} + {{- raise_exception('reasoning_effort error : ' + reasoning_effort + ', should be no_think/low/high') }} + {%- endif %} +{%- endif %} + +{%- if fallback_strategy is defined and fallback_strategy == 'reasoning_toolcall_retry' %} + {%- set reasoning_effort = 'high' %} + {%- set add_generation_prompt = false %} +{%- endif %} +{%- if not raw_last_assistant is defined %} + {%- set raw_last_assistant = false %} +{%- endif %} + +{%- macro visible_text(content) -%} + {%- if content is string -%} + {{- content }} + {%- elif content is iterable and content is not mapping -%} + {%- for item in content -%} + {%- if item is mapping and item.type == 'text' -%} + {{- item.text }} + {%- elif item is string -%} + {{- item }} + {%- endif -%} + {%- endfor -%} + {%- elif content is none -%} + {{- '' }} + {%- else -%} + {{- content }} + {%- endif -%} +{%- endmacro -%} + +{%- set ns = namespace(last_user_index=-1) %} +{%- set sp_ns = namespace(system_prompt='', is_first_sp=true) %} +{%- for message in messages %} + {%- if message['role'] == 'system' %} + {%- set sp_ns.system_prompt = sp_ns.system_prompt + visible_text(message['content']) %} + {%- endif %} + {%- if message['role'] == 'user' %} + {%- set ns.last_user_index = loop.index0 %} + {%- endif %} +{%- endfor %} +{%- if reasoning_effort is defined and reasoning_effort is string and reasoning_effort != '' and not tools %} + {%- set sp_ns.system_prompt = sp_ns.system_prompt + reasoning_mode_token + 'reasoning_effort:' + reasoning_effort %} +{%- endif %} +{{- bos_token }} +{{- sp_ns.system_prompt }} +{%- if tools %} + {%- if sp_ns.system_prompt != '' %} + {{- '\n\n# Tools\n\nYou may call one or more functions to assist with the user query.' }} + {%- else %} + {{- '# Tools\n\nYou may call one or more functions to assist with the user query.' }} + {%- endif %} + {{- '\n\nYou are provided with function signatures within XML tags:' }} + {{- '\n\n' }} + {%- for tool in tools %} + {%- if loop.index0 > 0 %} + {{- '\n' }} + {%- endif %} + {{- tool | tojson }} + {%- endfor %} + {{- '\n\n\n' }} + {{- 'For function call returns, you should first print ' + toolcalls_begin_token + '\n' }} + {{- 'For each function call, you should return object like:\n' }} + {{- toolcall_begin_token + '{function-name}' + toolsep_token + '\n' }} + {{- argkey_begin_token + '{arg-key-1}' + argkey_end_token + '\n' }} + {{- argvalue_begin_token + '{arg-value-1}' + argvalue_end_token + '\n' }} + {{- argkey_begin_token + '{arg-key-2}' + argkey_end_token + '\n' }} + {{- argvalue_begin_token + '{arg-value-2}' + argvalue_end_token + '\n' }} + {{- '...\n' }} + {{- toolcall_end_token + '\n' }} + {%- if reasoning_effort is defined and reasoning_effort is string and reasoning_effort != '' %} + {{- 'At the end of function call returns, you should print ' + toolcalls_end_token + reasoning_mode_token + 'reasoning_effort:' + reasoning_effort }} + {%- else %} + {{- 'At the end of function call returns, you should print ' + toolcalls_end_token }} + {%- endif %} +{%- endif %} + +{%- set prev_ns = namespace(is_tool=false, is_tool_first=true) %} +{%- set last_ns = namespace(last_is_assistant=false) %} +{%- for message in messages %} + {%- if message['role'] == 'user' %} + {%- if prev_ns.is_tool %} + {{- toolresponses_end_token }} + {%- endif %} + {{- user_token + visible_text(message['content']) }} + {%- set prev_ns.is_tool = false %} + {%- endif %} + {%- if message['role'] == 'assistant' %} + {%- if is_training %} + {%- if 'reasoning_content' in message and message['reasoning_content'] is string %} + {%- set rc = message['reasoning_content'] %} + {%- elif 'reasoning' in message and message['reasoning'] is string %} + {%- set rc = message['reasoning'] %} + {%- else %} + {%- set rc = none %} + {%- endif %} + {%- if rc is not none %} + {%- set content = think_begin_token + rc + think_end_token + visible_text(message['content']) %} + {%- else %} + {%- set content = think_begin_token + think_end_token + visible_text(message['content']) %} + {%- endif %} + {%- else %} + {%- if ((preserved_thinking is defined and preserved_thinking) or loop.index0 > ns.last_user_index) %} + {%- if 'reasoning_content' in message and message['reasoning_content'] is string %} + {%- set rc = message['reasoning_content'] %} + {%- elif 'reasoning' in message and message['reasoning'] is string %} + {%- set rc = message['reasoning'] %} + {%- else %} + {%- set rc = none %} + {%- endif %} + {%- if rc is not none %} + {%- set content = think_begin_token + rc + think_end_token + visible_text(message['content']) %} + {%- else %} + {%- set content = think_begin_token + think_end_token + visible_text(message['content']) %} + {%- endif %} + {%- else %} + {%- set content = think_begin_token + think_end_token + visible_text(message['content']) %} + {%- endif %} + {%- endif %} + {%- if prev_ns.is_tool %} + {{- toolresponses_end_token }} + {%- endif %} + {{- assistant_token }} + {%- if message['tool_calls'] is defined and message['tool_calls'] %} + {%- set prev_ns.is_tool_first = true %} + {{- content }} + {{- toolcalls_begin_token + '\n' }} + {%- for tool in message['tool_calls'] %} + {%- set arguments = tool['function']['arguments'] %} + {{- toolcall_begin_token + tool['function']['name'] + toolsep_token + '\n' }} + {%- for key, value in arguments.items() %} + {{- argkey_begin_token + key + argkey_end_token + '\n' }} + {%- if value is not string %} + {%- set value = value | tojson(ensure_ascii=False) %} + {%- endif %} + {{- argvalue_begin_token + value + argvalue_end_token + '\n' }} + {%- endfor %} + {{- toolcall_end_token + '\n' }} + {%- endfor %} + {{- toolcalls_end_token + eos_token }} + {%- else %} + {%- if loop.last and raw_last_assistant %} + {{- visible_text(message['content']) }} + {%- elif not loop.last or is_training %} + {{- content + eos_token }} + {%- else %} + {{- content }} + {%- endif %} + {%- endif %} + {%- set prev_ns.is_tool = false %} + {%- endif %} + {%- if message['role'] == 'tool' %} + {%- set prev_ns.is_tool = true %} + {%- if prev_ns.is_tool_first %} + {{- toolresponses_begin_token + '\n' }} + {%- set prev_ns.is_tool_first = false %} + {%- endif %} + {{- toolresponse_begin_token + '\n' + visible_text(message['content']) + '\n' + toolresponse_end_token + '\n' }} + {%- endif %} + {%- if loop.last and message['role'] == 'assistant' %} + {%- set last_ns.last_is_assistant = true %} + {%- endif %} + +{%- endfor %} +{%- if prev_ns.is_tool %} + {{- toolresponses_end_token }} +{%- endif %} +{%- if add_generation_prompt %} + {%- if not last_ns.last_is_assistant %} + {%- if reasoning_effort is defined and reasoning_effort in ['low', 'high'] %} + {{- assistant_token + think_begin_token }} + {%- elif reasoning_effort is defined and reasoning_effort == 'no_think' %} + {{- assistant_token + think_begin_token + think_end_token }} + {%- else %} + {{- assistant_token }} + {%- endif %} + {%- endif %} +{%- endif %} \ No newline at end of file diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index b890e66fc..72968607d 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -113,6 +113,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" }, { LLM_ARCH_HUNYUAN_DENSE, "hunyuan-dense" }, { LLM_ARCH_HUNYUAN_VL, "hunyuan_vl" }, + { LLM_ARCH_HY_V3, "hy_v3" }, { LLM_ARCH_SMOLLM3, "smollm3" }, { LLM_ARCH_OPENAI_MOE, "gpt-oss" }, { LLM_ARCH_LFM2, "lfm2" }, diff --git a/src/llama-arch.h b/src/llama-arch.h index a4f5091e7..b74d53af4 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -118,6 +118,7 @@ enum llm_arch { LLM_ARCH_HUNYUAN_MOE, LLM_ARCH_HUNYUAN_DENSE, LLM_ARCH_HUNYUAN_VL, + LLM_ARCH_HY_V3, LLM_ARCH_SMOLLM3, LLM_ARCH_OPENAI_MOE, LLM_ARCH_LFM2, diff --git a/src/llama-model.cpp b/src/llama-model.cpp index d87481381..eaf3f35d2 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -262,6 +262,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params return new llama_model_hunyuan_vl(params); case LLM_ARCH_HUNYUAN_DENSE: return new llama_model_hunyuan_dense(params); + case LLM_ARCH_HY_V3: + return new llama_model_hy_v3(params); case LLM_ARCH_SMOLLM3: return new llama_model_smollm3(params); case LLM_ARCH_OPENAI_MOE: @@ -2169,7 +2171,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, filter = [&](uint32_t il) { return il >= hparams.n_layer(); }; } - if (arch == LLM_ARCH_STEP35 && hparams.n_layer_nextn > 0) { + if ((arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_HY_V3) && hparams.n_layer_nextn > 0) { if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP) { filter = [&](uint32_t il) { return il >= hparams.n_layer(); }; } else { @@ -2525,6 +2527,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_JAIS2: case LLM_ARCH_OPENAI_MOE: case LLM_ARCH_HUNYUAN_DENSE: + case LLM_ARCH_HY_V3: case LLM_ARCH_LFM2: case LLM_ARCH_LFM2MOE: case LLM_ARCH_SMALLTHINKER: diff --git a/src/models/hy-v3.cpp b/src/models/hy-v3.cpp new file mode 100644 index 000000000..47a0beaf2 --- /dev/null +++ b/src/models/hy-v3.cpp @@ -0,0 +1,390 @@ +#include "models.h" + +void llama_model_hy_v3::load_arch_hparams(llama_model_loader & ml) { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false); + ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false); + + // HY V3 uses a sigmoid router with expert selection bias by default + if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) { + hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID; + } + + // NextN/MTP (HY V3): extra decoder block(s) appended beyond the main stack + ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false); + GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_all"); + + switch (hparams.n_layer()) { + case 48: type = LLM_TYPE_30B_A3B; break; + default: type = LLM_TYPE_UNKNOWN; + } +} + +void llama_model_hy_v3::load_arch_tensors(llama_model_loader & ml) { + LLAMA_LOAD_LOCALS; + + const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr); + // Trunk-only: the GGUF declares MTP layers in metadata but the actual MTP + // tensors live in a separate file (e.g. user split target/draft). Mark + // MTP tensors NOT_REQUIRED so the trunk loads cleanly. + const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight"; + const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr); + const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0; + const int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0; + + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + auto load_block = [&](int i, int flags) { + auto & layer = layers[i]; + const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / (n_expert_used > 0 ? n_expert_used : 1); + const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags); + + create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, flags); + + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags); + + // dense FFN (leading dense blocks, first_k_dense_replace) + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, TENSOR_NOT_REQUIRED); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED); + + // MoE routed experts (sigmoid router + expert selection bias) + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, i), {n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED); + create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, TENSOR_NOT_REQUIRED); + + // shared expert (always active, no gate) + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, TENSOR_NOT_REQUIRED); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_shexp}, TENSOR_NOT_REQUIRED); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED); + }; + + for (int i = 0; i < n_layer; ++i) { + load_block(i, trunk_flags); + } + + // NextN/MTP block(s): a full hy_v3 decoder block plus the NextN projections. + for (int i = n_layer; i < n_layer_all; ++i) { + auto & layer = layers[i]; + + load_block(i, mtp_flags); + + layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, mtp_flags); + layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, mtp_flags); + layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, mtp_flags); + layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED); + layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED); + // hy_v3 stores the MTP block's trailing final_layernorm here (applied + // after the decoder block, before the shared LM head). + layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED); + } +} + +std::unique_ptr llama_model_hy_v3::build_arch_graph(const llm_graph_params & params) const { + if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) { + return std::make_unique(*this, params); + } + return std::make_unique(*this, params); +} + +llama_model_hy_v3::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k()); + GGML_ASSERT(n_embd_head == n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + ggml_tensor * inp_pos = build_inp_pos(); + auto * inp_attn = build_attn_inp_kv(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); + + // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass. + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, n_head, n_head_kv, il); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + + if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + if (model.layers[il].ffn_gate_inp == nullptr) { + // dense FFN (leading dense blocks) + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, model.layers[il].ffn_up_s, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s, + nullptr, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_dense_out", il); + } else { + // MoE routed experts (sigmoid gating + expert selection bias) + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, + hparams.expert_weights_norm, + hparams.expert_weights_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il, + nullptr, model.layers[il].ffn_gate_up_exps, + model.layers[il].ffn_up_exps_s, + model.layers[il].ffn_gate_exps_s, + model.layers[il].ffn_down_exps_s); + cb(moe_out, "ffn_moe_out", il); + + // shared expert (always active, no gate) + ggml_tensor * sh_out = build_ffn(cur, + model.layers[il].ffn_up_shexp, nullptr, model.layers[il].ffn_up_shexp_s, + model.layers[il].ffn_gate_shexp, nullptr, model.layers[il].ffn_gate_shexp_s, + model.layers[il].ffn_down_shexp, nullptr, model.layers[il].ffn_down_shexp_s, + nullptr, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(sh_out, "ffn_shared_out", il); + + cur = ggml_add(ctx0, moe_out, sh_out); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + inpL = cur; + } + + cur = build_norm(inpL, model.output_norm, nullptr, LLM_NORM_RMS, -1); + + // Post-final-norm hidden state: what the MTP draft head's hnorm consumes. + // vLLM feeds the target model's normed output states, and the MTP layer + // itself returns final_layernorm(h), so the chained state is post-norm. + cb(cur, "h_nextn", -1); + res->t_h_nextn = cur; + + if (!cparams.embeddings_nextn_masked && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur, model.output_s); + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +// LLM_GRAPH_TYPE_DECODER_MTP draft head for HY V3 (MoE). +// Semantics mirror vLLM's HYV3MultiTokenPredictorLayer (hy_v3_mtp.py): +// enorm(embed) + hnorm(prev_hidden) -> concat(e, h) -> eh_proj -> +// hy_v3 decoder block -> final_layernorm (stored as nextn.shared_head_norm) -> +// shared LM head (the main model's lm_head; the checkpoint has no separate +// MTP head or MTP embeddings). +llama_model_hy_v3::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params) + : llm_graph_context(params) { + GGML_ASSERT(hparams.n_layer_nextn > 0 && "HY_V3 MTP requires n_layer_nextn > 0"); + + const int64_t n_embd_head = hparams.n_embd_head_v(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k()); + GGML_ASSERT(n_embd_head == n_rot); + + const int il = hparams.n_layer() + cparams.nextn_layer_offset; + GGML_ASSERT(cparams.nextn_layer_offset >= 0 && + cparams.nextn_layer_offset < (int) hparams.n_layer_nextn && + "nextn_layer_offset out of range [0, n_layer_nextn)"); + const auto & layer = model.layers[il]; + + GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj"); + GGML_ASSERT(layer.nextn.enorm && "MTP block missing nextn.enorm"); + GGML_ASSERT(layer.nextn.hnorm && "MTP block missing nextn.hnorm"); + + auto inp = std::make_unique(hparams.n_embd); + + inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + ggml_set_input(inp->tokens); + + inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens); + ggml_set_input(inp->embd); + ggml_set_name(inp->embd, "mtp_h_input"); + + ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd; + + ggml_tensor * h_input = inp->embd; + ggml_tensor * tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens); + cb(tok_embd, "mtp_tok_embd", il); + + res->add_input(std::move(inp)); + + ggml_tensor * inp_pos = build_inp_pos(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * h_norm = build_norm(h_input, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il); + cb(h_norm, "mtp_hnorm", il); + + ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il); + cb(e_norm, "mtp_enorm", il); + + ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0); + cb(concat, "mtp_concat", il); + + ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat); + cb(cur, "mtp_eh_proj", il); + + ggml_tensor * inpSA = cur; + + // mtp_block: a full hy_v3 decoder layer (mirrors the trunk graph) + cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "mtp_attn_norm", il); + + { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur, n_embd_head, n_head, n_head_kv, il); + + Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il); + Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); + + cur = build_attn(inp_attn, + layer.wo, layer.wo_b, layer.wo_s, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "mtp_attn_out", il); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "mtp_ffn_inp", il); + + cur = build_norm(ffn_inp, layer.ffn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "mtp_ffn_norm", il); + + if (layer.ffn_gate_inp == nullptr) { + cur = build_ffn(cur, + layer.ffn_up, layer.ffn_up_b, layer.ffn_up_s, + layer.ffn_gate, layer.ffn_gate_b, layer.ffn_gate_s, + layer.ffn_down, layer.ffn_down_b, layer.ffn_down_s, + nullptr, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "mtp_ffn_dense_out", il); + } else { + ggml_tensor * moe_out = build_moe_ffn(cur, + layer.ffn_gate_inp, + layer.ffn_up_exps, + layer.ffn_gate_exps, + layer.ffn_down_exps, + layer.ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, + hparams.expert_weights_norm, + hparams.expert_weights_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il, + nullptr, layer.ffn_gate_up_exps, + layer.ffn_up_exps_s, + layer.ffn_gate_exps_s, + layer.ffn_down_exps_s); + cb(moe_out, "mtp_ffn_moe_out", il); + + ggml_tensor * sh_out = build_ffn(cur, + layer.ffn_up_shexp, nullptr, layer.ffn_up_shexp_s, + layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s, + layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s, + nullptr, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(sh_out, "mtp_ffn_shared_out", il); + + cur = ggml_add(ctx0, moe_out, sh_out); + cb(cur, "mtp_ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "mtp_post_ffn", il); + + // final_layernorm applied after the decoder block, before the shared head. + // The post-norm hidden state seeds the next MTP step (matches vLLM, where + // HYV3MultiTokenPredictorLayer returns final_layernorm(h)). + ggml_tensor * head_norm_w = layer.nextn.shared_head_norm + ? layer.nextn.shared_head_norm + : model.output_norm; + GGML_ASSERT(head_norm_w && "HY_V3 MTP: missing both nextn.shared_head_norm and output_norm"); + cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1); + + cb(cur, "h_nextn", -1); + res->t_h_nextn = cur; + + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + cb(cur, "mtp_shared_head_norm", -1); + + ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output; + ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s; + GGML_ASSERT(head_w && "HY_V3 MTP: missing LM head (nextn.shared_head_head or model.output)"); + cur = build_lora_mm(head_w, cur, head_s); + cb(cur, "result_output", -1); + + res->t_logits = cur; + ggml_build_forward_expand(gf, cur); +} diff --git a/src/models/models.h b/src/models/models.h index 7a52e7bc1..beab9f6bc 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -1729,6 +1729,22 @@ struct llama_model_hunyuan_moe : public llama_model_base { std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; +struct llama_model_hy_v3 : public llama_model_base { + llama_model_hy_v3(const struct llama_model_params & params) : llama_model_base(params) {} + void load_arch_hparams(llama_model_loader & ml) override; + void load_arch_tensors(llama_model_loader & ml) override; + + struct graph : public llm_graph_context { + graph(const llama_model & model, const llm_graph_params & params); + }; + + struct graph_mtp : public llm_graph_context { + graph_mtp(const llama_model & model, const llm_graph_params & params); + }; + + std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; +}; + struct llama_model_hunyuan_vl : public llama_model_base { llama_model_hunyuan_vl(const struct llama_model_params & params) : llama_model_base(params) {} diff --git a/tests/test-chat-auto-parser.cpp b/tests/test-chat-auto-parser.cpp index d15fdd2c0..78e42c65a 100644 --- a/tests/test-chat-auto-parser.cpp +++ b/tests/test-chat-auto-parser.cpp @@ -1944,6 +1944,9 @@ static void test_role_markers_all_templates(testing & t) { // MiniMax M2: ]~b]{user|ai} { "MiniMax-M2.jinja", "]~b]user", "]~b]ai" }, + // HunYuan V3: <｜hy_User:opensource｜> / <｜hy_Assistant:opensource｜> + { "tencent-Hy3.jinja", "<｜hy_User:opensource｜>", "<｜hy_Assistant:opensource｜>" }, + // Nemotron Nano v2: {User|Assistant}; assistant marker // is followed by a prefilled block that gets included. { "NVIDIA-Nemotron-Nano-v2.jinja", "User", "Assistant" }, diff --git a/tests/test-jinja.cpp b/tests/test-jinja.cpp index d8d1892a9..90bdbc445 100644 --- a/tests/test-jinja.cpp +++ b/tests/test-jinja.cpp @@ -1376,6 +1376,36 @@ static void test_string_methods(testing & t) { "bXnXna" ); + test_template(t, "string.format() auto numbering", + "{{ '<{}|{}>'.format(s, 42) }}", + {{"s", "hello"}}, + "" + ); + + test_template(t, "string.format() manual numbering", + "{{ '{1}-{0}-{1}'.format('a', 'b') }}", + json::object(), + "b-a-b" + ); + + test_template(t, "string.format() named fields", + "{{ '{name} is {age}'.format(name='Bob', age=7) }}", + json::object(), + "Bob is 7" + ); + + test_template(t, "string.format() escaped braces", + "{{ '{{}} {} {{x}}'.format('mid') }}", + json::object(), + "{} mid {x}" + ); + + test_template(t, "string.format() no fields", + "{{ 'plain'.format() }}", + json::object(), + "plain" + ); + test_template(t, "undefined|capitalize", "{{ arr|capitalize }}", json::object(), diff --git a/tests/test-llama-archs.cpp b/tests/test-llama-archs.cpp index f39abe773..2cdf35739 100644 --- a/tests/test-llama-archs.cpp +++ b/tests/test-llama-archs.cpp @@ -346,6 +346,7 @@ static bool moe_mandatory(const llm_arch arch) { case LLM_ARCH_ERNIE4_5: case LLM_ARCH_ERNIE4_5_MOE: case LLM_ARCH_HUNYUAN_MOE: + case LLM_ARCH_HY_V3: case LLM_ARCH_OPENAI_MOE: case LLM_ARCH_LFM2MOE: case LLM_ARCH_SMALLTHINKER: From 14d3ba45f3369e75a308212399cfada5d349883b Mon Sep 17 00:00:00 2001 From: Pasha Khosravi Date: Mon, 13 Jul 2026 21:52:00 -0700 Subject: [PATCH 03/24] metal : add Q2_0 support (#25419) --- ggml/src/ggml-metal/ggml-metal-device.cpp | 10 ++ ggml/src/ggml-metal/ggml-metal-device.m | 2 + ggml/src/ggml-metal/ggml-metal-impl.h | 3 + ggml/src/ggml-metal/ggml-metal-ops.cpp | 1 + ggml/src/ggml-metal/ggml-metal.metal | 197 ++++++++++++++++++++++ 5 files changed, 213 insertions(+) diff --git a/ggml/src/ggml-metal/ggml-metal-device.cpp b/ggml/src/ggml-metal/ggml-metal-device.cpp index 15290c3d1..870f93df0 100644 --- a/ggml/src/ggml-metal/ggml-metal-device.cpp +++ b/ggml/src/ggml-metal/ggml-metal-device.cpp @@ -805,6 +805,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta nsg = N_SG_Q1_0; nr0 = N_R0_Q1_0; } break; + case GGML_TYPE_Q2_0: + { + nsg = N_SG_Q2_0; + nr0 = N_R0_Q2_0; + } break; case GGML_TYPE_Q4_0: { nsg = N_SG_Q4_0; @@ -1029,6 +1034,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_id(ggml_m nsg = N_SG_Q1_0; nr0 = N_R0_Q1_0; } break; + case GGML_TYPE_Q2_0: + { + nsg = N_SG_Q2_0; + nr0 = N_R0_Q2_0; + } break; case GGML_TYPE_Q4_0: { nsg = N_SG_Q4_0; diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m index 5d29250f6..1dfe0bdd5 100644 --- a/ggml/src/ggml-metal/ggml-metal-device.m +++ b/ggml/src/ggml-metal/ggml-metal-device.m @@ -1289,6 +1289,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te case GGML_TYPE_BF16: case GGML_TYPE_Q8_0: case GGML_TYPE_Q1_0: + case GGML_TYPE_Q2_0: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -1316,6 +1317,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te return false; } case GGML_TYPE_Q1_0: + case GGML_TYPE_Q2_0: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h index d6761023b..a45f3ac67 100644 --- a/ggml/src/ggml-metal/ggml-metal-impl.h +++ b/ggml/src/ggml-metal/ggml-metal-impl.h @@ -24,6 +24,9 @@ #define N_R0_Q1_0 8 #define N_SG_Q1_0 2 +#define N_R0_Q2_0 8 +#define N_SG_Q2_0 2 + #define N_R0_Q4_0 4 #define N_SG_Q4_0 2 diff --git a/ggml/src/ggml-metal/ggml-metal-ops.cpp b/ggml/src/ggml-metal/ggml-metal-ops.cpp index 45909c477..805bc4093 100644 --- a/ggml/src/ggml-metal/ggml-metal-ops.cpp +++ b/ggml/src/ggml-metal/ggml-metal-ops.cpp @@ -2077,6 +2077,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) { op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16 || op->src[0]->type == GGML_TYPE_Q1_0 || + op->src[0]->type == GGML_TYPE_Q2_0 || op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_Q4_1 || op->src[0]->type == GGML_TYPE_Q5_0 || diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal index 6b6f9fd87..38919ebad 100644 --- a/ggml/src/ggml-metal/ggml-metal.metal +++ b/ggml/src/ggml-metal/ggml-metal.metal @@ -170,6 +170,39 @@ void dequantize_q1_0_t4(device const block_q1_0 * xb, short il, thread type4 & r reg = (type4) reg_f; } +template +void dequantize_q2_0(device const block_q2_0 * xb, short il, thread type4x4 & reg) { + device const uint8_t * qs = xb->qs; + const float d = xb->d; + + const int byte_offset = il * 4; // il*16 elements = il*4 bytes (4 elements per byte) + float4x4 reg_f; + + for (int i = 0; i < 4; i++) { + const uint8_t b = qs[byte_offset + i]; + reg_f[i][0] = ((float)((b >> 0) & 3) - 1.0f) * d; + reg_f[i][1] = ((float)((b >> 2) & 3) - 1.0f) * d; + reg_f[i][2] = ((float)((b >> 4) & 3) - 1.0f) * d; + reg_f[i][3] = ((float)((b >> 6) & 3) - 1.0f) * d; + } + + reg = (type4x4) reg_f; +} + +template +void dequantize_q2_0_t4(device const block_q2_0 * xb, short il, thread type4 & reg) { + const float d = xb->d; + const uint8_t b = xb->qs[il]; + + float4 reg_f; + reg_f[0] = ((float)((b >> 0) & 3) - 1.0f) * d; + reg_f[1] = ((float)((b >> 2) & 3) - 1.0f) * d; + reg_f[2] = ((float)((b >> 4) & 3) - 1.0f) * d; + reg_f[3] = ((float)((b >> 6) & 3) - 1.0f) * d; + + reg = (type4) reg_f; +} + template void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) { device const uint16_t * qs = ((device const uint16_t *)xb + 1); @@ -221,6 +254,27 @@ void quantize_q1_0(device const float * src, device block_q1_0 & dst) { } } +void quantize_q2_0(device const float * src, device block_q2_0 & dst) { + float amax = 0.0f; + for (int j = 0; j < QK2_0; j++) { + float a = fabs(src[j]); + if (a > amax) amax = a; + } + const float d = amax; + dst.d = d; + + const float id = d > 0.0f ? 1.0f / d : 0.0f; + + for (int j = 0; j < QK2_0 / 4; j++) { + dst.qs[j] = 0; + } + for (int j = 0; j < QK2_0; j++) { + int q = (int)round(src[j] * id) + 1; + q = max(0, min(3, q)); + dst.qs[j / 4] |= (q << (2 * (j % 4))); + } +} + void quantize_q4_0(device const float * src, device block_q4_0 & dst) { #pragma METAL fp math_mode(safe) float amax = 0.0f; // absolute max @@ -3289,6 +3343,55 @@ inline float block_q_n_dot_y(device const block_q1_0 * qb_curr, float sumy, thre return qb_curr->d * (2.0f * acc - sumy); } +// Q2_0 dot: d * (sum_lo(y) + 2*sum_hi(y) - sumy) via per-bit conditional adds +inline float block_q_n_dot_y(device const block_q2_0 * qb_curr, float sumy, thread float * yl, int il) { + device const uint8_t * qs = qb_curr->qs + (il / 4); + const uint8_t b0 = qs[0]; + const uint8_t b1 = qs[1]; + const uint8_t b2 = qs[2]; + const uint8_t b3 = qs[3]; + + // Accumulate where low bit is set (bits 0,2,4,6 of each byte) + float acc_lo = 0.0f; + acc_lo += select(0.0f, yl[ 0], bool(b0 & 0x01)); + acc_lo += select(0.0f, yl[ 1], bool(b0 & 0x04)); + acc_lo += select(0.0f, yl[ 2], bool(b0 & 0x10)); + acc_lo += select(0.0f, yl[ 3], bool(b0 & 0x40)); + acc_lo += select(0.0f, yl[ 4], bool(b1 & 0x01)); + acc_lo += select(0.0f, yl[ 5], bool(b1 & 0x04)); + acc_lo += select(0.0f, yl[ 6], bool(b1 & 0x10)); + acc_lo += select(0.0f, yl[ 7], bool(b1 & 0x40)); + acc_lo += select(0.0f, yl[ 8], bool(b2 & 0x01)); + acc_lo += select(0.0f, yl[ 9], bool(b2 & 0x04)); + acc_lo += select(0.0f, yl[10], bool(b2 & 0x10)); + acc_lo += select(0.0f, yl[11], bool(b2 & 0x40)); + acc_lo += select(0.0f, yl[12], bool(b3 & 0x01)); + acc_lo += select(0.0f, yl[13], bool(b3 & 0x04)); + acc_lo += select(0.0f, yl[14], bool(b3 & 0x10)); + acc_lo += select(0.0f, yl[15], bool(b3 & 0x40)); + + // Accumulate where high bit is set (bits 1,3,5,7 of each byte) + float acc_hi = 0.0f; + acc_hi += select(0.0f, yl[ 0], bool(b0 & 0x02)); + acc_hi += select(0.0f, yl[ 1], bool(b0 & 0x08)); + acc_hi += select(0.0f, yl[ 2], bool(b0 & 0x20)); + acc_hi += select(0.0f, yl[ 3], bool(b0 & 0x80)); + acc_hi += select(0.0f, yl[ 4], bool(b1 & 0x02)); + acc_hi += select(0.0f, yl[ 5], bool(b1 & 0x08)); + acc_hi += select(0.0f, yl[ 6], bool(b1 & 0x20)); + acc_hi += select(0.0f, yl[ 7], bool(b1 & 0x80)); + acc_hi += select(0.0f, yl[ 8], bool(b2 & 0x02)); + acc_hi += select(0.0f, yl[ 9], bool(b2 & 0x08)); + acc_hi += select(0.0f, yl[10], bool(b2 & 0x20)); + acc_hi += select(0.0f, yl[11], bool(b2 & 0x80)); + acc_hi += select(0.0f, yl[12], bool(b3 & 0x02)); + acc_hi += select(0.0f, yl[13], bool(b3 & 0x08)); + acc_hi += select(0.0f, yl[14], bool(b3 & 0x20)); + acc_hi += select(0.0f, yl[15], bool(b3 & 0x80)); + + return qb_curr->d * (acc_lo + 2.0f * acc_hi - sumy); +} + // function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i]) // il indicates where the q4 quants begin (0 or QK4_0/4) // we assume that the yl's have been multiplied with the appropriate scale factor @@ -3592,6 +3695,86 @@ kernel void kernel_mul_mv_q1_0_f32( kernel_mul_mv_q1_0_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } +template +void kernel_mul_mv_q2_0_f32_impl( + args_t args, + device const char * src0, + device const char * src1, + device char * dst, + threadgroup char * shmem, + uint3 tgpig, + ushort tiisg, + ushort sgitg) { + const short NSG = FC_mul_mv_nsg; + + const int nb = args.ne00/QK2_0; + + const int r0 = tgpig.x; + const int r1 = tgpig.y; + const int im = tgpig.z; + + const int first_row = (r0 * NSG + sgitg) * nr0; + + const uint i12 = im%FC_mul_mv_ne12; + const uint i13 = im/FC_mul_mv_ne12; + + const uint64_t offset1 = r1*args.nb11 + (i12)*args.nb12 + (i13)*args.nb13; + + device const float * y = (device const float *) (src1 + offset1); + + device const block_q2_0 * ax[nr0]; + for (int row = 0; row < nr0; ++row) { + const uint64_t offset0 = (first_row + row)*args.nb01 + (i12/FC_mul_mv_r2)*args.nb02 + (i13/FC_mul_mv_r3)*args.nb03; + ax[row] = (device const block_q2_0 *) ((device char *) src0 + offset0); + } + + float yl[16]; + float sumf[nr0] = {0.f}; + + // group 64: 4 sub-blocks of 16 weights per Q2_0 block + const short ix = (tiisg/4); + const short il = (tiisg%4)*16; + + device const float * yb = y + ix*QK2_0 + il; + + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) { + float sumy = 0.f; + + FOR_UNROLL (short i = 0; i < 16; i++) { + yl[i] = yb[i]; + sumy += yb[i]; + } + + FOR_UNROLL (short row = 0; row < nr0; row++) { + sumf[row] += block_q_n_dot_y(ax[row] + ib, sumy, yl, il); + } + + yb += QK2_0 * (N_SIMDWIDTH/4); + } + + device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; + + for (int row = 0; row < nr0; ++row) { + const float tot = simd_sum(sumf[row]); + + if (tiisg == 0 && first_row + row < args.ne01) { + dst_f32[first_row + row] = tot; + } + } +} + +[[host_name("kernel_mul_mv_q2_0_f32")]] +kernel void kernel_mul_mv_q2_0_f32( + constant ggml_metal_kargs_mul_mv & args, + device const char * src0, + device const char * src1, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]) { + kernel_mul_mv_q2_0_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); +} + kernel void kernel_mul_mv_q4_0_f32( constant ggml_metal_kargs_mul_mv & args, device const char * src0, @@ -3989,6 +4172,11 @@ template [[host_name("kernel_mul_mv_ext_q1_0_f32_r1_3")]] kernel mul_mv_ext_q4 template [[host_name("kernel_mul_mv_ext_q1_0_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q1_0, 128, dequantize_q1_0_t4>; template [[host_name("kernel_mul_mv_ext_q1_0_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q1_0, 128, dequantize_q1_0_t4>; +template [[host_name("kernel_mul_mv_ext_q2_0_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q2_0, 64, dequantize_q2_0_t4>; +template [[host_name("kernel_mul_mv_ext_q2_0_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q2_0, 64, dequantize_q2_0_t4>; +template [[host_name("kernel_mul_mv_ext_q2_0_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q2_0, 64, dequantize_q2_0_t4>; +template [[host_name("kernel_mul_mv_ext_q2_0_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q2_0, 64, dequantize_q2_0_t4>; + template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q4_0, 32, dequantize_q4_0_t4>; template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q4_0, 32, dequantize_q4_0_t4>; template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q4_0, 32, dequantize_q4_0_t4>; @@ -7700,6 +7888,7 @@ typedef decltype(kernel_cpy_f32_q) cpy_f_q_ template [[host_name("kernel_cpy_f32_q8_0")]] kernel cpy_f_q_t kernel_cpy_f32_q; template [[host_name("kernel_cpy_f32_q1_0")]] kernel cpy_f_q_t kernel_cpy_f32_q; +template [[host_name("kernel_cpy_f32_q2_0")]] kernel cpy_f_q_t kernel_cpy_f32_q; template [[host_name("kernel_cpy_f32_q4_0")]] kernel cpy_f_q_t kernel_cpy_f32_q; template [[host_name("kernel_cpy_f32_q4_1")]] kernel cpy_f_q_t kernel_cpy_f32_q; template [[host_name("kernel_cpy_f32_q5_0")]] kernel cpy_f_q_t kernel_cpy_f32_q; @@ -7745,6 +7934,7 @@ kernel void kernel_cpy_q_f32( typedef decltype(kernel_cpy_q_f32) cpy_q_f_t; template [[host_name("kernel_cpy_q1_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; +template [[host_name("kernel_cpy_q2_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; template [[host_name("kernel_cpy_q4_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; template [[host_name("kernel_cpy_q4_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; template [[host_name("kernel_cpy_q5_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; @@ -7752,6 +7942,7 @@ template [[host_name("kernel_cpy_q5_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32< template [[host_name("kernel_cpy_q8_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; template [[host_name("kernel_cpy_q1_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32; +template [[host_name("kernel_cpy_q2_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32; template [[host_name("kernel_cpy_q4_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32; template [[host_name("kernel_cpy_q4_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32; template [[host_name("kernel_cpy_q5_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32; @@ -9596,6 +9787,7 @@ template [[host_name("kernel_get_rows_bf16")]] kernel get_rows_f_t kernel_get_ro typedef decltype(kernel_get_rows_q) get_rows_q_t; template [[host_name("kernel_get_rows_q1_0")]] kernel get_rows_q_t kernel_get_rows_q; +template [[host_name("kernel_get_rows_q2_0")]] kernel get_rows_q_t kernel_get_rows_q; template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_q_t kernel_get_rows_q; template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_q_t kernel_get_rows_q; template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_q_t kernel_get_rows_q; @@ -10466,6 +10658,7 @@ template [[host_name("kernel_mul_mm_f16_f32")]] kernel mul_mm_t kernel_mul_m template [[host_name("kernel_mul_mm_bf16_f32")]] kernel mul_mm_t kernel_mul_mm; #endif template [[host_name("kernel_mul_mm_q1_0_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q2_0_f32")]] kernel mul_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mul_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mul_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mul_mm_t kernel_mul_mm; @@ -10490,6 +10683,7 @@ template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mul_mm_t kernel_mul_m template [[host_name("kernel_mul_mm_f32_f16")]] kernel mul_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_f16_f16")]] kernel mul_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q1_0_f16")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q2_0_f16")]] kernel mul_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q4_0_f16")]] kernel mul_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q4_1_f16")]] kernel mul_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q5_0_f16")]] kernel mul_mm_t kernel_mul_mm; @@ -10523,6 +10717,7 @@ template [[host_name("kernel_mul_mm_id_f16_f32")]] kernel mul_mm_id kernel_m template [[host_name("kernel_mul_mm_id_bf16_f32")]] kernel mul_mm_id kernel_mul_mm_id; #endif template [[host_name("kernel_mul_mm_id_q1_0_f32")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q2_0_f32")]] kernel mul_mm_id kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mul_mm_id kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mul_mm_id kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mul_mm_id kernel_mul_mm_id; @@ -10547,6 +10742,7 @@ template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mul_mm_id kernel_m template [[host_name("kernel_mul_mm_id_f32_f16")]] kernel mul_mm_id kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_f16_f16")]] kernel mul_mm_id kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_q1_0_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q2_0_f16")]] kernel mul_mm_id kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_q4_0_f16")]] kernel mul_mm_id kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_q4_1_f16")]] kernel mul_mm_id kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_q5_0_f16")]] kernel mul_mm_id kernel_mul_mm_id; @@ -10702,6 +10898,7 @@ template [[host_name("kernel_mul_mv_id_bf16_f32_4")]] kernel kernel_mul_mv_id_4 template [[host_name("kernel_mul_mv_id_q8_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; template [[host_name("kernel_mul_mv_id_q1_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_q2_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; template [[host_name("kernel_mul_mv_id_q4_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; template [[host_name("kernel_mul_mv_id_q4_1_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; template [[host_name("kernel_mul_mv_id_q5_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; From c1063ac9d75fe9238c82de94fc4b1f58b2cc6196 Mon Sep 17 00:00:00 2001 From: Titaniumtown Date: Tue, 14 Jul 2026 05:00:00 -0400 Subject: [PATCH 04/24] sycl: set fattn_vec_nthreads to 256 for Battlemage (#25205) Currently detects lunarlake + battlemage / xe2 and sets the value to 256. Keeps default at 128, Intel's ARC Alchemist's prefered value. --- ggml/src/ggml-sycl/fattn-vec.hpp | 46 +++++++++++++++++++------------- 1 file changed, 27 insertions(+), 19 deletions(-) diff --git a/ggml/src/ggml-sycl/fattn-vec.hpp b/ggml/src/ggml-sycl/fattn-vec.hpp index 8031acfdf..04baac441 100644 --- a/ggml/src/ggml-sycl/fattn-vec.hpp +++ b/ggml/src/ggml-sycl/fattn-vec.hpp @@ -15,13 +15,11 @@ namespace syclex = sycl::ext::oneapi::experimental; -static int ggml_sycl_fattn_vec_get_nthreads_host(const int cc) { - return 128; - GGML_UNUSED(cc); -} - -static constexpr int ggml_sycl_fattn_vec_get_nthreads_device() { - return 128; +static int ggml_sycl_fattn_vec_get_nthreads_device(gpu_arch arch) { + // Xe2 (Battlemage, Lunar Lake) runs the flash-attention vec kernel best with a 256-thread work group. + return (arch == gpu_arch::intel_gpu_bmg_g21 || + arch == gpu_arch::intel_gpu_bmg_g31 || + arch == gpu_arch::intel_gpu_lnl_m) ? 256 : 128; } // Currenlty llvm with the amdgcn target dose not support unrolling loops @@ -36,7 +34,8 @@ template // D == head size + int warp_size, + int nthreads> // D == head size static void flash_attn_ext_vec(const char* __restrict__ Q, const char* __restrict__ K, const char* __restrict__ V, @@ -99,7 +98,6 @@ static void flash_attn_ext_vec(const char* __restrict__ Q, constexpr int nthreads_KQ_q = (D/4 < warp_size ? D/4 : warp_size); constexpr int nthreads_V_q = (D/4 < warp_size ? D/4 : warp_size); - constexpr int nthreads = ggml_sycl_fattn_vec_get_nthreads_device(); constexpr int nthreads_KQ = type_K == GGML_TYPE_F16 ? 128 / cpy_nb : nthreads_KQ_q; constexpr int nthreads_V = type_V == GGML_TYPE_F16 ? 128 / cpy_nb : nthreads_V_q; @@ -581,24 +579,34 @@ static void flash_attn_ext_vec(const char* __restrict__ Q, #endif // __clang__ + template void ggml_sycl_flash_attn_ext_vec_case_impl(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - const int warp_size = WARP_16_SIZE; //better performance than WARP_32_SIZE - - const int cc = ggml_sycl_info().devices[ggml_sycl_get_device()].cc; - - const int nthreads = ggml_sycl_fattn_vec_get_nthreads_host(cc); - const int nwarps = nthreads / warp_size; + constexpr int warp_size = WARP_16_SIZE; //better performance than WARP_32_SIZE const bool need_f16_K = type_K == GGML_TYPE_F16; const bool need_f16_V = type_V == GGML_TYPE_F16; constexpr size_t nbytes_shared = 0; - launch_fattn, warp_size>( - ctx, dst, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false); + const auto arch = ggml_sycl_info().devices[ctx.device].hw_info.arch; + const int nthreads = ggml_sycl_fattn_vec_get_nthreads_device(arch); + // 256 threads would overflow the 64 KB work-group local memory at D == 512, so keep 128 there. + if (D <= 256 && nthreads == 256) { + constexpr int nthreads_hw = 256; + constexpr int nwarps = nthreads_hw / warp_size; + launch_fattn, warp_size>( + ctx, dst, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false); + } else { + constexpr int nthreads_hw = 128; + constexpr int nwarps = nthreads_hw / warp_size; + launch_fattn, warp_size>( + ctx, dst, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false); + } } template From ec0dbef81632bb4233971fe83d91cf0accbfd882 Mon Sep 17 00:00:00 2001 From: Christian Kastner Date: Tue, 14 Jul 2026 11:03:22 +0200 Subject: [PATCH 05/24] arg: Flush log before exiting after usage() (#25504) Under certain conditions, it's possible for messages emitted via LOG() to get lost before exit, apparently because they are emitted by another thread. common_params_print_usage() uses printf directly, and is not affected. Flushing the log before exit seems to resolve this. --- common/arg.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/common/arg.cpp b/common/arg.cpp index 71118b308..9aefcd202 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -1077,6 +1077,7 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e if (ctx_arg.print_usage) { ctx_arg.print_usage(argc, argv); } + common_log_flush(common_log_main()); exit(0); } if (ctx_arg.params.completion) { From cb489bc0fb789c2cb7a9cc9dc44fa71893fe0988 Mon Sep 17 00:00:00 2001 From: Thiago Padilha Date: Tue, 14 Jul 2026 06:43:15 -0300 Subject: [PATCH 06/24] convert_hf_to_gguf: support split MTP export for HY V3 (#25641) - Add a supports_mtp_export capability to ModelBase so architectures can opt into --mtp and --no-mtp without extending a central class allowlist. - Enable the capability for the existing Qwen3.5/3.6 and Step3.5/3.7 implementations, and for HY V3, whose converter already supports filtering the appended MTP layers. --- conversion/base.py | 4 +++- conversion/hunyuan.py | 1 + conversion/qwen.py | 1 + conversion/step3.py | 1 + convert_hf_to_gguf.py | 6 ++---- 5 files changed, 8 insertions(+), 5 deletions(-) diff --git a/conversion/base.py b/conversion/base.py index 0421aa4bc..1b85ef0a1 100644 --- a/conversion/base.py +++ b/conversion/base.py @@ -109,7 +109,9 @@ class ModelBase: sentence_transformers_dense_modules: bool = False # MTP (multi-token prediction) export modes; set by main() before instantiation. - # Architectures opt in by overriding the handling (see _Qwen35MtpMixin). + # Architectures that implement the filtering/export behavior opt in by + # setting supports_mtp_export = True on their model class or a mixin. + supports_mtp_export: bool = False mtp_only: bool = False no_mtp: bool = False diff --git a/conversion/hunyuan.py b/conversion/hunyuan.py index 4d2545f8b..65d294fbe 100644 --- a/conversion/hunyuan.py +++ b/conversion/hunyuan.py @@ -361,6 +361,7 @@ class HunyuanVLTextModel(HunYuanModel): @ModelBase.register("HYV3ForCausalLM") class HYV3Model(TextModel): model_arch = gguf.MODEL_ARCH.HY_V3 + supports_mtp_export = True # Trunk layer count, stashed before indexing so the classmethod # filter_tensors can identify the appended MTP block(s) (mirrors diff --git a/conversion/qwen.py b/conversion/qwen.py index 0356bd2da..82d42fcc1 100644 --- a/conversion/qwen.py +++ b/conversion/qwen.py @@ -541,6 +541,7 @@ class _Qwen35MtpMixin: `mtp.*` to the standard layer-indexed nextn naming so the existing tensor_map handles them.""" + supports_mtp_export = True hparams: dict[str, Any] model_arch: gguf.MODEL_ARCH gguf_writer: gguf.GGUFWriter diff --git a/conversion/step3.py b/conversion/step3.py index 49bb5244a..f7cdc997e 100644 --- a/conversion/step3.py +++ b/conversion/step3.py @@ -98,6 +98,7 @@ class Step3VLTextModel(Qwen3Model): @ModelBase.register("Step3p5ForCausalLM", "Step3p7ForConditionalGeneration") class Step35Model(TextModel): model_arch = gguf.MODEL_ARCH.STEP35 + supports_mtp_export = True # The --mtp / --no-mtp toggles are ModelBase.mtp_only / no_mtp (set in # convert_hf_to_gguf.py main()). Unlike Qwen3.5, which stores MTP under a diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 3b23d5ebc..2c5e62a16 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -259,10 +259,8 @@ def main() -> None: sys.exit(1) if args.mtp or args.no_mtp: - from conversion.qwen import _Qwen35MtpMixin - from conversion.step3 import Step35Model - if not (issubclass(model_class, _Qwen35MtpMixin) or issubclass(model_class, Step35Model)): - logger.error("--mtp / --no-mtp are only supported for Qwen3.5/3.6 and Step3.5 text variants today") + if not model_class.supports_mtp_export: + logger.error("--mtp / --no-mtp are not supported for %s", model_architecture) sys.exit(1) if args.no_mtp: model_class.no_mtp = True From c9330ed0cf726be38308822ebbfada07f437c35d Mon Sep 17 00:00:00 2001 From: Pascal Date: Tue, 14 Jul 2026 12:05:40 +0200 Subject: [PATCH 07/24] ui: add reasoning effort control to mobile add sheet (#25539) The mobile "+" sheet was missing the reasoning effort section present in the desktop dropdown, so thinking could not be toggled on touch. Extract the shared derivation and selection logic into useReasoningMenu and consume it from both the desktop submenu and the mobile sheet, keeping a single source of truth and preserving each surface idiom. --- .../ChatFormActionAddReasoningSubmenu.svelte | 92 ++++-------------- .../ChatFormActionAddSheet.svelte | 69 ++++++++++++- .../lib/hooks/use-reasoning-menu.svelte.ts | 96 +++++++++++++++++++ 3 files changed, 184 insertions(+), 73 deletions(-) create mode 100644 tools/ui/src/lib/hooks/use-reasoning-menu.svelte.ts diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddReasoningSubmenu.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddReasoningSubmenu.svelte index 070fd3ac6..caea7a022 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddReasoningSubmenu.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddReasoningSubmenu.svelte @@ -2,85 +2,31 @@ import { Lightbulb, LightbulbOff, Check, Info } from '@lucide/svelte'; import * as DropdownMenu from '$lib/components/ui/dropdown-menu'; import * as Tooltip from '$lib/components/ui/tooltip'; - import { ReasoningEffort } from '$lib/enums'; - import { REASONING_EFFORT_TOKENS } from '$lib/constants/reasoning-effort-tokens'; - import { REASONING_EFFORT_LEVELS } from '$lib/constants/reasoning-effort'; - import type { ReasoningEffortLevel } from '$lib/types'; - import { - modelsStore, - checkModelSupportsThinking, - supportsThinking, - propsCacheVersion, - loadedModelIds - } from '$lib/stores/models.svelte'; - import { chatStore } from '$lib/stores/chat.svelte'; - import { conversationsStore, activeMessages } from '$lib/stores/conversations.svelte'; - import { isRouterMode } from '$lib/stores/server.svelte'; - import type { DatabaseMessage } from '$lib/types/database'; + import { useReasoningMenu } from '$lib/hooks/use-reasoning-menu.svelte'; let subOpen = $state(false); - let conversationModel = $derived( - chatStore.getConversationModel(activeMessages() as DatabaseMessage[]) - ); - - let modelSupportsThinkingFromMessages = $derived.by(() => { - const modelId = isRouterMode() ? modelsStore.selectedModelName || conversationModel : null; - if (!modelId) return false; - - const messages = conversationsStore.activeMessages; - - return messages.some( - (m) => m.role === 'assistant' && m.model === modelId && !!m.reasoningContent - ); - }); - - let modelSupportsThinking = $derived.by(() => { - loadedModelIds(); - propsCacheVersion(); - - if (isRouterMode()) { - const modelId = modelsStore.selectedModelName || conversationModel; - return checkModelSupportsThinking(modelId ?? '') || modelSupportsThinkingFromMessages; - } - - return supportsThinking() || modelSupportsThinkingFromMessages; - }); - - let thinkingEnabled = $derived(conversationsStore.getThinkingEnabled()); - let currentEffort = $derived(conversationsStore.getReasoningEffort()); - let isOff = $derived(!thinkingEnabled); - - function isSelected(item: ReasoningEffortLevel): boolean { - if (item.isOff) return isOff; - return thinkingEnabled && currentEffort === item.value; - } - - function handleSelection(item: ReasoningEffortLevel) { - if (item.isOff) { - conversationsStore.setThinkingEnabled(false); - } else { - conversationsStore.setThinkingEnabled(true); - conversationsStore.setReasoningEffort(item.value as ReasoningEffort); - } - subOpen = false; - } + const reasoning = useReasoningMenu(); -{#if modelSupportsThinking} +{#if reasoning.modelSupportsThinking} - {#if thinkingEnabled} + {#if reasoning.thinkingEnabled} {:else} {/if} - + Reasoning - {thinkingEnabled ? currentEffort : 'off'} + {reasoning.thinkingEnabled ? reasoning.currentEffort : 'off'} @@ -88,14 +34,18 @@ - {#each REASONING_EFFORT_LEVELS as level (level.value)} + {#each reasoning.levels as level (level.value)} + {@const tokenLabel = reasoning.tokenLabel(level)} + {/each} + + + + {/if} + (filesExpanded = open)}> {#if filesExpanded} diff --git a/tools/ui/src/lib/hooks/use-reasoning-menu.svelte.ts b/tools/ui/src/lib/hooks/use-reasoning-menu.svelte.ts new file mode 100644 index 000000000..6a459e66b --- /dev/null +++ b/tools/ui/src/lib/hooks/use-reasoning-menu.svelte.ts @@ -0,0 +1,96 @@ +import { ReasoningEffort } from '$lib/enums'; +import { REASONING_EFFORT_LEVELS } from '$lib/constants/reasoning-effort'; +import { REASONING_EFFORT_TOKENS } from '$lib/constants/reasoning-effort-tokens'; +import type { ReasoningEffortLevel } from '$lib/types'; +import type { DatabaseMessage } from '$lib/types/database'; +import { + modelsStore, + checkModelSupportsThinking, + supportsThinking, + propsCacheVersion, + loadedModelIds +} from '$lib/stores/models.svelte'; +import { chatStore } from '$lib/stores/chat.svelte'; +import { conversationsStore, activeMessages } from '$lib/stores/conversations.svelte'; +import { isRouterMode } from '$lib/stores/server.svelte'; + +export interface UseReasoningMenuReturn { + readonly modelSupportsThinking: boolean; + readonly thinkingEnabled: boolean; + readonly currentEffort: ReasoningEffort; + readonly levels: ReasoningEffortLevel[]; + isSelected(level: ReasoningEffortLevel): boolean; + tokenLabel(level: ReasoningEffortLevel): string | null; + select(level: ReasoningEffortLevel): void; +} + +/** + * Shared reactive state and helpers for the reasoning effort menu. + * + * Used by both the desktop dropdown (`ChatFormActionAddReasoningSubmenu`) + * and the mobile sheet (`ChatFormActionAddSheet`) to avoid duplicating the + * thinking-support derivation and the effort selection logic. + */ +export function useReasoningMenu(): UseReasoningMenuReturn { + const conversationModel = $derived( + chatStore.getConversationModel(activeMessages() as DatabaseMessage[]) + ); + + // a router chat can carry reasoning from an earlier turn before the props + // cache is primed, so a model that already produced thinking still qualifies + const modelSupportsThinkingFromMessages = $derived.by(() => { + const modelId = isRouterMode() ? modelsStore.selectedModelName || conversationModel : null; + if (!modelId) return false; + + return conversationsStore.activeMessages.some( + (m) => m.role === 'assistant' && m.model === modelId && !!m.reasoningContent + ); + }); + + const modelSupportsThinking = $derived.by(() => { + loadedModelIds(); + propsCacheVersion(); + + if (isRouterMode()) { + const modelId = modelsStore.selectedModelName || conversationModel; + return checkModelSupportsThinking(modelId ?? '') || modelSupportsThinkingFromMessages; + } + + return supportsThinking() || modelSupportsThinkingFromMessages; + }); + + const thinkingEnabled = $derived(conversationsStore.getThinkingEnabled()); + const currentEffort = $derived(conversationsStore.getReasoningEffort()); + + return { + get modelSupportsThinking() { + return modelSupportsThinking; + }, + get thinkingEnabled() { + return thinkingEnabled; + }, + get currentEffort() { + return currentEffort; + }, + get levels() { + return REASONING_EFFORT_LEVELS; + }, + isSelected(level: ReasoningEffortLevel): boolean { + if (level.isOff) return !thinkingEnabled; + return thinkingEnabled && currentEffort === level.value; + }, + tokenLabel(level: ReasoningEffortLevel): string | null { + if (level.isOff) return null; + const tokens = REASONING_EFFORT_TOKENS[level.value]; + return tokens === -1 ? 'Unlimited' : `Max ${tokens.toLocaleString()} tokens`; + }, + select(level: ReasoningEffortLevel): void { + if (level.isOff) { + conversationsStore.setThinkingEnabled(false); + return; + } + conversationsStore.setThinkingEnabled(true); + conversationsStore.setReasoningEffort(level.value as ReasoningEffort); + } + }; +} From 47c786924ad1ab7e91da2cdc72fcdb563780c2bd Mon Sep 17 00:00:00 2001 From: Charles Xu Date: Tue, 14 Jul 2026 12:12:18 +0200 Subject: [PATCH 08/24] kleidiai : add SME2 f32 kernel (#24414) * kleidiai : add SME2 f32 kernel * enable dynamic scheduling for SME2 f32 kernel --- ggml/src/ggml-cpu/CMakeLists.txt | 7 + ggml/src/ggml-cpu/kleidiai/kernels.cpp | 82 ++++++ ggml/src/ggml-cpu/kleidiai/kernels.h | 9 + ggml/src/ggml-cpu/kleidiai/kleidiai.cpp | 334 ++++++++++++++++++++---- 4 files changed, 375 insertions(+), 57 deletions(-) diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index f7c557af4..1316978e2 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -638,6 +638,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name) ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/) set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}") @@ -687,9 +688,15 @@ function(ggml_add_cpu_backend_variant_impl tag_name) ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.c ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa_asm.S + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa_asm.S ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme_asm.S + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme_asm.S ${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S) set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2+sme2+fp16") endif() diff --git a/ggml/src/ggml-cpu/kleidiai/kernels.cpp b/ggml/src/ggml-cpu/kleidiai/kernels.cpp index 8c4d7bc92..bf03fd766 100644 --- a/ggml/src/ggml-cpu/kleidiai/kernels.cpp +++ b/ggml/src/ggml-cpu/kleidiai/kernels.cpp @@ -20,14 +20,17 @@ #include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.h" #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.h" #include "kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.h" +#include "kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.h" #include "kai_lhs_pack_bf16p2vlx2_f32_sme.h" +#include "kai_lhs_pack_f32p2vlx1_f32_sme.h" #include "kai_lhs_quant_pack_qsi8d32p_f32.h" #include "kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.h" #include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h" #include "kai_lhs_quant_pack_qai8dxp_f32.h" #include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h" +#include "kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.h" #include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h" #include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h" #include "kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.h" @@ -865,6 +868,65 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = { { /* Sentinel */ } }; +static ggml_kleidiai_kernels ggml_kleidiai_kernels_f32[] = { +#if defined(__ARM_FEATURE_SME) + { + /* SME GEMM */ + { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_lhs_offset_ex = */ &kernel_offs_fn2, + /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2, + /* .run_kernel_ex = */ &kernel_run_fn10, + }, + /* .gemm_lhs_info = */ { + /* .get_offset = */ kai_get_lhs_offset_lhs_pack_f32p2vlx1_f32_sme, + /* .get_packed_offset_ex = */ &lhs_offs_fn5, + /* .packed_size_ex = */ &lhs_ps_fn5, + /* .pack_func_ex = */ &lhs_pack_void_fn9, + }, + /* SME GEMV */ + { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa, + /* .get_lhs_offset_ex = */ nullptr, + /* .get_rhs_packed_offset_ex = */ nullptr, + /* .run_kernel_ex = */ nullptr, + }, + /* .gemv_lhs_info = */ { + /* .get_offset = */ kai_get_lhs_offset_lhs_pack_f32p2vlx1_f32_sme, + /* .get_packed_offset_ex = */ &lhs_offs_fn5, + /* .packed_size_ex = */ &lhs_ps_fn5, + /* .pack_func_ex = */ &lhs_pack_void_fn9, + }, + /* .rhs_info = */ { + /* .packed_stride = */ nullptr, + /* .to_float = */ nullptr, + /* .packed_size_ex = */ &rhs_ps_fn2, + /* .packed_stride_ex = */ &rhs_stride_fn1, + /* .pack_func_ex = */ &rhs_pack_fn13, + }, + /* .required_cpu = */ CPU_FEATURE_SME, + /* .lhs_type = */ GGML_TYPE_F32, + /* .rhs_type = */ GGML_TYPE_F32, + /* .op_type = */ GGML_TYPE_F32, + }, +#endif + { /* Sentinel */ } +}; + ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) { ggml_kleidiai_kernels * kernel = nullptr; @@ -888,12 +950,15 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c if (tensor->src[0]->type == GGML_TYPE_Q8_0) { try_table(gemm_gemv_kernels_q8); + } else if (tensor->src[0]->type == GGML_TYPE_F32) { + try_table(ggml_kleidiai_kernels_f32); } else { try_table(gemm_gemv_kernels); } #else GGML_UNUSED(gemm_gemv_kernels); GGML_UNUSED(gemm_gemv_kernels_q8); + GGML_UNUSED(ggml_kleidiai_kernels_f32); GGML_UNUSED(cpu_features); #endif } @@ -937,3 +1002,20 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q8_0(cpu_feature features) return kernels; } + +ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_f32(cpu_feature features) { + ggml_kleidiai_kernels * kernels = nullptr; + +#if defined(__ARM_FEATURE_SME) + for (size_t i = 0; i < NELEMS(ggml_kleidiai_kernels_f32) - 1; ++i) { + if ((features & ggml_kleidiai_kernels_f32[i].required_cpu) == ggml_kleidiai_kernels_f32[i].required_cpu) { + kernels = &ggml_kleidiai_kernels_f32[i]; + break; + } + } +#else + GGML_UNUSED(features); +#endif + + return kernels; +} diff --git a/ggml/src/ggml-cpu/kleidiai/kernels.h b/ggml/src/ggml-cpu/kleidiai/kernels.h index 129245400..a46f837ac 100644 --- a/ggml/src/ggml-cpu/kleidiai/kernels.h +++ b/ggml/src/ggml-cpu/kleidiai/kernels.h @@ -55,6 +55,12 @@ struct lhs_packing_info { size_t m_idx_start, const void * lhs, size_t lhs_stride, void * lhs_packed); }; +enum rhs_repack_mode { + RHS_REPACK_PER_KERNEL, + RHS_REPACK_SHARED, + RHS_REPACK_SINGLE_ONLY, +}; + struct rhs_packing_info { size_t (*packed_stride)(size_t k, size_t nr, size_t kr, size_t bl); @@ -68,6 +74,8 @@ struct rhs_packing_info { void (*pack_func_ex)(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, size_t rhs_stride, const void * rhs, const void * bias, const void * scale, void * rhs_packed, size_t extra_bytes, const void * params); + + rhs_repack_mode repack_mode = RHS_REPACK_PER_KERNEL; }; struct ggml_kleidiai_kernels { @@ -88,3 +96,4 @@ struct ggml_kleidiai_kernels { ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor); ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features); ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q8_0(cpu_feature features); +ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_f32(cpu_feature features); diff --git a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp index 9e54b676b..a8de7df25 100644 --- a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +++ b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp @@ -60,10 +60,11 @@ struct ggml_kleidiai_context { cpu_feature features; ggml_kleidiai_kernels * kernels_q4; ggml_kleidiai_kernels * kernels_q8; + ggml_kleidiai_kernels * kernels_f32; int sme_thread_cap; // <= 0 means “SME disabled/unknown”; int thread_hint; // <= 0 means “no hint” int chunk_multiplier; -} static ctx = { CPU_FEATURE_NONE, nullptr, nullptr, 0, -1, 4 }; +} static ctx = { CPU_FEATURE_NONE, nullptr, nullptr, nullptr, 0, -1, 4 }; static const char* cpu_feature_to_string(cpu_feature f) { if (f == CPU_FEATURE_NONE) { @@ -156,10 +157,10 @@ static size_t detect_num_smcus() { } } } - return 1; + return 0; #else - return 1; + return 0; #endif } @@ -192,7 +193,6 @@ static void init_kleidiai_context(void) { const char *env_threads = getenv("GGML_TOTAL_THREADS"); const char *env_chunk_mult = getenv("GGML_KLEIDIAI_CHUNK_MULTIPLIER"); - const bool cpu_has_sme = ggml_cpu_has_sme(); size_t detected_smcus = 0; ctx.features = (ggml_cpu_has_dotprod() ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) | @@ -216,56 +216,47 @@ static void init_kleidiai_context(void) { } // SME policy: - // - If CPU doesn't support SME: SME always off. - // - Else: - // - env unset => auto-detect cores; enable if detected > 0. - // - env=0 => force off. - // - env>0 => force N cores (skip detection). + // - env unset => auto-detect SMCUs; enable SME only if detected > 0. + // - env=0 => force off. + // - env>0 => force N cores, if the binary was built with SME. int sme_cores = 0; bool sme_env_ok = false; bool sme_env_set = (env_sme != nullptr); - if (!cpu_has_sme) { - if (sme_env_set) { - bool ok = false; - int req = parse_uint_env(env_sme, "GGML_KLEIDIAI_SME", &ok); - if (ok && req > 0) { - GGML_LOG_WARN("kleidiai: GGML_KLEIDIAI_SME=%d but SME is not supported on this CPU; disabling SME\n", req); - } - } - sme_cores = 0; - } else { - if (sme_env_set) { - bool ok = false; - int v = parse_uint_env(env_sme, "GGML_KLEIDIAI_SME", &ok); - sme_env_ok = ok; + if (sme_env_set) { + bool ok = false; + int v = parse_uint_env(env_sme, "GGML_KLEIDIAI_SME", &ok); + sme_env_ok = ok; - if (!ok) { - GGML_LOG_WARN("kleidiai: GGML_KLEIDIAI_SME set but parsing failed; falling back to runtime SME-core detection\n"); - detected_smcus = detect_num_smcus(); - sme_cores = detected_smcus > 0 ? (int)detected_smcus : 0; - } else if (v == 0) { - sme_cores = 0; - } else { - sme_cores = v; - } - } else { + if (!ok) { + GGML_LOG_WARN("kleidiai: GGML_KLEIDIAI_SME set but parsing failed; falling back to runtime SME-core detection\n"); detected_smcus = detect_num_smcus(); sme_cores = detected_smcus > 0 ? (int)detected_smcus : 0; + } else if (v == 0) { + sme_cores = 0; + } else if (!ggml_cpu_has_sme()) { + GGML_LOG_WARN("kleidiai: GGML_KLEIDIAI_SME=%d but the binary was not built with SME; disabling SME\n", v); + sme_cores = 0; + } else { + sme_cores = v; } + } else { + detected_smcus = detect_num_smcus(); + sme_cores = detected_smcus > 0 ? (int)detected_smcus : 0; + } - if (!sme_env_set && sme_cores == 0) { - GGML_LOG_WARN("kleidiai: SME supported but runtime SME-core detection returned 0; falling back to NEON\n"); - } + if (!sme_env_set && ggml_cpu_has_sme() && sme_cores == 0) { + GGML_LOG_WARN("kleidiai: runtime SME-core detection returned 0; falling back to NEON\n"); + } - if (sme_cores > 0) { - ctx.features |= CPU_FEATURE_SME; - } + if (sme_cores > 0) { + ctx.features |= CPU_FEATURE_SME; } // Kernel selection - ctx.kernels_q4 = ggml_kleidiai_select_kernels_q4_0(ctx.features); - ctx.kernels_q8 = ggml_kleidiai_select_kernels_q8_0(ctx.features); + ctx.kernels_q4 = ggml_kleidiai_select_kernels_q4_0(ctx.features); + ctx.kernels_q8 = ggml_kleidiai_select_kernels_q8_0(ctx.features); + ctx.kernels_f32 = ggml_kleidiai_select_kernels_f32(ctx.features); if (!ctx.kernels_q4) { GGML_LOG_INFO("kleidiai: no compatible q4 kernels found for CPU features mask %d\n", (int)ctx.features); @@ -279,6 +270,12 @@ static void init_kleidiai_context(void) { GGML_LOG_INFO("kleidiai: primary q8 kernel feature %s\n", cpu_feature_to_string(ctx.kernels_q8->required_cpu)); } + if (!ctx.kernels_f32) { + GGML_LOG_INFO("kleidiai: no compatible f32 kernels found for CPU features mask %d\n", (int)ctx.features); + } else { + GGML_LOG_INFO("kleidiai: primary f32 kernel feature %s\n", cpu_feature_to_string(ctx.kernels_f32->required_cpu)); + } + ctx.sme_thread_cap = (ctx.features & CPU_FEATURE_SME) ? sme_cores : 0; if (ctx.features & CPU_FEATURE_SME) { @@ -334,6 +331,13 @@ static inline size_t ceil_div_size(size_t a, size_t b) { return b == 0 ? 0 : (a + b - 1) / b; } +static inline size_t kleidiai_chunk_cols(size_t n, int nth_total, bool disable_chunking, size_t n_step) { + const size_t multiplier = (nth_total == 1 || disable_chunking) ? 1 : std::max(1, (size_t) ctx.chunk_multiplier); + const size_t divisor = std::max(1, (size_t) nth_total * multiplier); + const size_t chunk_cols = align_up(std::max(1, ceil_div_size(n, divisor)), n_step); + return chunk_cols ? chunk_cols : n_step; +} + struct kleidiai_block_args { size_t lhs_bl; size_t rhs_bl; @@ -418,6 +422,10 @@ static inline ggml_kleidiai_kernels * kleidiai_primary_kernel_q8() { return ctx.kernels_q8; } +static inline ggml_kleidiai_kernels * kleidiai_primary_kernel_f32() { + return ctx.kernels_f32; +} + template static int kleidiai_collect_kernel_chain_common( ggml_kleidiai_kernels * primary, @@ -430,11 +438,16 @@ static int kleidiai_collect_kernel_chain_common( } out[count++] = primary; + if (primary->rhs_info.repack_mode == RHS_REPACK_SINGLE_ONLY) { + return count; + } + if ((primary->required_cpu & CPU_FEATURE_SME) == CPU_FEATURE_SME) { const cpu_feature fallback_mask = static_cast(features & ~CPU_FEATURE_SME); if (fallback_mask != CPU_FEATURE_NONE) { ggml_kleidiai_kernels * fallback = select_fallback(fallback_mask); if (fallback && fallback != primary && + fallback->rhs_info.repack_mode != RHS_REPACK_SINGLE_ONLY && fallback->lhs_type == primary->lhs_type && fallback->rhs_type == primary->rhs_type && fallback->op_type == primary->op_type) { @@ -465,6 +478,12 @@ static int kleidiai_collect_q8_chain(std::array & out) { + ggml_kleidiai_kernels * primary = kleidiai_primary_kernel_f32(); + return kleidiai_collect_kernel_chain_common(primary, ctx.features, out, + [&](cpu_feature mask) { return ggml_kleidiai_select_kernels_f32(mask); }); +} + static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) { GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS); return tensor->ne[dim]; @@ -539,6 +558,36 @@ class tensor_traits : public ggml::cpu::tensor_traits { return true; } + if (op->src[0]->type == GGML_TYPE_F32) { + size_t cursor = 0; + bool any_slot = false; + + for (int slot = 0; slot < slot_count; ++slot) { + ggml_kleidiai_kernels * kernels = kernel_chain[slot]; + lhs_packing_info * lhs_info = &kernels->gemm_lhs_info; + kernel_info * kernel = &kernels->gemm; + + if (!lhs_info || !lhs_info->packed_size_ex || !kernel) { + return false; + } + + const size_t mr = kernel->get_mr(); + const size_t kr = kernel->get_kr(); + const size_t sr = kernel->get_sr(); + + cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN); + cursor += lhs_info->packed_size_ex(m, k, 0, mr, kr, sr); + any_slot = true; + } + + if (!any_slot) { + return false; + } + + size = cursor; + return true; + } + if (op->src[0]->type == GGML_TYPE_F16) { const int64_t lhs_batch_size0 = op->src[1]->ne[2]; const int64_t rhs_batch_size0 = op->src[0]->ne[2]; @@ -595,6 +644,8 @@ class tensor_traits : public ggml::cpu::tensor_traits { if (dst->op == GGML_OP_MUL_MAT) { if (dst->src[0]->type == GGML_TYPE_Q4_0 || dst->src[0]->type == GGML_TYPE_Q8_0) { return compute_forward_qx(params, dst); + } else if (dst->src[0]->type == GGML_TYPE_F32) { + return compute_forward_f32(params, dst); } else if (dst->src[0]->type == GGML_TYPE_F16) { return compute_forward_fp16(params, dst); } @@ -606,6 +657,144 @@ class tensor_traits : public ggml::cpu::tensor_traits { return false; } + bool compute_forward_f32(ggml_compute_params * params, struct ggml_tensor * dst) { + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + if (src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) { + return false; + } + + ggml_kleidiai_kernels * kernels = kleidiai_primary_kernel_f32(); + if (!kernels) { + return false; + } + + kernel_info * kernel = &kernels->gemm; + lhs_packing_info * lhs_info = &kernels->gemm_lhs_info; + + if (!kernel || !lhs_info || !lhs_info->get_offset || !lhs_info->get_packed_offset_ex || + !lhs_info->packed_size_ex || !lhs_info->pack_func_ex || + !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex || !kernel->get_dst_offset) { + return false; + } + + const kleidiai_weight_header * header = kleidiai_weight_header_from_ptr(src0->data); + const bool has_header = kleidiai_is_weight_header_valid(header); + + const uint8_t * rhs_base = has_header ? kleidiai_weight_slot_ptr(header, 0) + : static_cast(src0->data); + if (!rhs_base) { + return false; + } + + const int nth = params->nth > 0 ? params->nth : 1; + const int ith = params->ith; + + const size_t k = ne00; + const size_t m = ne11; + const size_t n = ne01; + + const size_t mr = kernel->get_mr(); + const size_t kr = kernel->get_kr(); + const size_t sr = kernel->get_sr(); + + const size_t lhs_packed_size = lhs_info->packed_size_ex(m, k, 0, mr, kr, sr); + GGML_ASSERT(lhs_packed_size <= params->wsize); + + uint8_t * lhs_packed = static_cast(params->wdata); + const size_t dst_stride = dst->nb[1]; + const size_t n_step = kernel->get_n_step() ? kernel->get_n_step() : 1; + const bool disable_chunking = ggml_is_numa(); + GGML_ASSERT(n <= (size_t) INT_MAX); + + for (int64_t batch_idx = 0; batch_idx < ne12; ++batch_idx) { + const uint8_t * lhs_batch_base = static_cast(src1->data) + batch_idx * src1->nb[2]; + uint8_t * dst_batch_base = static_cast(dst->data) + batch_idx * dst->nb[2]; + + { + const int64_t m_roundup_mr = kai_roundup((int64_t)m, (int64_t)mr); + int64_t max_threads = mr ? (m_roundup_mr / (int64_t)mr) : nth; + max_threads = std::max(1, max_threads); + const int64_t use_threads = std::min(nth, max_threads); + + if (ith < use_threads) { + const int64_t num_m_per_thread0 = round_down((size_t)(m_roundup_mr / use_threads), mr); + const int64_t num_m_per_threadN_1 = (int64_t)m - (use_threads - 1) * num_m_per_thread0; + + const int64_t m_start = (int64_t)ith * num_m_per_thread0; + const int64_t m_count = (ith == use_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0; + + const size_t base_packed_off = lhs_info->get_packed_offset_ex(m_start, k, 0, mr, kr, sr); + const size_t next_block_off = lhs_info->get_packed_offset_ex(m_start + mr, k, 0, mr, kr, sr); + const size_t row_stride_bytes = mr ? (next_block_off - base_packed_off) / mr : 0; + + int64_t remaining = m_count; + int64_t cur = m_start; + + while (remaining > 0) { + const int64_t take = std::min((int64_t)m - cur, remaining); + const size_t src_off = lhs_info->get_offset(cur, src1->nb[1]); + const void * src_ptr = lhs_batch_base + src_off; + const size_t dst_off = base_packed_off + (size_t)(cur - m_start) * row_stride_bytes; + void * dst_ptr = lhs_packed + dst_off; + + lhs_info->pack_func_ex(take, k, 0, mr, kr, sr, 0, src_ptr, src1->nb[1], dst_ptr); + + cur += take; + remaining -= take; + } + } + } + + if (ith == 0) { + ggml_threadpool_chunk_set(params->threadpool, 0); + } + + ggml_barrier(params->threadpool); + + const size_t chunk_cols = kleidiai_chunk_cols(n, nth, disable_chunking, n_step); + GGML_ASSERT(chunk_cols <= (size_t) INT_MAX); + + int current_col = ggml_threadpool_chunk_add(params->threadpool, (int) chunk_cols); + while ((size_t) current_col < n) { + const size_t n_start = (size_t) current_col; + const size_t n_to_process = std::min(chunk_cols, n - n_start); + + if (n_to_process > 0) { + const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr); + const size_t rhs_packed_offset = kernel->get_rhs_packed_offset_ex(n_start, k, 0); + const size_t dst_offset = kernel->get_dst_offset(0, n_start, dst_stride); + + const void * lhs_ptr = lhs_packed + lhs_packed_offset; + const void * rhs_ptr = rhs_base + rhs_packed_offset; + float * dst_ptr = reinterpret_cast(dst_batch_base + dst_offset); + + kernel->run_kernel_ex(m, n_to_process, k, 0, + lhs_ptr, + rhs_ptr, + dst_ptr, + dst_stride, + sizeof(float), + -FLT_MAX, + FLT_MAX); + } + + current_col = ggml_threadpool_chunk_add(params->threadpool, (int) chunk_cols); + } + + if (batch_idx != ne12 - 1) { + ggml_barrier(params->threadpool); + } + } + + return true; + } + bool compute_forward_fp16(ggml_compute_params * params, struct ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; const ggml_tensor * src1 = dst->src[1]; @@ -1214,7 +1403,7 @@ class tensor_traits : public ggml::cpu::tensor_traits { public: int repack(struct ggml_tensor * tensor, const void * data, size_t data_size) { - GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q8_0); + GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q8_0 || tensor->type == GGML_TYPE_F32); const size_t n = tensor->ne[1]; const size_t k = tensor->ne[0]; @@ -1233,12 +1422,15 @@ public: std::array kernel_chain; const bool want_q8 = tensor->type == GGML_TYPE_Q8_0; - const int slot_total = want_q8 ? kleidiai_collect_q8_chain(kernel_chain) - : kleidiai_collect_q4_chain(kernel_chain); + const bool want_f32 = tensor->type == GGML_TYPE_F32; + const int slot_total = want_f32 ? kleidiai_collect_f32_chain(kernel_chain) + : want_q8 ? kleidiai_collect_q8_chain(kernel_chain) + : kleidiai_collect_q4_chain(kernel_chain); const bool allow_fallback = kleidiai_pack_fallback_allowed(); std::vector qdata; std::vector scales; + std::vector bias; if (want_q8 && slot_total > 0) { qdata.resize(n * k, 0); @@ -1286,6 +1478,10 @@ public: } } + if (want_f32 && slot_total > 0) { + bias.resize(n, 0.0f); + } + for (int slot = 0; slot < slot_total && slot < GGML_KLEIDIAI_MAX_KERNEL_SLOTS; ++slot) { if (!allow_fallback && slot > 0) { break; @@ -1302,8 +1498,9 @@ public: const size_t sr = kernel->get_sr(); const ggml_type rhs_type = kernels->rhs_type; const size_t block_len = rhs_type == GGML_TYPE_Q8_0 ? QK8_0 : - rhs_type == GGML_TYPE_Q4_0 ? QK4_0 : 0; - if (block_len == 0) { + rhs_type == GGML_TYPE_Q4_0 ? QK4_0 : + rhs_type == GGML_TYPE_F32 ? 0 : SIZE_MAX; + if (block_len == SIZE_MAX) { continue; } @@ -1326,6 +1523,10 @@ public: rhs_info->pack_func_ex(1, n, k, nr, kr, sr, 0, 0, qdata.data(), nullptr, scales.data(), dst_ptr, 0, ¶ms); + } else if (rhs_type == GGML_TYPE_F32) { + rhs_info->pack_func_ex(1, n, k, nr, kr, sr, 0, tensor->nb[1], + data, bias.data(), nullptr, + dst_ptr, 0, nullptr); } else { continue; } @@ -1400,7 +1601,7 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_b static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) { GGML_UNUSED(buft); - if (tensor->type != GGML_TYPE_Q4_0 && tensor->type != GGML_TYPE_Q8_0) { + if (tensor->type != GGML_TYPE_Q4_0 && tensor->type != GGML_TYPE_Q8_0 && tensor->type != GGML_TYPE_F32) { return ggml_nbytes(tensor); } @@ -1412,8 +1613,10 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_ std::array kernel_chain; const bool want_q8 = tensor->type == GGML_TYPE_Q8_0; - const int slot_total = want_q8 ? kleidiai_collect_q8_chain(kernel_chain) - : kleidiai_collect_q4_chain(kernel_chain); + const bool want_f32 = tensor->type == GGML_TYPE_F32; + const int slot_total = want_f32 ? kleidiai_collect_f32_chain(kernel_chain) + : want_q8 ? kleidiai_collect_q8_chain(kernel_chain) + : kleidiai_collect_q4_chain(kernel_chain); const bool allow_fallback = kleidiai_pack_fallback_allowed(); size_t slot_count = 0; @@ -1433,8 +1636,9 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_ const ggml_type rhs_type = kernels->rhs_type; const size_t block_len = rhs_type == GGML_TYPE_Q4_0 ? QK4_0 : - rhs_type == GGML_TYPE_Q8_0 ? QK8_0 : 0; - if (block_len == 0) { + rhs_type == GGML_TYPE_Q8_0 ? QK8_0 : + rhs_type == GGML_TYPE_F32 ? 0 : SIZE_MAX; + if (block_len == SIZE_MAX) { continue; } @@ -1455,25 +1659,41 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type { bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override { std::array kernel_chain; const int slot_total = kleidiai_collect_kernel_chain(op, kernel_chain); - if ((op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) && - (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_Q8_0) && + const bool src0_is_kleidiai = op->src[0]->buffer && (ggml_n_dims(op->src[0]) == 2) && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && - slot_total > 0) { + slot_total > 0; + + if ((op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) && + (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_Q8_0 || op->src[0]->type == GGML_TYPE_F32) && + src0_is_kleidiai) { if (op->src[0]->type == GGML_TYPE_Q4_0 && ctx.kernels_q4 == nullptr) { return false; } if (op->src[0]->type == GGML_TYPE_Q8_0 && ctx.kernels_q8 == nullptr) { return false; } + if (op->src[0]->type == GGML_TYPE_F32 && ctx.kernels_f32 == nullptr) { + return false; + } if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) { return false; } - if ((op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_I32) && - ggml_ne(op->src[1], 3) == 1) { - return true; + + if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_Q8_0) { + if ((op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_I32) && + ggml_ne(op->src[1], 3) == 1) { + return true; + } + return false; } + + if (op->op != GGML_OP_MUL_MAT || op->src[1]->type != GGML_TYPE_F32 || op->type != GGML_TYPE_F32) { + return false; + } + + return true; } return false; } From 47a39665e7081dc482feec169961acc09750a5c4 Mon Sep 17 00:00:00 2001 From: JusteLeo Date: Tue, 14 Jul 2026 12:13:13 +0200 Subject: [PATCH 09/24] ggml: uniformize im2col dst_type for all conv ops (#23660) * ggml: uniformize im2col dst_type for all conv ops * Update ggml/src/ggml.c Co-authored-by: Georgi Gerganov * ggml : uniformize im2col casting logic across all conv ops * fix : allow im2col_f16 to accept any kernel type --------- Co-authored-by: Georgi Gerganov --- ggml/src/ggml-cpu/ops.cpp | 3 --- ggml/src/ggml.c | 10 +++++----- tests/test-backend-ops.cpp | 1 + 3 files changed, 6 insertions(+), 8 deletions(-) diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp index 85eaebb57..7c54cb6f4 100644 --- a/ggml/src/ggml-cpu/ops.cpp +++ b/ggml/src/ggml-cpu/ops.cpp @@ -6362,7 +6362,6 @@ static void ggml_compute_forward_im2col_f16( const ggml_tensor * src0 = dst->src[0]; const ggml_tensor * src1 = dst->src[1]; - GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F16); @@ -6393,7 +6392,6 @@ static void ggml_compute_forward_im2col_f16( int ofs0 = is_2D ? nb13 : nb12; int ofs1 = is_2D ? nb12 : nb11; - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); GGML_ASSERT(nb10 == ggml_type_size(src1->type)); // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] @@ -6563,7 +6561,6 @@ static void ggml_compute_forward_im2col_3d_f16( const ggml_tensor * src0 = dst->src[0]; const ggml_tensor * src1 = dst->src[1]; - GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F16); diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 5d8804991..f9cc7465e 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -4507,7 +4507,7 @@ struct ggml_tensor * ggml_conv_1d( int s0, int p0, int d0) { - struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16); // [N, OL, IC * K] + struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, 0, p0, 0, d0, 0, false, a->type == GGML_TYPE_BF16 ? GGML_TYPE_F32 : GGML_TYPE_F16); // [N, OL, IC * K] struct ggml_tensor * result = ggml_mul_mat(ctx, @@ -4541,7 +4541,7 @@ struct ggml_tensor * ggml_conv_1d_dw( int d0) { struct ggml_tensor * new_b = ggml_reshape_4d(ctx, b, b->ne[0], 1, b->ne[1], b->ne[2]); - struct ggml_tensor * im2col = ggml_im2col(ctx, a, new_b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16); + struct ggml_tensor * im2col = ggml_im2col(ctx, a, new_b, s0, 0, p0, 0, d0, 0, false, a->type == GGML_TYPE_BF16 ? GGML_TYPE_F32 : GGML_TYPE_F16); struct ggml_tensor * result = ggml_mul_mat(ctx, im2col, a); @@ -4647,7 +4647,7 @@ struct ggml_tensor * ggml_conv_2d( int p1, int d0, int d1) { - struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true, a->type); // [N, OH, OW, IC * KH * KW] + struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true, a->type == GGML_TYPE_BF16 ? GGML_TYPE_F32 : GGML_TYPE_F16); // [N, OH, OW, IC * KH * KW] struct ggml_tensor * result = ggml_mul_mat(ctx, @@ -4729,7 +4729,7 @@ struct ggml_tensor * ggml_conv_3d( int d1, // dilation height int d2 // dilation depth ) { - struct ggml_tensor * im2col = ggml_im2col_3d(ctx, a, b, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2, a->type); // [N*OD, OH, OW, IC * KD * KH * KW] + struct ggml_tensor * im2col = ggml_im2col_3d(ctx, a, b, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2, a->type == GGML_TYPE_BF16 ? GGML_TYPE_F32 : GGML_TYPE_F16); // [N*OD, OH, OW, IC * KD * KH * KW] int64_t OC = a->ne[3] / IC; int64_t N = b->ne[3] / IC; @@ -4779,7 +4779,7 @@ struct ggml_tensor * ggml_conv_2d_dw( struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]); struct ggml_tensor * im2col = ggml_im2col(ctx, new_a, ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]), - s0, s1, p0, p1, d0, d1, true, GGML_TYPE_F16); // [N * IC, OH, OW, KH * KW] + s0, s1, p0, p1, d0, d1, true, a->type == GGML_TYPE_BF16 ? GGML_TYPE_F32 : GGML_TYPE_F16); // [N * IC, OH, OW, KH * KW] struct ggml_tensor * new_b = ggml_reshape_4d(ctx, im2col, im2col->ne[0], im2col->ne[2] * im2col->ne[1], b->ne[2], b->ne[3]); // [N * IC, OH, OW, KH * KW] => [N, IC, OH * OW, KH * KW] new_a = ggml_reshape_4d(ctx, new_a, (new_a->ne[0] * new_a->ne[1]), new_a->ne[2], new_a->ne[3], 1); // [OC，1, KH, KW] => [1, OC, 1, KH * KW] diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 86d17b3f6..a83cb302b 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -8013,6 +8013,7 @@ static std::vector> make_test_cases_eval() { // im2col 2D test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32)); + test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F16)); test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32)); test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16)); for (int s0 : {1, 3}) { From 657e01125aa49577a62a5531fde24cbcc007006d Mon Sep 17 00:00:00 2001 From: Christian Kastner Date: Tue, 14 Jul 2026 12:15:41 +0200 Subject: [PATCH 10/24] tests: export-graph-ops: exit gracefully when called w/o arguments (#25619) Fixes a segfault when `test-export-graph-ops` is called without any arguments. --- tests/test-export-graph-ops.cpp | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/tests/test-export-graph-ops.cpp b/tests/test-export-graph-ops.cpp index 7d8118dcd..46ded1398 100644 --- a/tests/test-export-graph-ops.cpp +++ b/tests/test-export-graph-ops.cpp @@ -152,6 +152,10 @@ int main(int argc, char ** argv) { init_result = common_init_from_params(params); ctx = init_result->context(); + if (!ctx) { + LOG_ERR("failed to initialize params\n"); + return 1; + } } else { #ifdef LLAMA_HF_FETCH auto [hf_repo, hf_quant] = common_download_split_repo_tag(params.model.hf_repo); From a7312ae94f801fc9c6786dc56e38df57b964f697 Mon Sep 17 00:00:00 2001 From: fairydreaming <166155368+fairydreaming@users.noreply.github.com> Date: Tue, 14 Jul 2026 14:37:52 +0200 Subject: [PATCH 11/24] ggml : add a set of functions for checking contiguity of inner tensor dimensions (#25650) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-authored-by: Stanisław Szymczyk --- ggml/include/ggml.h | 4 ++++ ggml/src/ggml.c | 24 ++++++++++++++++++------ 2 files changed, 22 insertions(+), 6 deletions(-) diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index b2859ebe7..92ba65222 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -780,6 +780,10 @@ extern "C" { GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1 GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2 + GGML_API bool ggml_is_contiguous_to_1(const struct ggml_tensor * tensor); // contiguous for dims < 1 + GGML_API bool ggml_is_contiguous_to_2(const struct ggml_tensor * tensor); // contiguous for dims < 2 + GGML_API bool ggml_is_contiguous_to_3(const struct ggml_tensor * tensor); // contiguous for dims < 3 + // returns whether the tensor elements are allocated as one contiguous block of memory (no gaps, but permutation ok) GGML_API bool ggml_is_contiguously_allocated(const struct ggml_tensor * tensor); diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index f9cc7465e..086b6ab08 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1464,14 +1464,14 @@ bool ggml_is_transposed(const struct ggml_tensor * tensor) { return tensor->nb[0] > tensor->nb[1]; } -static bool ggml_is_contiguous_n(const struct ggml_tensor * tensor, int n) { +static bool ggml_is_contiguous_m_n(const struct ggml_tensor * tensor, int m, int n) { size_t next_nb = ggml_type_size(tensor->type); if (tensor->ne[0] != ggml_blck_size(tensor->type) && tensor->nb[0] != next_nb) { return false; } next_nb *= tensor->ne[0]/ggml_blck_size(tensor->type); - for (int i = 1; i < GGML_MAX_DIMS; i++) { - if (i > n) { + for (int i = 1; i < n; i++) { + if (i > m) { if (tensor->ne[i] != 1 && tensor->nb[i] != next_nb) { return false; } @@ -1489,15 +1489,27 @@ bool ggml_is_contiguous(const struct ggml_tensor * tensor) { } bool ggml_is_contiguous_0(const struct ggml_tensor * tensor) { - return ggml_is_contiguous_n(tensor, 0); + return ggml_is_contiguous_m_n(tensor, 0, GGML_MAX_DIMS); } bool ggml_is_contiguous_1(const struct ggml_tensor * tensor) { - return ggml_is_contiguous_n(tensor, 1); + return ggml_is_contiguous_m_n(tensor, 1, GGML_MAX_DIMS); } bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) { - return ggml_is_contiguous_n(tensor, 2); + return ggml_is_contiguous_m_n(tensor, 2, GGML_MAX_DIMS); +} + +bool ggml_is_contiguous_to_1(const struct ggml_tensor * tensor) { + return ggml_is_contiguous_m_n(tensor, 0, 1); +} + +bool ggml_is_contiguous_to_2(const struct ggml_tensor * tensor) { + return ggml_is_contiguous_m_n(tensor, 0, 2); +} + +bool ggml_is_contiguous_to_3(const struct ggml_tensor * tensor) { + return ggml_is_contiguous_m_n(tensor, 0, 3); } bool ggml_is_contiguously_allocated(const struct ggml_tensor * tensor) { From 8ff8c4299db80225ff075683b51b63274ba7ea9a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Adrien=20Gallou=C3=ABt?= Date: Tue, 14 Jul 2026 15:20:53 +0200 Subject: [PATCH 12/24] tokenize : align usage by using common args (#25516) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Migrate the tokenize tool to common_params_parse, replacing its hand-rolled argv parsing, Windows UTF-8 handling and file reading with the shared common helpers. Expose the model-sourcing flags (-m, -mu, -dr, -hf, -hff, --offline, HF_TOKEN) to LLAMA_EXAMPLE_TOKENIZE, and register --ids, --stdin, --no-bos, --no-parse-special and --show-count as common args. parse_special defaults to true for TOKENIZE to preserve the old behavior. Errors now go through LOG_ERR instead of fprintf(stderr). Signed-off-by: Adrien Gallouët --- common/arg.cpp | 51 +++++- common/common.h | 7 + tools/tokenize/tokenize.cpp | 327 ++++++++---------------------------- 3 files changed, 120 insertions(+), 265 deletions(-) diff --git a/common/arg.cpp b/common/arg.cpp index 9aefcd202..9676adafe 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -1179,6 +1179,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.sampling.temp = 0.2; // lower temp by default for better quality } else if (ex == LLAMA_EXAMPLE_SERVER) { params.n_parallel = -1; // auto by default + } else if (ex == LLAMA_EXAMPLE_TOKENIZE) { + params.parse_special = true; // parse special tokens by default, like the old tokenize tool } params.use_color = tty_can_use_colors(); @@ -2746,14 +2748,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params, const std::string & value) { params.model.path = value; } - ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_MODEL")); + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_TOKENIZE}).set_env("LLAMA_ARG_MODEL")); add_opt(common_arg( {"-mu", "--model-url"}, "MODEL_URL", "model download url (default: unused)", [](common_params & params, const std::string & value) { params.model.url = value; } - ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_MODEL_URL")); + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_TOKENIZE}).set_env("LLAMA_ARG_MODEL_URL")); add_opt(common_arg( { "-dr", "--docker-repo" }, "[/][:quant]", "Docker Hub model repository. repo is optional, default to ai/. quant is optional, default to :latest.\n" @@ -2762,7 +2764,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params, const std::string & value) { params.model.docker_repo = value; } - ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_DOCKER_REPO")); + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_TOKENIZE}).set_env("LLAMA_ARG_DOCKER_REPO")); add_opt(common_arg( {"-hf", "-hfr", "--hf-repo"}, "/[:quant]", "Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n" @@ -2772,14 +2774,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params, const std::string & value) { params.model.hf_repo = value; } - ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_HF_REPO")); + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_TOKENIZE}).set_env("LLAMA_ARG_HF_REPO")); add_opt(common_arg( {"-hff", "--hf-file"}, "FILE", "Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)", [](common_params & params, const std::string & value) { params.model.hf_file = value; } - ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_HF_FILE")); + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_TOKENIZE}).set_env("LLAMA_ARG_HF_FILE")); add_opt(common_arg( {"-hfv", "-hfrv", "--hf-repo-v"}, "/[:quant]", "Hugging Face model repository for the vocoder model (default: unused)", @@ -2800,7 +2802,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params, const std::string & value) { params.hf_token = value; } - ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("HF_TOKEN")); + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_TOKENIZE}).set_env("HF_TOKEN")); add_opt(common_arg( {"--mtp"}, "also download the multi-token prediction (MTP) head, if available (default: unused)", @@ -2916,6 +2918,41 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.parse_special = true; } ).set_examples({LLAMA_EXAMPLE_IMATRIX})); + add_opt(common_arg( + {"--ids"}, + string_format("only print the token IDs, in a Python-parseable list form like [1, 2, 3] (default: %s)", params.tokenize_ids ? "true" : "false"), + [](common_params & params) { + params.tokenize_ids = true; + } + ).set_examples({LLAMA_EXAMPLE_TOKENIZE})); + add_opt(common_arg( + {"--stdin"}, + string_format("read the prompt from stdin (mutually exclusive with -f/--file and -p/--prompt) (default: %s)", params.tokenize_stdin ? "true" : "false"), + [](common_params & params) { + params.tokenize_stdin = true; + } + ).set_examples({LLAMA_EXAMPLE_TOKENIZE})); + add_opt(common_arg( + {"--no-bos"}, + string_format("do not add a BOS token to the prompt, even if the model normally uses one (default: %s)", params.tokenize_no_bos ? "true" : "false"), + [](common_params & params) { + params.tokenize_no_bos = true; + } + ).set_examples({LLAMA_EXAMPLE_TOKENIZE})); + add_opt(common_arg( + {"--no-parse-special"}, + string_format("do not parse special tokens (chat, tool, etc) (default: %s)", !params.parse_special ? "true" : "false"), + [](common_params & params) { + params.parse_special = false; + } + ).set_examples({LLAMA_EXAMPLE_TOKENIZE})); + add_opt(common_arg( + {"--show-count"}, + string_format("print the total number of tokens (default: %s)", params.tokenize_show_count ? "true" : "false"), + [](common_params & params) { + params.tokenize_show_count = true; + } + ).set_examples({LLAMA_EXAMPLE_TOKENIZE})); add_opt(common_arg( {"-pps"}, string_format("is the prompt shared across parallel sequences (default: %s)", params.is_pp_shared ? "true" : "false"), @@ -3506,7 +3543,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params) { params.offline = true; } - ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_OFFLINE")); + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_TOKENIZE}).set_env("LLAMA_ARG_OFFLINE")); add_opt(common_arg( {"-lv", "--verbosity", "--log-verbosity"}, "N", string_format("Set the verbosity threshold. Messages with a higher verbosity will be ignored. Values:\n" diff --git a/common/common.h b/common/common.h index 7ed1a9827..66760005c 100644 --- a/common/common.h +++ b/common/common.h @@ -105,6 +105,7 @@ enum llama_example { LLAMA_EXAMPLE_RESULTS, LLAMA_EXAMPLE_EXPORT_GRAPH_OPS, LLAMA_EXAMPLE_DOWNLOAD, + LLAMA_EXAMPLE_TOKENIZE, LLAMA_EXAMPLE_COUNT, }; @@ -716,6 +717,12 @@ struct common_params { // batched-bench params bool batched_bench_output_jsonl = false; + // tokenize params + bool tokenize_ids = false; // if true, only print the token IDs + bool tokenize_stdin = false; // if true, read the prompt from stdin + bool tokenize_no_bos = false; // if true, do not add the BOS token + bool tokenize_show_count = false; // if true, print the total token count + // common params std::string out_file; // output filename for all example programs // optional callback for model loading progress and cancellation: diff --git a/tools/tokenize/tokenize.cpp b/tools/tokenize/tokenize.cpp index 32cf8c8eb..23120ad2e 100644 --- a/tools/tokenize/tokenize.cpp +++ b/tools/tokenize/tokenize.cpp @@ -1,5 +1,6 @@ +#include "arg.h" #include "common.h" -//#include "log.h" // TODO: start using log.h +#include "log.h" #include "llama.h" #include @@ -8,115 +9,22 @@ #include #include #include -#include // TODO: remove me +#include +#include #if defined(_WIN32) #define WIN32_LEAN_AND_MEAN #include -#include // For CommandLineToArgvW #endif -static void print_usage_information(const char * argv0) { - printf("usage: %s [options]\n\n", argv0); - printf("The tokenize program tokenizes a prompt using a given model,\n"); - printf("and prints the resulting tokens to standard output.\n\n"); - printf("It needs a model file, a prompt, and optionally other flags\n"); - printf("to control the behavior of the tokenizer.\n\n"); - printf(" The possible options are:\n"); - printf("\n"); - printf(" -h, --help print this help and exit\n"); - printf(" -m MODEL_PATH, --model MODEL_PATH path to model.\n"); - printf(" --ids if given, only print numerical token IDs, and not token strings.\n"); - printf(" The output format looks like [1, 2, 3], i.e. parseable by Python.\n"); - printf(" -f PROMPT_FNAME, --file PROMPT_FNAME read prompt from a file.\n"); - printf(" -p PROMPT, --prompt PROMPT read prompt from the argument.\n"); - printf(" --stdin read prompt from standard input.\n"); - printf(" --no-bos do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n"); - printf(" --no-escape do not escape input (such as \\n, \\t, etc.).\n"); - printf(" --no-parse-special do not parse control tokens.\n"); - printf(" --log-disable disable logs. Makes stderr quiet when loading the model.\n"); - printf(" --show-count print the total number of tokens.\n"); -} +static void print_usage(int argc, char ** argv) { + (void) argc; -static void llama_log_callback_null(ggml_log_level level, const char * text, void * user_data) { - (void) level; - (void) text; - (void) user_data; -} - -static std::string read_prompt_from_file(const char * filepath, bool & success) { - success = false; - - std::ifstream in(filepath, std::ios::binary); - if (!in) { - fprintf(stderr, "%s: could not open file '%s' for reading: %s\n", __func__, filepath, strerror(errno)); - return std::string(); - } - // do not assume the file is seekable (e.g. /dev/stdin) - std::stringstream buffer; - buffer << in.rdbuf(); - if (in.fail()) { - fprintf(stderr, "%s: could not read the entire file '%s': %s\n", __func__, filepath, strerror(errno)); - return std::string(); - } - - success = true; - return buffer.str(); -} - -// -// Function: ingest_args(...) -> vector -// -// Takes argc and argv arguments, and converts them to a vector of UTF-8 encoded -// strings, as an STL vector. -// -// In particular, it handles character encoding shenanigans on Windows. -// -// Note: raw_argc and raw_argv are not actually read at all on Windows. -// On Windows we call GetCommandLineW to get the arguments in wchar_t -// format, ignoring the regular argc/argv arguments to main(). -// -// TODO: potential opportunity to roll common stuff into common/console.cpp -// in relation to Windows wchar_t shenanigans. -static std::vector ingest_args(int raw_argc, char ** raw_argv) { - std::vector argv; - - // Handle Windows, if given non-ASCII arguments. - // We convert wchar_t arguments into UTF-8 char* on this platform. - // Lets you invoke 'tokenize' on Windows cmd.exe with non-ASCII characters - // without throwing tantrums. -#if defined(_WIN32) - int argc; - const LPWSTR cmdline_wargv = GetCommandLineW(); - LPWSTR * wargv = CommandLineToArgvW(cmdline_wargv, &argc); - - // silence unused arg warnings - (void) raw_argc; - (void) raw_argv; - - for (int i = 0; i < argc; ++i) { - int length_needed = WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), 0, 0, NULL, NULL); - char * output_buf = (char *) calloc(length_needed+1, sizeof(char)); - GGML_ASSERT(output_buf); - - WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), output_buf, length_needed, NULL, NULL); - output_buf[length_needed] = '\0'; - - argv.push_back(output_buf); - free(output_buf); - } - - LocalFree((HLOCAL) wargv); -#else - int argc = raw_argc; - for (int i = 0; i < argc; ++i) { - argv.push_back(raw_argv[i]); - } -#endif - - GGML_ASSERT((unsigned int) argc == argv.size()); - - return argv; + LOG("\nexample usage:\n"); + LOG("\n %s -m your_model.gguf -p \"Hello world\"\n", argv[0]); + LOG("\n %s -m your_model.gguf -f prompt.txt --ids\n", argv[0]); + LOG("\n cat prompt.txt | %s -m your_model.gguf --stdin --show-count\n", argv[0]); + LOG("\n"); } // @@ -184,166 +92,69 @@ static void write_utf8_cstr_to_stdout(const char * str, bool & invalid_utf8) { #endif } -int main(int raw_argc, char ** raw_argv) { +int main(int argc, char ** argv) { std::setlocale(LC_NUMERIC, "C"); - const std::vector argv = ingest_args(raw_argc, raw_argv); - const int argc = argv.size(); + common_params params; - if (argc <= 1) { - print_usage_information(argv[0].c_str()); + common_init(); + + if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_TOKENIZE, print_usage)) { return 1; } - ////// - // Read out all the command line arguments. - ////// + // which prompt source was requested? + // -p/--prompt and -f/--file both end up in params.prompt (common's -f also + // strips a single trailing newline), but -f additionally records the path + // in params.prompt_file, so we use that to tell them apart. + const bool use_stdin = params.tokenize_stdin; + const bool use_file = !params.prompt_file.empty(); - // variables where to put any arguments we see. - bool printing_ids = false; - bool no_bos = false; - bool no_escape = false; - bool no_parse_special = false; - bool disable_logging = false; - bool show_token_count = false; - const char * model_path = NULL; - const char * prompt_path = NULL; - const char * prompt_arg = NULL; - - // track which arguments were explicitly given - // used for sanity checking down the line - bool model_path_set = false; - bool prompt_path_set = false; - bool prompt_set = false; - bool stdin_set = false; - - int iarg = 1; - for (; iarg < argc; ++iarg) { - std::string arg{argv[iarg]}; - if (arg == "-h" || arg == "--help") { - print_usage_information(argv[0].c_str()); - return 0; - } - else if (arg == "--ids") { - printing_ids = true; - } - else if (arg == "-m" || arg == "--model") { - if (model_path_set) { - fprintf(stderr, "Error: -m or --model specified multiple times.\n"); - return 1; - } - model_path = argv[++iarg].c_str(); - model_path_set = true; - } - else if (arg == "--no-bos") { - no_bos = true; - } - else if (arg == "--no-escape") { - no_escape = true; - } - else if (arg == "--no-parse-special") { - no_parse_special = true; - } - else if (arg == "-p" || arg == "--prompt") { - if (prompt_set) { - fprintf(stderr, "Error: -p or --prompt specified multiple times.\n"); - return 1; - } - prompt_arg = argv[++iarg].c_str(); - prompt_set = true; - } - else if (arg == "-f" || arg == "--file") { - if (prompt_path_set) { - fprintf(stderr, "Error: -f or --file specified multiple times.\n"); - return 1; - } - prompt_path = argv[++iarg].c_str(); - prompt_path_set = true; - } - else if (arg == "--stdin") { - stdin_set = true; - } - else if (arg == "--log-disable") { - disable_logging = true; - } - else if (arg == "--show-count") { - show_token_count = true; - } - else { - fprintf(stderr, "Error: unknown option '%s'\n", argv[iarg].c_str()); - return 1; - } - } - - ////// - // Sanity check the command line arguments. - ////// - - // Check that we have the required stuff set. - if (model_path_set && model_path == NULL) { - fprintf(stderr, "Error: --model requires an argument.\n"); - return 1; - } - if (!model_path_set) { - fprintf(stderr, "Error: must specify --model.\n"); - return 1; - } - if (prompt_path_set && prompt_path == NULL) { - fprintf(stderr, "Error: --file requires an argument.\n"); - return 1; - } - if (prompt_set && prompt_arg == NULL) { - fprintf(stderr, "Error: --prompt requires an argument.\n"); - return 1; - } - const int prompts_set = !!(prompt_path_set) + !!(prompt_set) + !!(stdin_set); - if (prompts_set > 1) { - fprintf(stderr, "Error: --stdin, --file and --prompt are mutually exclusive.\n"); - return 1; - } - // Must have some prompt. - if (prompts_set == 0) { - fprintf(stderr, "Error: must specify one of: --stdin, --file or --prompt.\n"); + // sanity check: --stdin is mutually exclusive with -f/--file and -p/--prompt + if (use_stdin && (use_file || !params.prompt.empty())) { + LOG_ERR("error: --stdin is mutually exclusive with --file and --prompt\n"); return 1; } - GGML_ASSERT(model_path); - GGML_ASSERT(prompt_path || prompt_arg || stdin_set); - - ////// - // Figure out where will the prompt come from. - ////// + // must have some prompt + if (!use_stdin && !use_file && params.prompt.empty()) { + LOG_ERR("error: must specify one of: --stdin, --file or --prompt\n"); + return 1; + } std::string prompt; - if (prompt_path_set) { - bool success = false; - prompt = read_prompt_from_file(prompt_path, success); - if (!success) { + if (use_file) { + // read the file verbatim: common's -f handler strips a single trailing + // newline, but for a tokenizer the input bytes must be preserved exactly + // (a trailing newline is itself a token). escapes are applied locally + // to match the behavior of -p/--prompt and --stdin. + std::ifstream in(params.prompt_file, std::ios::binary); + if (!in) { + LOG_ERR("error: could not open file '%s' for reading\n", params.prompt_file.c_str()); return 1; } - } else if (prompt_set) { - prompt = prompt_arg; - } else { - GGML_ASSERT(stdin_set); - // we read stdin *after* loading model (early exit if model cannot - // be loaded, which can be a nicer user experience) - } - - ////// - // Start actually doing the tokenizing stuff. - ////// - - if (disable_logging) { - llama_log_set(llama_log_callback_null, NULL); + std::stringstream ss; + ss << in.rdbuf(); + prompt = ss.str(); + if (params.escape) { + string_process_escapes(prompt); + } + } else if (!use_stdin) { + // -p/--prompt is already escape-processed by common_params_parse() + // (controlled by --escape/--no-escape), so use it verbatim here. + prompt = params.prompt; } + // else: we read stdin *after* loading the model (early exit if the + // model cannot be loaded, which is a nicer user experience) llama_backend_init(); + // load only the vocabulary (no weights), since tokenizing does not need them llama_model_params model_params = llama_model_default_params(); model_params.vocab_only = true; - llama_model * model = llama_model_load_from_file(model_path, model_params); + llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params); if (!model) { - fprintf(stderr, "Error: could not load model from file '%s'.\n", model_path); + LOG_ERR("error: could not load model from file '%s'.\n", params.model.path.c_str()); return 1; } @@ -352,42 +163,41 @@ int main(int raw_argc, char ** raw_argv) { llama_context_params ctx_params = llama_context_default_params(); llama_context * ctx = llama_init_from_model(model, ctx_params); if (!ctx) { - fprintf(stderr, "Error: could not create context.\n"); + LOG_ERR("error: could not create context.\n"); return 1; } // read entire prompt from stdin? - if (stdin_set) { - GGML_ASSERT(!prompt_path_set && !prompt_set); - + if (params.tokenize_stdin) { std::stringstream stdin_buffer; stdin_buffer << std::cin.rdbuf(); if (std::cin.fail()) { - fprintf(stderr, "Error: could not read the entire standard input.\n"); + LOG_ERR("error: could not read the entire standard input.\n"); return 1; } prompt = stdin_buffer.str(); + + // stdin is not seen by common_params_parse(), so apply escape handling + // here to match the behavior of -p/--prompt and -f/--file. + if (params.escape) { + string_process_escapes(prompt); + } } const bool model_wants_add_bos = llama_vocab_get_add_bos(vocab); - const bool add_bos = model_wants_add_bos && !no_bos; - const bool parse_special = !no_parse_special; - const bool escape = !no_escape; - - if (escape) { - string_process_escapes(prompt); - } + const bool add_bos = model_wants_add_bos && !params.tokenize_no_bos; + const bool parse_special = params.parse_special; std::vector tokens; tokens = common_tokenize(vocab, prompt, add_bos, parse_special); - if (printing_ids) { + if (params.tokenize_ids) { printf("["); } for (int i = 0; i < (int) tokens.size(); i++) { - if (printing_ids) { + if (params.tokenize_ids) { if (i > 0) { printf(", "); } @@ -404,13 +214,14 @@ int main(int raw_argc, char ** raw_argv) { } } - if (printing_ids) { + if (params.tokenize_ids) { printf("]\n"); } - if (show_token_count) { + if (params.tokenize_show_count) { printf("Total number of tokens: %zu\n", tokens.size()); } + // silence valgrind llama_free(ctx); llama_model_free(model); From 7cbd61002d0815d440a48db011360998ca61346b Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Tue, 14 Jul 2026 08:26:55 -0500 Subject: [PATCH 13/24] vulkan/cpu: Support f16 as SET_ROWS src. (#25432) * vulkan/cpu: Support f16 as SET_ROWS src. This adds full support for f16 SET_ROWS (equivalent to f32) to vulkan and CPU backends, and adds more backend tests. * Set DenormPreserve 16 when supported, to try to fix failures on Intel * tune error threshold * update metal supports_op --- ggml/src/ggml-cpu/ggml-cpu.c | 6 + ggml/src/ggml-cpu/ops.cpp | 27 ++-- ggml/src/ggml-metal/ggml-metal-device.m | 6 +- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 119 +++++++++++------- .../vulkan-shaders/copy_to_quant.comp | 52 ++++---- .../vulkan-shaders/vulkan-shaders-gen.cpp | 10 +- tests/test-backend-ops.cpp | 30 +++-- 7 files changed, 150 insertions(+), 100 deletions(-) diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index 2745a7dbb..d9347e3c2 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -2863,6 +2863,12 @@ struct ggml_cplan ggml_graph_plan( cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; } } break; + case GGML_OP_SET_ROWS: + { + if (node->src[0]->type == GGML_TYPE_F16 && node->type != GGML_TYPE_F16) { + cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; + } + } break; case GGML_OP_SOFT_MAX: case GGML_OP_ROPE: case GGML_OP_ROPE_BACK: diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp index 7c54cb6f4..c49719374 100644 --- a/ggml/src/ggml-cpu/ops.cpp +++ b/ggml/src/ggml-cpu/ops.cpp @@ -5041,7 +5041,7 @@ static void ggml_compute_forward_set_rows_impl( assert(ne0 == nc); assert(ne2 == ne02); assert(ne3 == ne03); - GGML_ASSERT(src0->type == GGML_TYPE_F32 || (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16)); + GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); assert(ne02 % ne11 == 0); assert(ne03 % ne12 == 0); @@ -5075,10 +5075,19 @@ static void ggml_compute_forward_set_rows_impl( (const float *) ((char *) src0->data + i*nb01 + i02*nb02 + i03*nb03), ((char *) dst->data + i1*nb1 + i02*nb2 + i03*nb3), nc); } else if constexpr (std::is_same_v) { - memcpy( + if (dst->type == GGML_TYPE_F16) { + memcpy( ((char *) dst->data + i1*nb1 + i02*nb2 + i03*nb3), ((char *) src0->data + i*nb01 + i02*nb02 + i03*nb03), rs); + } else { + float * wdata = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith; + ggml_fp16_to_fp32_row( + (const ggml_fp16_t *) ((char *) src0->data + i*nb01 + i02*nb02 + i03*nb03), + wdata, nc); + from_float(wdata, + ((char *) dst->data + i1*nb1 + i02*nb2 + i03*nb3), nc); + } } else { GGML_ABORT("src0->type = %d (%s) not supported", src0->type, ggml_type_name(src0->type)); } @@ -5107,16 +5116,12 @@ void ggml_compute_forward_set_rows( } break; case GGML_TYPE_F16: { - if (dst->type == GGML_TYPE_F16) { - if (src1->type == GGML_TYPE_I64) { - ggml_compute_forward_set_rows_impl(params, dst); - } else if (src1->type == GGML_TYPE_I32) { - ggml_compute_forward_set_rows_impl(params, dst); - } else { - GGML_ABORT("src1->type = %d (%s) not supported", src1->type, ggml_type_name(src1->type)); - } + if (src1->type == GGML_TYPE_I64) { + ggml_compute_forward_set_rows_impl(params, dst); + } else if (src1->type == GGML_TYPE_I32) { + ggml_compute_forward_set_rows_impl(params, dst); } else { - GGML_ABORT("dst->type = %d (%s) not supported with src0->type = %d (%s)", dst->type, ggml_type_name(dst->type), src0->type, ggml_type_name(src0->type)); + GGML_ABORT("src1->type = %d (%s) not supported", src1->type, ggml_type_name(src1->type)); } } break; default: diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m index 1dfe0bdd5..80e47f2c2 100644 --- a/ggml/src/ggml-metal/ggml-metal-device.m +++ b/ggml/src/ggml-metal/ggml-metal-device.m @@ -1340,7 +1340,11 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te return op->src[0]->type != GGML_TYPE_NVFP4; case GGML_OP_SET_ROWS: { - if (op->src[0]->type != GGML_TYPE_F32 && op->src[0]->type != GGML_TYPE_F16) { + if (op->src[0]->type == GGML_TYPE_F16) { + return op->type == GGML_TYPE_F16; + } + + if (op->src[0]->type != GGML_TYPE_F32) { return false; } diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 932c561e0..1704da07e 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -723,6 +723,7 @@ struct vk_device_struct { bool uma; bool prefer_host_memory; bool float_controls_rte_fp16; + bool float_controls_denorm_preserve_fp16; bool subgroup_basic; bool subgroup_arithmetic; bool subgroup_shuffle; @@ -868,8 +869,9 @@ struct vk_device_struct { vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT]; vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT]; vk_pipeline pipeline_cpy_transpose_16, pipeline_cpy_transpose_32; - vk_pipeline pipeline_set_rows_i32[GGML_TYPE_COUNT]; - vk_pipeline pipeline_set_rows_i64[GGML_TYPE_COUNT]; + // [src0 0=fp32,1=fp16][dst] + vk_pipeline pipeline_set_rows_i32[2][GGML_TYPE_COUNT]; + vk_pipeline pipeline_set_rows_i64[2][GGML_TYPE_COUNT]; vk_pipeline pipeline_norm_f32; vk_pipeline pipeline_group_norm_f32; vk_pipeline pipeline_rms_norm_f32; @@ -2595,10 +2597,10 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin vk::ShaderModuleCreateInfo shader_module_create_info({}, spv_size, reinterpret_cast(spv_data)); - // Patch SPIR-V to enable RTE rounding for FP16, avoiding the need for - // separate shader variants compiled with -DRTE16. + // Patch SPIR-V to enable supported FP16 float controls, avoiding the need + // for separate shader variants. std::vector spirv; - if (device->float_controls_rte_fp16) { + if (device->float_controls_rte_fp16 || device->float_controls_denorm_preserve_fp16) { const uint32_t* spv_words = reinterpret_cast(spv_data); size_t word_count = spv_size / sizeof(uint32_t); spirv.assign(spv_words, spv_words + word_count); @@ -2635,9 +2637,17 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin // Insert from latest position first so earlier indices stay valid. - // OpExecutionMode %entrypoint RoundingModeRTE 16 - uint32_t exec_mode[] = { (4u << spv::WordCountShift) | spv::OpExecutionMode, entry_point_id, spv::ExecutionModeRoundingModeRTE, 16 }; - spirv.insert(spirv.begin() + exec_insert_pos, std::begin(exec_mode), std::end(exec_mode)); + if (device->float_controls_rte_fp16) { + // OpExecutionMode %entrypoint RoundingModeRTE 16 + uint32_t exec_mode[] = { (4u << spv::WordCountShift) | spv::OpExecutionMode, entry_point_id, spv::ExecutionModeRoundingModeRTE, 16 }; + spirv.insert(spirv.begin() + exec_insert_pos, std::begin(exec_mode), std::end(exec_mode)); + } + + if (device->float_controls_denorm_preserve_fp16) { + // OpExecutionMode %entrypoint DenormPreserve 16 + uint32_t exec_mode[] = { (4u << spv::WordCountShift) | spv::OpExecutionMode, entry_point_id, spv::ExecutionModeDenormPreserve, 16 }; + spirv.insert(spirv.begin() + exec_insert_pos, std::begin(exec_mode), std::end(exec_mode)); + } // OpExtension "SPV_KHR_float_controls" const char ext_str[] = "SPV_KHR_float_controls"; @@ -2647,9 +2657,17 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin memcpy(&extension[1], ext_str, sizeof(ext_str)); spirv.insert(spirv.begin() + ext_insert_pos, extension.begin(), extension.end()); - // OpCapability RoundingModeRTE - uint32_t capability[] = { (2u << spv::WordCountShift) | spv::OpCapability, spv::CapabilityRoundingModeRTE }; - spirv.insert(spirv.begin() + cap_insert_pos, std::begin(capability), std::end(capability)); + if (device->float_controls_rte_fp16) { + // OpCapability RoundingModeRTE + uint32_t capability[] = { (2u << spv::WordCountShift) | spv::OpCapability, spv::CapabilityRoundingModeRTE }; + spirv.insert(spirv.begin() + cap_insert_pos, std::begin(capability), std::end(capability)); + } + + if (device->float_controls_denorm_preserve_fp16) { + // OpCapability DenormPreserve + uint32_t capability[] = { (2u << spv::WordCountShift) | spv::OpCapability, spv::CapabilityDenormPreserve }; + spirv.insert(spirv.begin() + cap_insert_pos, std::begin(capability), std::end(capability)); + } shader_module_create_info = vk::ShaderModuleCreateInfo({}, spirv.size() * sizeof(uint32_t), spirv.data()); } @@ -5187,20 +5205,22 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) { ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1); -#define SET_ROWS(itype) \ - ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_F32], "set_rows_f32" #itype, set_rows_f32 ## itype ## _len, set_rows_f32 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ - ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_F16], "set_rows_f16" #itype, set_rows_f16 ## itype ## _len, set_rows_f16 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ - ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_BF16], "set_rows_bf16" #itype, set_rows_bf16 ## itype ## _len, set_rows_bf16 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ - ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q1_0], "set_rows_q1_0" #itype, set_rows_q1_0 ## itype ## _len, set_rows_q1_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ - ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q4_0], "set_rows_q4_0" #itype, set_rows_q4_0 ## itype ## _len, set_rows_q4_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ - ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q4_1], "set_rows_q4_1" #itype, set_rows_q4_1 ## itype ## _len, set_rows_q4_1 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ - ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q5_0], "set_rows_q5_0" #itype, set_rows_q5_0 ## itype ## _len, set_rows_q5_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ - ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q5_1], "set_rows_q5_1" #itype, set_rows_q5_1 ## itype ## _len, set_rows_q5_1 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ - ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q8_0], "set_rows_q8_0" #itype, set_rows_q8_0 ## itype ## _len, set_rows_q8_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ - ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_IQ4_NL], "set_rows_iq4_nl" #itype, set_rows_iq4_nl ## itype ## _len, set_rows_iq4_nl ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); +#define SET_ROWS(src_idx, src, itype) \ + ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_F32], "set_rows_" #src "_f32" #itype, set_rows_ ## src ## _f32 ## itype ## _len, set_rows_ ## src ## _f32 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_F16], "set_rows_" #src "_f16" #itype, set_rows_ ## src ## _f16 ## itype ## _len, set_rows_ ## src ## _f16 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_BF16], "set_rows_" #src "_bf16" #itype, set_rows_ ## src ## _bf16 ## itype ## _len, set_rows_ ## src ## _bf16 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q1_0], "set_rows_" #src "_q1_0" #itype, set_rows_ ## src ## _q1_0 ## itype ## _len, set_rows_ ## src ## _q1_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q4_0], "set_rows_" #src "_q4_0" #itype, set_rows_ ## src ## _q4_0 ## itype ## _len, set_rows_ ## src ## _q4_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q4_1], "set_rows_" #src "_q4_1" #itype, set_rows_ ## src ## _q4_1 ## itype ## _len, set_rows_ ## src ## _q4_1 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q5_0], "set_rows_" #src "_q5_0" #itype, set_rows_ ## src ## _q5_0 ## itype ## _len, set_rows_ ## src ## _q5_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q5_1], "set_rows_" #src "_q5_1" #itype, set_rows_ ## src ## _q5_1 ## itype ## _len, set_rows_ ## src ## _q5_1 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q8_0], "set_rows_" #src "_q8_0" #itype, set_rows_ ## src ## _q8_0 ## itype ## _len, set_rows_ ## src ## _q8_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_IQ4_NL], "set_rows_" #src "_iq4_nl" #itype, set_rows_ ## src ## _iq4_nl ## itype ## _len, set_rows_ ## src ## _iq4_nl ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); - SET_ROWS(_i32) - SET_ROWS(_i64) + SET_ROWS(0, f32, _i32) + SET_ROWS(0, f32, _i64) + SET_ROWS(1, f16, _i32) + SET_ROWS(1, f16, _i64) #undef SET_ROWS @@ -6031,6 +6051,7 @@ static vk_device ggml_vk_get_device(size_t idx) { device->shader_core_count = 0; } device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16; + device->float_controls_denorm_preserve_fp16 = vk12_props.shaderDenormPreserveFloat16; device->subgroup_basic = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) && (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eBasic); @@ -10843,10 +10864,17 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const case GGML_OP_DUP: return ggml_vk_get_cpy_pipeline(ctx, src0, dst, dst->type); case GGML_OP_SET_ROWS: - if (src1->type == GGML_TYPE_I64) { - return ctx->device->pipeline_set_rows_i64[dst->type]; - } else { - return ctx->device->pipeline_set_rows_i32[dst->type]; + { + if (src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) { + return nullptr; + } + const int src_idx = src0->type == GGML_TYPE_F16; + if (src1->type == GGML_TYPE_I64) { + return ctx->device->pipeline_set_rows_i64[src_idx][dst->type]; + } else if (src1->type == GGML_TYPE_I32) { + return ctx->device->pipeline_set_rows_i32[src_idx][dst->type]; + } + return nullptr; } case GGML_OP_SILU_BACK: if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { @@ -17500,24 +17528,25 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32; case GGML_OP_SET_ROWS: { - if (op->src[0]->type == GGML_TYPE_F32) { - switch (op->type) { - case GGML_TYPE_F32: - case GGML_TYPE_F16: - case GGML_TYPE_BF16: - case GGML_TYPE_Q1_0: - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_Q5_0: - case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: - case GGML_TYPE_IQ4_NL: - return true; - default: - return false; - } + if ((op->src[0]->type != GGML_TYPE_F32 && op->src[0]->type != GGML_TYPE_F16) || + (op->src[1]->type != GGML_TYPE_I32 && op->src[1]->type != GGML_TYPE_I64)) { + return false; + } + switch (op->type) { + case GGML_TYPE_F32: + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_Q1_0: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_IQ4_NL: + return true; + default: + return false; } - return false; } case GGML_OP_CONT: case GGML_OP_CPY: diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp index 710c15296..c92798f8e 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp @@ -10,7 +10,7 @@ layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; const uint BLOCK_SIZE = 32; #endif -layout (binding = 0) readonly buffer S {float data_s[];}; +layout (binding = 0) readonly buffer S {S_TYPE data_s[];}; #if defined(SET_ROWS) #include "generic_binary_head.glsl" @@ -35,7 +35,7 @@ void quantize(uint dst_idx, uint src_idx) float vmax = 0.0; [[unroll]] for (int j = 0; j < QUANT_K_Q4_0; ++j) { - const float v = data_s[src_idx + j]; + const float v = float(data_s[src_idx + j]); if (amax < abs(v)) { amax = abs(v); vmax = v; @@ -48,8 +48,8 @@ void quantize(uint dst_idx, uint src_idx) data_q[dst_idx].d = float16_t(d); [[unroll]] for (int j = 0; j < QUANT_K_Q4_0/2; ++j) { - const float x0 = data_s[src_idx + 0 + j]*id; - const float x1 = data_s[src_idx + QUANT_K_Q4_0/2 + j]*id; + const float x0 = float(data_s[src_idx + 0 + j])*id; + const float x1 = float(data_s[src_idx + QUANT_K_Q4_0/2 + j])*id; const uint xi0 = min(15, int(x0 + 8.5)); const uint xi1 = min(15, int(x1 + 8.5)); @@ -66,7 +66,7 @@ void quantize(uint dst_idx, uint src_idx) float vmax = -vmin; [[unroll]] for (int j = 0; j < QUANT_K_Q4_1; ++j) { - const float v = data_s[src_idx + j]; + const float v = float(data_s[src_idx + j]); if (v < vmin) vmin = v; if (v > vmax) vmax = v; @@ -79,8 +79,8 @@ void quantize(uint dst_idx, uint src_idx) data_q[dst_idx].m = float16_t(vmin); [[unroll]] for (int j = 0; j < QUANT_K_Q4_1/2; ++j) { - const float x0 = (data_s[src_idx + 0 + j] - vmin)*id; - const float x1 = (data_s[src_idx + QUANT_K_Q4_1/2 + j] - vmin)*id; + const float x0 = (float(data_s[src_idx + 0 + j]) - vmin)*id; + const float x1 = (float(data_s[src_idx + QUANT_K_Q4_1/2 + j]) - vmin)*id; const uint xi0 = min(15, int(x0 + 0.5)); const uint xi1 = min(15, int(x1 + 0.5)); @@ -97,7 +97,7 @@ void quantize(uint dst_idx, uint src_idx) float vmax = 0.0; [[unroll]] for (int j = 0; j < QUANT_K_Q5_0; ++j) { - const float v = data_s[src_idx + j]; + const float v = float(data_s[src_idx + j]); if (amax < abs(v)) { amax = abs(v); vmax = v; @@ -111,8 +111,8 @@ void quantize(uint dst_idx, uint src_idx) uint32_t qh = 0; [[unroll]] for (int j = 0; j < QUANT_K_Q5_0/2; ++j) { - const float x0 = data_s[src_idx + 0 + j]*id; - const float x1 = data_s[src_idx + QUANT_K_Q5_0/2 + j]*id; + const float x0 = float(data_s[src_idx + 0 + j])*id; + const float x1 = float(data_s[src_idx + QUANT_K_Q5_0/2 + j])*id; const uint xi0 = min(31, int(x0 + 16.5)); const uint xi1 = min(31, int(x1 + 16.5)); @@ -129,11 +129,11 @@ void quantize(uint dst_idx, uint src_idx) #if defined(DATA_A_Q5_1) void quantize(uint dst_idx, uint src_idx) { - float min = data_s[src_idx + 0]; + float min = float(data_s[src_idx + 0]); float max = min; [[unroll]] for (int j = 1; j < QUANT_K_Q5_1; ++j) { - const float v = data_s[src_idx + j]; + const float v = float(data_s[src_idx + j]); min = v < min ? v : min; max = v > max ? v : max; } @@ -146,8 +146,8 @@ void quantize(uint dst_idx, uint src_idx) uint32_t qh = 0; [[unroll]] for (int j = 0; j < QUANT_K_Q5_1/2; ++j) { - const float x0 = (data_s[src_idx + 0 + j] - min)*id; - const float x1 = (data_s[src_idx + QUANT_K_Q5_1/2 + j] - min)*id; + const float x0 = (float(data_s[src_idx + 0 + j]) - min)*id; + const float x1 = (float(data_s[src_idx + QUANT_K_Q5_1/2 + j]) - min)*id; const uint xi0 = uint(x0 + 0.5); const uint xi1 = uint(x1 + 0.5); @@ -166,7 +166,7 @@ void quantize(uint dst_idx, uint src_idx) float amax = 0.0; // absolute max [[unroll]] for (int j = 0; j < QUANT_K_Q8_0; j++) { - const float v = data_s[src_idx + j]; + const float v = float(data_s[src_idx + j]); amax = max(amax, abs(v)); } @@ -176,7 +176,7 @@ void quantize(uint dst_idx, uint src_idx) data_q[dst_idx].d = float16_t(d); [[unroll]] for (int j = 0; j < QUANT_K_Q8_0; ++j) { - const float x0 = data_s[src_idx + j]*id; + const float x0 = float(data_s[src_idx + j])*id; data_q[dst_idx].qs[j] = int8_t(round(x0)); } @@ -189,7 +189,7 @@ void quantize(uint dst_idx, uint src_idx) float sum_abs = 0.0; [[unroll]] for (int j = 0; j < QUANT_K_Q1_0; j++) { - sum_abs += abs(data_s[src_idx + j]); + sum_abs += abs(float(data_s[src_idx + j])); } const float d = sum_abs / QUANT_K_Q1_0; @@ -201,7 +201,7 @@ void quantize(uint dst_idx, uint src_idx) } [[unroll]] for (int j = 0; j < QUANT_K_Q1_0; ++j) { - if (data_s[src_idx + j] >= 0.0) { + if (float(data_s[src_idx + j]) >= 0.0) { data_q[dst_idx].qs[j / 8] |= uint8_t(1 << (j % 8)); } } @@ -226,7 +226,7 @@ void quantize(uint dst_idx, uint src_idx) float vmax = 0.0; [[unroll]] for (int j = 0; j < QUANT_K_IQ4_NL; ++j) { - const float v = data_s[src_idx + j]; + const float v = float(data_s[src_idx + j]); if (amax < abs(v)) { amax = abs(v); vmax = v; @@ -238,16 +238,16 @@ void quantize(uint dst_idx, uint src_idx) float sumqx = 0, sumq2 = 0; [[unroll]] for (int j = 0; j < QUANT_K_IQ4_NL/2; ++j) { - const float x0 = data_s[src_idx + 0 + j]*id; - const float x1 = data_s[src_idx + QUANT_K_IQ4_NL/2 + j]*id; + const float x0 = float(data_s[src_idx + 0 + j])*id; + const float x1 = float(data_s[src_idx + QUANT_K_IQ4_NL/2 + j])*id; const uint xi0 = best_index(x0); const uint xi1 = best_index(x1); data_q[dst_idx].qs[j] = uint8_t(xi0 | (xi1 << 4)); const float v0 = kvalues_iq4nl[xi0]; const float v1 = kvalues_iq4nl[xi1]; - const float w0 = data_s[src_idx + 0 + j]*data_s[src_idx + 0 + j]; - const float w1 = data_s[src_idx + QUANT_K_IQ4_NL/2 + j]*data_s[src_idx + QUANT_K_IQ4_NL/2 + j]; - sumqx += w0*v0*data_s[src_idx + j] + w1*v1*data_s[src_idx + QUANT_K_IQ4_NL/2 + j]; + const float w0 = float(data_s[src_idx + 0 + j])*float(data_s[src_idx + 0 + j]); + const float w1 = float(data_s[src_idx + QUANT_K_IQ4_NL/2 + j])*float(data_s[src_idx + QUANT_K_IQ4_NL/2 + j]); + sumqx += w0*v0*float(data_s[src_idx + j]) + w1*v1*float(data_s[src_idx + QUANT_K_IQ4_NL/2 + j]); sumq2 += w0*v0*v0 + w1*v1*v1; } @@ -259,14 +259,14 @@ void quantize(uint dst_idx, uint src_idx) #if defined(DATA_A_F32) || defined(DATA_A_F16) void quantize(uint dst_idx, uint src_idx) { - data_q[dst_idx] = A_TYPE(data_s[src_idx]); + data_q[dst_idx] = A_TYPE(float(data_s[src_idx])); } #endif #if defined(DATA_A_BF16) void quantize(uint dst_idx, uint src_idx) { - data_q[dst_idx] = A_TYPE(fp32_to_bf16(data_s[src_idx])); + data_q[dst_idx] = A_TYPE(fp32_to_bf16(float(data_s[src_idx]))); } #endif diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index 7d09ac4a5..240e1d1b3 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -824,13 +824,15 @@ void process_shaders() { string_to_spv("cpy_transpose_32", "copy_transpose.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}}); for (std::string t : {"q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) { - string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"S_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); } - for (std::string t : {"f32", "f16", "bf16", "q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) { - string_to_spv("set_rows_" + t + "_i32", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uint"}, {"B_SIZE", "32"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); - string_to_spv("set_rows_" + t + "_i64", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"B_SIZE", "64"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + for (auto src : {std::pair{"f32", "float"}, std::pair{"f16", "float16_t"}}) { + for (std::string dst : {"f32", "f16", "bf16", "q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) { + string_to_spv("set_rows_" + std::string(src.first) + "_" + dst + "_i32", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(dst), "1"}, {"B_TYPE", "uint"}, {"B_SIZE", "32"}, {"S_TYPE", src.second}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + string_to_spv("set_rows_" + std::string(src.first) + "_" + dst + "_i64", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(dst), "1"}, {"B_TYPE", "uvec2"}, {"B_SIZE", "64"}, {"S_TYPE", src.second}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + } } auto get_type_str = [](bool f16) { diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index a83cb302b..ae49d2d0d 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -2423,11 +2423,10 @@ struct test_set_rows : public test_case { void initialize_tensors(ggml_context * ctx) override { for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { + if (ggml_is_view_op(t->op)) { + continue; + } if (t->type == GGML_TYPE_I64 || t->type == GGML_TYPE_I32) { - if (ggml_is_view_op(t->op)) { - continue; - } - init_set_rows_row_ids(t, ne[1]); } else { init_tensor_uniform(t); @@ -2450,6 +2449,9 @@ struct test_set_rows : public test_case { err_estimate /= 8.0f; } err_estimate *= err_estimate; + if (type_src == GGML_TYPE_F16) { + err_estimate *= 16.0f; + } err_estimate /= 0.25f*float(ne[0] * r * ne[2]*nr23[0] * ne[3]*nr23[1]); return err_estimate; } @@ -7928,17 +7930,19 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_I64, { 1, 8, 1, 3 }, { 1, 1 }, 2, false)); test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_I32, { 1, 8, 1, 3 }, { 1, 1 }, 2, false)); test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_Q8_0, GGML_TYPE_I32, { 256, 5, 1, 3 }, { 1, 1, }, 1, false)); - for (ggml_type type : all_types) { - for (int b : {1, 7}) { - for (bool v : {false, true}) { - test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 256, 5, b, 3 }, { 1, 1, }, 1, v)); - test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 256, 11, 1, b }, { 2, 3, }, 7, v)); + for (ggml_type src_type : {GGML_TYPE_F16, GGML_TYPE_F32}) { + for (ggml_type type : all_types) { + for (int b : {1, 7}) { + for (bool v : {false, true}) { + test_cases.emplace_back(new test_set_rows(src_type, type, GGML_TYPE_I64, { 256, 5, b, 3 }, { 1, 1, }, 1, v)); + test_cases.emplace_back(new test_set_rows(src_type, type, GGML_TYPE_I64, { 256, 11, 1, b }, { 2, 3, }, 7, v)); - test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 3*ggml_blck_size(type), 3, b, 1 }, { 2, 3, }, 2, v)); + test_cases.emplace_back(new test_set_rows(src_type, type, GGML_TYPE_I64, { 3*ggml_blck_size(type), 3, b, 1 }, { 2, 3, }, 2, v)); - if (ggml_blck_size(type) == 1) { - test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 31, 3, b, 1 }, { 2, 3, }, 2, v)); - test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 33, 5, 1, b }, { 2, 3, }, 1, v)); + if (ggml_blck_size(type) == 1) { + test_cases.emplace_back(new test_set_rows(src_type, type, GGML_TYPE_I64, { 31, 3, b, 1 }, { 2, 3, }, 2, v)); + test_cases.emplace_back(new test_set_rows(src_type, type, GGML_TYPE_I64, { 33, 5, 1, b }, { 2, 3, }, 1, v)); + } } } } From 7f575c39d6a29a40c0ef22278eca6bd4a573c8a6 Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Tue, 14 Jul 2026 21:45:36 +0800 Subject: [PATCH 14/24] DeepseekV4: fix seq_rm (#25588) * DeepseekV4: fix seq_rm * implement proper seq_cp * create actual update context --- src/llama-kv-cache-dsv4.cpp | 100 ++++++++++++++++++++++++++++++------ src/llama-kv-cache-dsv4.h | 25 +++++++-- 2 files changed, 106 insertions(+), 19 deletions(-) diff --git a/src/llama-kv-cache-dsv4.cpp b/src/llama-kv-cache-dsv4.cpp index ebafac091..7cb6cc18d 100644 --- a/src/llama-kv-cache-dsv4.cpp +++ b/src/llama-kv-cache-dsv4.cpp @@ -720,7 +720,7 @@ llama_dsv4_comp_state::llama_dsv4_comp_state( auto it = ctx_map.find(buft); if (it == ctx_map.end()) { ggml_init_params params = { - /*.mem_size =*/ size_t(2u*hparams.n_layer()*ggml_tensor_overhead()), + /*.mem_size =*/ size_t(2u*(1 + n_stream)*hparams.n_layer()*ggml_tensor_overhead()), /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, }; @@ -767,9 +767,17 @@ llama_dsv4_comp_state::llama_dsv4_comp_state( ggml_format_name(kv, "dsv4_%s_state_kv_l%d", name, il); ggml_format_name(score, "dsv4_%s_state_score_l%d", name, il); + std::vector kv_stream; + std::vector score_stream; + + for (uint32_t s = 0; s < n_stream; ++s) { + kv_stream.push_back(ggml_view_2d(ctx, kv, n_embd_state, state_size, kv->nb[1], s*kv->nb[2])); + score_stream.push_back(ggml_view_2d(ctx, score, n_embd_state, state_size, score->nb[1], s*score->nb[2])); + } + map_layer_ids[il] = layers.size(); - layers.push_back({ il, kv, score }); + layers.push_back({ il, kv, score, std::move(kv_stream), std::move(score_stream) }); } for (auto & [buft, ctx] : ctx_map) { @@ -809,6 +817,30 @@ void llama_dsv4_comp_state::clear(llama_seq_id seq_id, bool data) { } } +void llama_dsv4_comp_state::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst) { + GGML_ASSERT(seq_id_src >= 0 && (uint32_t) seq_id_src < n_stream); + GGML_ASSERT(seq_id_dst >= 0 && (uint32_t) seq_id_dst < n_stream); + + if (seq_id_src == seq_id_dst) { + return; + } + + sc_info.ssrc.push_back((uint32_t) seq_id_src); + sc_info.sdst.push_back((uint32_t) seq_id_dst); +} + +void llama_dsv4_comp_state::apply_copies(const stream_copy_info & sc_info) const { + for (size_t i = 0; i < sc_info.ssrc.size(); ++i) { + const uint32_t ssrc = sc_info.ssrc[i]; + const uint32_t sdst = sc_info.sdst[i]; + + for (const auto & layer : layers) { + ggml_backend_tensor_copy(layer.kv_stream[ssrc], layer.kv_stream[sdst]); + ggml_backend_tensor_copy(layer.score_stream[ssrc], layer.score_stream[sdst]); + } + } +} + uint32_t llama_dsv4_comp_state::get_ratio() const { return ratio; } @@ -1154,7 +1186,13 @@ llama_memory_context_ptr llama_kv_cache_dsv4::init_full() { } llama_memory_context_ptr llama_kv_cache_dsv4::init_update(llama_context * lctx, bool optimize) { - return std::make_unique(this, lctx, optimize); + return std::make_unique( + this, + lctx, + optimize, + std::move(csa_state->sc_info), + std::move(hca_state->sc_info), + std::move(lid_state->sc_info)); } bool llama_kv_cache_dsv4::get_can_shift() const { @@ -1174,14 +1212,19 @@ bool llama_kv_cache_dsv4::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1 } if (p0 > 0) { - // DSV4 compressed cache rows are derived from running compressor state, - // so arbitrary rollback is not reconstructible from the raw cache alone. - // Allow the common prompt-cache cleanup no-op: remove [end, infinity). - if (seq_id >= 0 && p0 > kv_raw->seq_pos_max(seq_id)) { - return true; + if (seq_id < 0 || (uint32_t) seq_id >= n_seq_max || + p0 <= kv_raw->seq_pos_max(seq_id)) { + return false; } - return false; + bool res = true; + + res = res & kv_raw->seq_rm(seq_id, p0, -1); + res = res & kv_csa->seq_rm(seq_id, p0/DSV4_CSA_RATIO, -1); + res = res & kv_hca->seq_rm(seq_id, p0/DSV4_HCA_RATIO, -1); + res = res & kv_lid->seq_rm(seq_id, p0/DSV4_CSA_RATIO, -1); + + return res; } const bool res = kv_raw->seq_rm(seq_id, p0, p1); @@ -1194,7 +1237,16 @@ bool llama_kv_cache_dsv4::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1 } void llama_kv_cache_dsv4::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) { + GGML_ASSERT(p0 <= 0 && p1 < 0 && "DSV4 only supports full sequence copies"); + kv_raw->seq_cp(seq_id_src, seq_id_dst, p0, p1); + kv_csa->seq_cp(seq_id_src, seq_id_dst, -1, -1); + kv_hca->seq_cp(seq_id_src, seq_id_dst, -1, -1); + kv_lid->seq_cp(seq_id_src, seq_id_dst, -1, -1); + + csa_state->seq_cp(seq_id_src, seq_id_dst); + hca_state->seq_cp(seq_id_src, seq_id_dst); + lid_state->seq_cp(seq_id_src, seq_id_dst); } void llama_kv_cache_dsv4::seq_keep(llama_seq_id seq_id) { @@ -1639,20 +1691,26 @@ llama_kv_cache_dsv4_context::llama_kv_cache_dsv4_context( llama_kv_cache_dsv4_context::llama_kv_cache_dsv4_context( llama_kv_cache_dsv4 * kv, llama_context * lctx, - bool optimize) : + bool optimize, + stream_copy_info sc_info_csa, + stream_copy_info sc_info_hca, + stream_copy_info sc_info_lid) : ctx_raw(std::make_unique(kv->get_raw(), lctx, optimize)), ctx_csa_mem(kv->get_csa()->init_update(lctx, optimize)), ctx_hca_mem(kv->get_hca()->init_update(lctx, optimize)), ctx_lid_mem(kv->get_lid()->init_update(lctx, optimize)), - ctx_csa(std::make_unique(kv->get_csa())), - ctx_hca(std::make_unique(kv->get_hca())), - ctx_lid(std::make_unique(kv->get_lid())), csa_state(kv->get_csa_state()), hca_state(kv->get_hca_state()), lid_state(kv->get_lid_state()), + sc_info_csa(std::move(sc_info_csa)), + sc_info_hca(std::move(sc_info_hca)), + sc_info_lid(std::move(sc_info_lid)), status(llama_memory_status_combine( - llama_memory_status_combine(ctx_raw->get_status(), ctx_csa_mem->get_status()), - llama_memory_status_combine(ctx_hca_mem->get_status(), ctx_lid_mem->get_status()))) { + llama_memory_status_combine( + llama_memory_status_combine(ctx_raw->get_status(), ctx_csa_mem->get_status()), + llama_memory_status_combine(ctx_hca_mem->get_status(), ctx_lid_mem->get_status())), + this->sc_info_csa.empty() && this->sc_info_hca.empty() && this->sc_info_lid.empty() ? + LLAMA_MEMORY_STATUS_NO_UPDATE : LLAMA_MEMORY_STATUS_SUCCESS)) { } llama_kv_cache_dsv4_context::llama_kv_cache_dsv4_context( @@ -1720,6 +1778,18 @@ bool llama_kv_cache_dsv4_context::apply() { res = res & ctx_raw->apply(); + if (ctx_csa_mem) { + res = res & ctx_csa_mem->apply(); + res = res & ctx_hca_mem->apply(); + res = res & ctx_lid_mem->apply(); + } + + if (ubatches.empty()) { + csa_state->apply_copies(sc_info_csa); + hca_state->apply_copies(sc_info_hca); + lid_state->apply_copies(sc_info_lid); + } + return res; } diff --git a/src/llama-kv-cache-dsv4.h b/src/llama-kv-cache-dsv4.h index 91619ce9e..76b1daf57 100644 --- a/src/llama-kv-cache-dsv4.h +++ b/src/llama-kv-cache-dsv4.h @@ -10,6 +10,10 @@ class llama_dsv4_comp_state { public: + using stream_copy_info = llama_kv_cache::stream_copy_info; + + stream_copy_info sc_info; + llama_dsv4_comp_state( const llama_model & model, bool offload, @@ -22,6 +26,8 @@ public: const llama_memory_i::layer_filter_cb & filter); void clear(llama_seq_id seq_id, bool data); + void seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst); + void apply_copies(const stream_copy_info & sc_info) const; uint32_t get_ratio() const; uint32_t get_state_size() const; @@ -44,6 +50,9 @@ private: ggml_tensor * kv; ggml_tensor * score; + + std::vector kv_stream; + std::vector score_stream; }; const uint32_t ratio; @@ -245,6 +254,7 @@ private: class llama_kv_cache_dsv4_context : public llama_memory_context_i { public: using slot_info_vec_t = llama_kv_cache::slot_info_vec_t; + using stream_copy_info = llama_kv_cache::stream_copy_info; struct comp_plan { // Per-ubatch recipe for updating compressor state, committing completed @@ -291,7 +301,10 @@ public: llama_kv_cache_dsv4_context( llama_kv_cache_dsv4 * kv, llama_context * lctx, - bool optimize); + bool optimize, + stream_copy_info sc_info_csa, + stream_copy_info sc_info_hca, + stream_copy_info sc_info_lid); llama_kv_cache_dsv4_context( llama_kv_cache_dsv4 * kv, @@ -351,9 +364,13 @@ private: const std::unique_ptr ctx_hca; const std::unique_ptr ctx_lid; - const llama_dsv4_comp_state * csa_state = nullptr; - const llama_dsv4_comp_state * hca_state = nullptr; - const llama_dsv4_comp_state * lid_state = nullptr; + llama_dsv4_comp_state * csa_state = nullptr; + llama_dsv4_comp_state * hca_state = nullptr; + llama_dsv4_comp_state * lid_state = nullptr; + + stream_copy_info sc_info_csa; + stream_copy_info sc_info_hca; + stream_copy_info sc_info_lid; bool reserve_plans = false; mutable comp_plan reserve_plan_csa; From 17a05e451fd175d1d82a1ee96b266338f6b07005 Mon Sep 17 00:00:00 2001 From: Pascal Date: Tue, 14 Jul 2026 16:50:44 +0200 Subject: [PATCH 15/24] ui: fix mcp panel for toggle + timeout + proxy + ON/OFF state (#25631) * ui: fix MCP panel regressions after settings rework Restore the llama-server proxy switch in the Add New Server dialog. The dialog never passed useProxy/onUseProxyChange to McpServerForm, which only renders the proxy switch when the handler is provided. The flag is now wired, persisted on addServer, and reset on close. Bound the MCP connection handshake with the configured timeout. handshakeTimeoutMs was set in the server config but never consumed. The SDK timeout only covers the initialize request, not transport.start(), which can hang forever on an unreachable host. The whole handshake now races against the timeout and closes the transport on expiry so the underlying fetch or socket is aborted. Keep disabled MCP servers visible in management and chat-add UIs. Collapsing mcpDefaultServerOverrides into mcpServers[i].enabled turned the visibleMcpServers enabled filter into a visibility trap: toggling a server off outside a conversation hid it from every surface with no way to re-enable it. The filter is dropped, tools derived from health checks still skip disabled servers, and the settings page and server card render the real card instead of a skeleton for disabled servers that never receive a startup health check. * ui: clarify MCP server list semantics and add regression test Remove the visibleMcpServers getter, a filterless alias of getServers whose name invites the next refactor to put a filter back. Call sites read getServers directly, the duplicate list in the chat submenu is merged, and the misleading local variable in the sheet is renamed. A parser unit test pins the invariant: enabled is an on/off state, never a visibility filter, so disabled servers stay listed and toggleable. * ui: apply the MCP request timeout setting live to all servers The per-server requestTimeoutSeconds field was never editable in any UI and froze the global setting at server creation time, so changing the timeout in Settings was a no-op for existing servers. The field is removed from the data model and parsers, the timeout is read live from the global setting wherever a request config is built, and the misleading "Can be overridden per server" help text is dropped. A parser unit test guards against reintroducing the stored field. * ui: move the MCP request timeout into the Agentic settings section The MCP section held a single setting. The timeout is a global tool execution parameter like the other Agentic entries, so it moves there and the section is removed. Same settings key, no migration needed. * ui: remove the dead tool preview lines setting The agenticMaxToolPreviewLines setting was read into AgenticConfig and consumed by nothing: the agentic loop only uses enabled and maxTurns. Its help text described a previous architecture where only truncated previews and the final response survived the loop; tool results and intermediate turns now persist as full DB messages, so the setting had no effect at any value. Stale keys in localStorage or a server ui-config are ignored. * ui: resolve absent MCP per-chat overrides to the server enabled flag New conversations started with every MCP server off: the settings rework stopped seeding a per-conversation override list, assuming the enabled check would fall back to mcpServers[i].enabled, but it fell back to false, and the send path passed the raw stored list with no fallback at all. The per-conversation list is now sparse by contract, holding only explicit toggles, and every access point resolves a missing entry to the server's own enabled flag: the toggle display, the resolved list handed to the agentic flow, and the enabled check itself. --- .../ChatFormActionAddMcpServersSubmenu.svelte | 8 +-- .../ChatFormActionAddSheet.svelte | 8 +-- .../app/dialogs/DialogMcpServerAddNew.svelte | 7 ++- .../mcp/McpServerCard/McpServerCard.svelte | 4 +- .../app/settings/SettingsMcpServers.svelte | 15 ++++-- tools/ui/src/lib/constants/agentic.ts | 3 +- tools/ui/src/lib/constants/routes.ts | 1 - tools/ui/src/lib/constants/settings-keys.ts | 1 - .../ui/src/lib/constants/settings-registry.ts | 32 ++--------- tools/ui/src/lib/services/mcp.service.ts | 25 ++++++++- tools/ui/src/lib/stores/agentic.svelte.ts | 5 +- tools/ui/src/lib/stores/chat.svelte.ts | 2 +- .../ui/src/lib/stores/conversations.svelte.ts | 53 ++++++++----------- tools/ui/src/lib/stores/mcp.svelte.ts | 38 ++++++------- tools/ui/src/lib/stores/tools.svelte.ts | 3 +- tools/ui/src/lib/types/agentic.d.ts | 1 - tools/ui/src/lib/types/mcp.d.ts | 2 - tools/ui/src/lib/utils/mcp.ts | 5 -- .../unit/parse-mcp-server-settings.test.ts | 40 ++++++++------ 19 files changed, 126 insertions(+), 127 deletions(-) diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpServersSubmenu.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpServersSubmenu.svelte index f86222823..ea4494ce5 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpServersSubmenu.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpServersSubmenu.svelte @@ -17,10 +17,10 @@ let { onMcpSettingsClick }: Props = $props(); let mcpSearchQuery = $state(''); - let allMcpServers = $derived(mcpStore.getServers()); - let mcpServers = $derived(mcpStore.visibleMcpServers); + // Every configured server is listed; `enabled` is an on/off state, + // not a visibility filter, so a disabled server stays toggleable. + let mcpServers = $derived(mcpStore.getServers()); let hasMcpServers = $derived(mcpServers.length > 0); - // let hasAnyMcpServers = $derived(allMcpServers.length > 0); let filteredMcpServers = $derived.by(() => { const query = mcpSearchQuery.toLowerCase().trim(); if (!query) return mcpServers; @@ -46,7 +46,7 @@ function handleMcpSubMenuOpen(open: boolean) { if (open) { mcpSearchQuery = ''; - mcpStore.runHealthChecksForServers(allMcpServers); + mcpStore.runHealthChecksForServers(mcpServers); } } diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddSheet.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddSheet.svelte index c280a01a0..6ee8eb578 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddSheet.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddSheet.svelte @@ -84,7 +84,7 @@ const sheetItemRowClass = 'flex w-full items-center justify-between gap-2 rounded-md px-3 py-2 text-left text-sm transition-colors hover:bg-accent'; - let visibleMcpServers = $derived(mcpStore.visibleMcpServers); + let mcpServers = $derived(mcpStore.getServers());
@@ -218,13 +218,13 @@ MCP Servers - {visibleMcpServers.length} server{visibleMcpServers.length !== 1 ? 's' : ''} + {mcpServers.length} server{mcpServers.length !== 1 ? 's' : ''}
- {#each visibleMcpServers as server (server.id)} + {#each mcpServers as server (server.id)} {@const healthState = mcpStore.getHealthCheckState(server.id)} {@const hasError = healthState.status === HealthCheckStatus.ERROR} {@const displayName = mcpStore.getServerLabel(server)} @@ -267,7 +267,7 @@ {/each} - {#if visibleMcpServers.length === 0} + {#if mcpServers.length === 0}
No MCP servers configured
diff --git a/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte b/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte index ee99e6b5e..7438b8fd6 100644 --- a/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte +++ b/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte @@ -16,6 +16,7 @@ let newServerUrl = $state(''); let newServerHeaders = $state(''); + let newServerUseProxy = $state(false); let newServerUrlError = $derived.by(() => { if (!newServerUrl.trim()) return 'URL is required'; try { @@ -35,6 +36,7 @@ if (!value) { newServerUrl = ''; newServerHeaders = ''; + newServerUseProxy = false; } open = value; onOpenChange?.(value); @@ -49,7 +51,8 @@ id: newServerId, enabled: true, url: newServerUrl.trim(), - headers: newServerHeaders.trim() || undefined + headers: newServerHeaders.trim() || undefined, + useProxy: newServerUseProxy }); conversationsStore.setMcpServerOverride(newServerId, true); @@ -74,8 +77,10 @@ (newServerUrl = v)} onHeadersChange={(v) => (newServerHeaders = v)} + onUseProxyChange={(v) => (newServerUseProxy = v)} urlError={newServerUrl ? newServerUrlError : null} id="new-server" /> diff --git a/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCard.svelte b/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCard.svelte index d238d1ffb..f7b6e2471 100644 --- a/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCard.svelte +++ b/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCard.svelte @@ -32,7 +32,9 @@ let isHealthChecking = $derived(healthState.status === HealthCheckStatus.CONNECTING); let isConnected = $derived(healthState.status === HealthCheckStatus.SUCCESS); let isError = $derived(healthState.status === HealthCheckStatus.ERROR); - let showSkeleton = $derived(isIdle || isHealthChecking); + // Disabled servers stay IDLE (no startup health check), so the body + // skeleton only applies while a check is running or expected to run. + let showSkeleton = $derived(isHealthChecking || (isIdle && server.enabled)); let errorMessage = $derived( healthState.status === HealthCheckStatus.ERROR ? healthState.message : undefined ); diff --git a/tools/ui/src/lib/components/app/settings/SettingsMcpServers.svelte b/tools/ui/src/lib/components/app/settings/SettingsMcpServers.svelte index 120e7914f..bff2caa5b 100644 --- a/tools/ui/src/lib/components/app/settings/SettingsMcpServers.svelte +++ b/tools/ui/src/lib/components/app/settings/SettingsMcpServers.svelte @@ -22,7 +22,9 @@ let { class: className }: Props = $props(); - let servers = $derived(mcpStore.visibleMcpServers); + // Every configured server is listed; `enabled` is an on/off state, + // not a visibility filter, so a disabled server stays toggleable. + let servers = $derived(mcpStore.getServers()); let isAddingServer = $state(false); @@ -58,9 +60,14 @@ // Each card decides for itself whether to render based on its own // health-check state, so adding a server only flashes the new card // (not every other already-loaded card) until its health check resolves. - function isServerPending(serverId: string): boolean { + // Disabled servers never receive a startup health check, so IDLE only + // counts as pending when the server is enabled; otherwise the real card + // renders and keeps the enable toggle reachable. + function isServerPending(serverId: string, enabled: boolean): boolean { const status = mcpStore.getHealthCheckState(serverId).status; - return status === HealthCheckStatus.IDLE || status === HealthCheckStatus.CONNECTING; + return ( + status === HealthCheckStatus.CONNECTING || (status === HealthCheckStatus.IDLE && enabled) + ); } @@ -109,7 +116,7 @@ style="grid-template-columns: repeat(auto-fill, minmax(min(32rem, calc(100dvw - 2rem)), 1fr));" > {#each servers as server (server.id)} - {#if isServerPending(server.id)} + {#if isServerPending(server.id, server.enabled)} {:else} = { } }, { - key: SETTINGS_KEYS.AGENTIC_MAX_TOOL_PREVIEW_LINES, - label: 'Max lines per tool preview', - help: 'Number of lines shown in tool output previews (last N lines). Only these previews and the final LLM response persist after the agentic loop completes.', - defaultValue: 25, + key: SETTINGS_KEYS.MCP_REQUEST_TIMEOUT_SECONDS, + label: 'MCP request timeout (seconds)', + help: 'Timeout for individual MCP tool calls.', + defaultValue: DEFAULT_MCP_CONFIG.requestTimeoutSeconds, type: SettingsFieldType.INPUT, section: SETTINGS_SECTION_SLUGS.AGENTIC, isPositiveInteger: true, sync: { - serverKey: SETTINGS_KEYS.AGENTIC_MAX_TOOL_PREVIEW_LINES, + serverKey: SETTINGS_KEYS.MCP_REQUEST_TIMEOUT_SECONDS, paramType: SyncableParameterType.NUMBER } } @@ -735,26 +733,6 @@ const SETTINGS_REGISTRY: Record = { } } ] - }, - [SETTINGS_SECTION_SLUGS.MCP]: { - title: SETTINGS_SECTION_TITLES.MCP, - slug: SETTINGS_SECTION_SLUGS.MCP, - icon: McpLogo, - settings: [ - { - key: SETTINGS_KEYS.MCP_REQUEST_TIMEOUT_SECONDS, - label: 'Request timeout (seconds)', - help: 'Default timeout for individual MCP tool calls. Can be overridden per server.', - defaultValue: DEFAULT_MCP_CONFIG.requestTimeoutSeconds, - type: SettingsFieldType.INPUT, - section: SETTINGS_SECTION_SLUGS.MCP, - isPositiveInteger: true, - sync: { - serverKey: SETTINGS_KEYS.MCP_REQUEST_TIMEOUT_SECONDS, - paramType: SyncableParameterType.NUMBER - } - } - ] } } as const; diff --git a/tools/ui/src/lib/services/mcp.service.ts b/tools/ui/src/lib/services/mcp.service.ts index ae98632a6..e36faaac2 100644 --- a/tools/ui/src/lib/services/mcp.service.ts +++ b/tools/ui/src/lib/services/mcp.service.ts @@ -692,8 +692,31 @@ export class MCPService { this.createLog(MCPConnectionPhase.INITIALIZING, 'Sending initialize request...') ); + // The SDK timeout only covers the initialize request, not transport.start(), + // which can hang forever on an unreachable host (SSE endpoint wait, WebSocket + // handshake, proxied fetch). This race bounds the whole handshake and closes + // the transport on expiry so the underlying fetch or socket is aborted. + const handshakeTimeoutMs = + serverConfig.handshakeTimeoutMs ?? DEFAULT_MCP_CONFIG.connectionTimeoutMs; + try { - await client.connect(transport); + let handshakeTimer: ReturnType | undefined; + const handshakeDeadline = new Promise((_, reject) => { + handshakeTimer = setTimeout(() => { + void transport.close().catch(() => {}); + reject(new Error(`Connection timed out after ${Math.round(handshakeTimeoutMs / 1000)}s`)); + }, handshakeTimeoutMs); + }); + + try { + await Promise.race([ + client.connect(transport, { timeout: handshakeTimeoutMs }), + handshakeDeadline + ]); + } finally { + clearTimeout(handshakeTimer); + } + // Transport diagnostics are only for the initial handshake, not long-lived traffic. stopPhaseLogging(); client.onerror = runtimeErrorHandler; diff --git a/tools/ui/src/lib/stores/agentic.svelte.ts b/tools/ui/src/lib/stores/agentic.svelte.ts index 1a677602f..a1beb4029 100644 --- a/tools/ui/src/lib/stores/agentic.svelte.ts +++ b/tools/ui/src/lib/stores/agentic.svelte.ts @@ -280,16 +280,13 @@ class AgenticStore { getConfig(settings: SettingsConfigType, perChatOverrides?: McpServerOverride[]): AgenticConfig { const maxTurns = Number(settings.agenticMaxTurns) || DEFAULT_AGENTIC_CONFIG.maxTurns; - const maxToolPreviewLines = - Number(settings.agenticMaxToolPreviewLines) || DEFAULT_AGENTIC_CONFIG.maxToolPreviewLines; const hasTools = mcpStore.hasEnabledServers(perChatOverrides) || toolsStore.builtinTools.length > 0 || toolsStore.customTools.length > 0; return { enabled: hasTools && DEFAULT_AGENTIC_CONFIG.enabled, - maxTurns, - maxToolPreviewLines + maxTurns }; } diff --git a/tools/ui/src/lib/stores/chat.svelte.ts b/tools/ui/src/lib/stores/chat.svelte.ts index fcd07c4fd..0a496f2c7 100644 --- a/tools/ui/src/lib/stores/chat.svelte.ts +++ b/tools/ui/src/lib/stores/chat.svelte.ts @@ -1334,7 +1334,7 @@ class ChatStore { } }; - const perChatOverrides = conversationsStore.activeConversation?.mcpServerOverrides; + const perChatOverrides = conversationsStore.getAllMcpServerOverrides(); { const agenticResult = await agenticStore.runAgenticFlow({ diff --git a/tools/ui/src/lib/stores/conversations.svelte.ts b/tools/ui/src/lib/stores/conversations.svelte.ts index ea2c11e14..c7bfd00b8 100644 --- a/tools/ui/src/lib/stores/conversations.svelte.ts +++ b/tools/ui/src/lib/stores/conversations.svelte.ts @@ -243,9 +243,9 @@ class ConversationsStore { const conversationName = name || `Chat ${new Date().toLocaleString()}`; const conversation = await DatabaseService.createConversation(conversationName); - // New conversations inherit per-server enabled defaults directly from - // `mcpServers[i].enabled` (see #checkServerEnabled). No per-conversation - // override list needs to be seeded. + // No MCP override list is seeded: getAllMcpServerOverrides resolves + // servers without a per-conversation override to `mcpServers[i].enabled`, + // and only explicit toggles are stored on the conversation. // Inherit global thinking/reasoning defaults into the new conversation const thinkingEnabled = this.getThinkingEnabled(); @@ -601,48 +601,41 @@ class ConversationsStore { */ /** - /** - * Resolve the per-server enabled value when no active conversation exists. - * The default for new chats is the server's own `enabled` flag in `mcpServers`. + * Resolve the default enabled value for a server: its own `enabled` + * flag in `mcpServers`, so the global on/off state lives in one place. */ - #getDefaultOverrideForNoConversation(serverId: string): McpServerOverride | undefined { + #getDefaultOverride(serverId: string): McpServerOverride | undefined { const server = mcpStore.getServers().find((s) => s.id === serverId); if (!server) return undefined; return { serverId, enabled: server.enabled }; } /** - * Default overrides for new chats are derived from `mcpServers[i].enabled`, - * so the global on/off state lives in one place. - */ - #getAllDefaultOverridesForNoConversation(): McpServerOverride[] { - return mcpStore.getServers().map((s) => ({ serverId: s.id, enabled: s.enabled })); - } - - /** - * Gets MCP server override for a specific server in the active conversation. - * Falls back to `mcpServers[i].enabled` if no active conversation exists. + * Gets the effective MCP server override for a specific server. + * A per-conversation override wins when present; a server without one + * resolves to its `mcpServers[i].enabled` default. * @param serverId - The server ID to check - * @returns The override if set, undefined if no matching server + * @returns The effective override, undefined if no matching server */ getMcpServerOverride(serverId: string): McpServerOverride | undefined { - if (this.activeConversation) { - return this.activeConversation.mcpServerOverrides?.find( - (o: McpServerOverride) => o.serverId === serverId - ); - } - return this.#getDefaultOverrideForNoConversation(serverId); + const override = this.activeConversation?.mcpServerOverrides?.find( + (o: McpServerOverride) => o.serverId === serverId + ); + if (override) return override; + return this.#getDefaultOverride(serverId); } /** - * Get all MCP server overrides for the current conversation. - * When no active conversation, derives from `mcpServers[i].enabled`. + * Gets the effective override list for the current conversation: + * one entry per configured server, resolved per server. The stored + * per-conversation list is sparse and only holds explicit toggles. */ getAllMcpServerOverrides(): McpServerOverride[] { - if (this.activeConversation?.mcpServerOverrides) { - return this.activeConversation.mcpServerOverrides; - } - return this.#getAllDefaultOverridesForNoConversation(); + const overrides = this.activeConversation?.mcpServerOverrides; + return mcpStore.getServers().map((s) => { + const override = overrides?.find((o: McpServerOverride) => o.serverId === s.id); + return { serverId: s.id, enabled: override?.enabled ?? s.enabled }; + }); } /** diff --git a/tools/ui/src/lib/stores/mcp.svelte.ts b/tools/ui/src/lib/stores/mcp.svelte.ts index c752e7a36..4f05ad13e 100644 --- a/tools/ui/src/lib/stores/mcp.svelte.ts +++ b/tools/ui/src/lib/stores/mcp.svelte.ts @@ -148,15 +148,22 @@ class MCPStore { enabled: Boolean((entry as { enabled?: unknown })?.enabled), url, name: (entry as { name?: string })?.name, - requestTimeoutSeconds: - (entry as { requestTimeoutSeconds?: number })?.requestTimeoutSeconds ?? - DEFAULT_MCP_CONFIG.requestTimeoutSeconds, headers: headers || undefined, useProxy: Boolean((entry as { useProxy?: unknown })?.useProxy) } satisfies MCPServerSettingsEntry; }); } + /** + * Request timeout in milliseconds, read live from the global setting + * so a change in Settings applies to every server immediately. + */ + #requestTimeoutMs(): number { + const seconds = + Number(config().mcpRequestTimeoutSeconds) || DEFAULT_MCP_CONFIG.requestTimeoutSeconds; + return Math.round(seconds * 1000); + } + /** * Builds server configuration from a settings entry. */ @@ -183,7 +190,7 @@ class MCPStore { url: entry.url, transport: detectMcpTransportFromUrl(entry.url), handshakeTimeoutMs: connectionTimeoutMs, - requestTimeoutMs: Math.round(entry.requestTimeoutSeconds * 1000), + requestTimeoutMs: this.#requestTimeoutMs(), headers, useProxy: entry.useProxy }; @@ -191,15 +198,15 @@ class MCPStore { /** * Checks if a server is enabled for a given chat. - * Only per-chat overrides (persisted in localStorage for new chats, - * or in IndexedDB for existing conversations) control enabled state. + * A per-chat override wins when present; a server without one resolves + * to its own `enabled` flag in `mcpServers`. */ #checkServerEnabled( server: MCPServerSettingsEntry, perChatOverrides?: McpServerOverride[] ): boolean { const override = perChatOverrides?.find((o) => o.serverId === server.id); - return override?.enabled ?? false; + return override?.enabled ?? server.enabled; } /** @@ -230,7 +237,7 @@ class MCPStore { protocolVersion: DEFAULT_MCP_CONFIG.protocolVersion, capabilities: DEFAULT_MCP_CONFIG.capabilities, clientInfo: DEFAULT_MCP_CONFIG.clientInfo, - requestTimeoutMs: Math.round(DEFAULT_MCP_CONFIG.requestTimeoutSeconds * 1000), + requestTimeoutMs: this.#requestTimeoutMs(), servers }; } @@ -500,7 +507,7 @@ class MCPStore { } addServer( - serverData: Omit & { id?: string } + serverData: Omit & { id?: string } ): MCPServerSettingsEntry { const servers = this.getServers(); const newServer: MCPServerSettingsEntry = { @@ -509,8 +516,6 @@ class MCPStore { url: serverData.url.trim(), name: serverData.name, headers: serverData.headers?.trim() || undefined, - requestTimeoutSeconds: - Number(config().mcpRequestTimeoutSeconds) || DEFAULT_MCP_CONFIG.requestTimeoutSeconds, useProxy: serverData.useProxy }; settingsStore.updateConfig(SETTINGS_KEYS.MCP_SERVERS, JSON.stringify([...servers, newServer])); @@ -551,14 +556,6 @@ class MCPStore { }); } - /** - * MCP servers selectable in chat-add UIs and the settings page, - * in the order they were added to the config. - */ - get visibleMcpServers(): MCPServerSettingsEntry[] { - return this.getServers().filter((server) => server.enabled); - } - async ensureInitialized(perChatOverrides?: McpServerOverride[]): Promise { if (!browser) { return false; @@ -1226,7 +1223,6 @@ class MCPStore { id: string; enabled: boolean; url: string; - requestTimeoutSeconds: number; headers?: string; }[], skipIfChecked = true, @@ -1317,7 +1313,7 @@ class MCPStore { logs: [] }); - const timeoutMs = Math.round(server.requestTimeoutSeconds * 1000); + const timeoutMs = this.#requestTimeoutMs(); const headers = this.parseHeaders(server.headers); try { diff --git a/tools/ui/src/lib/stores/tools.svelte.ts b/tools/ui/src/lib/stores/tools.svelte.ts index dcaab5f42..53b6ad9ee 100644 --- a/tools/ui/src/lib/stores/tools.svelte.ts +++ b/tools/ui/src/lib/stores/tools.svelte.ts @@ -412,7 +412,8 @@ class ToolsStore { tools: { name: string; description?: string }[]; }[] { const result: ReturnType = []; - for (const server of mcpStore.visibleMcpServers) { + for (const server of mcpStore.getServers()) { + if (!server.enabled) continue; const health = mcpStore.getHealthCheckState(server.id); if (health.status === HealthCheckStatus.SUCCESS && health.tools.length > 0) { result.push({ diff --git a/tools/ui/src/lib/types/agentic.d.ts b/tools/ui/src/lib/types/agentic.d.ts index bcec10c67..0b5a8ef6c 100644 --- a/tools/ui/src/lib/types/agentic.d.ts +++ b/tools/ui/src/lib/types/agentic.d.ts @@ -15,7 +15,6 @@ import type { DatabaseMessage, DatabaseMessageExtra, McpServerOverride } from '. export interface AgenticConfig { enabled: boolean; maxTurns: number; - maxToolPreviewLines: number; } /** diff --git a/tools/ui/src/lib/types/mcp.d.ts b/tools/ui/src/lib/types/mcp.d.ts index dca57d396..2666394a0 100644 --- a/tools/ui/src/lib/types/mcp.d.ts +++ b/tools/ui/src/lib/types/mcp.d.ts @@ -174,7 +174,6 @@ export interface HealthCheckParams { id: string; enabled: boolean; url: string; - requestTimeoutSeconds: number; headers?: string; useProxy?: boolean; } @@ -220,7 +219,6 @@ export interface MCPServerDisplayInfo { export type MCPServerSettingsEntry = MCPServerDisplayInfo & { enabled: boolean; - requestTimeoutSeconds: number; headers?: string; iconUrl?: string; useProxy?: boolean; diff --git a/tools/ui/src/lib/utils/mcp.ts b/tools/ui/src/lib/utils/mcp.ts index 7ce7f19fe..b9a16ac92 100644 --- a/tools/ui/src/lib/utils/mcp.ts +++ b/tools/ui/src/lib/utils/mcp.ts @@ -9,7 +9,6 @@ import { MimeTypeText } from '$lib/enums'; import { - DEFAULT_MCP_CONFIG, MCP_SERVER_ID_PREFIX, IMAGE_FILE_EXTENSION_REGEX, CODE_FILE_EXTENSION_REGEX, @@ -64,7 +63,6 @@ export function detectMcpTransportFromUrl(url: string): MCPTransportType { /** * Parses MCP server settings from a JSON string or array. - * Preserves per-server requestTimeoutSeconds if stored, otherwise falls back to the global default. * @param rawServers - The raw servers to parse * @returns An empty array if the input is invalid. */ @@ -103,9 +101,6 @@ export function parseMcpServerSettings(rawServers: unknown): MCPServerSettingsEn enabled: Boolean((entry as { enabled?: unknown })?.enabled), url, name: (entry as { name?: string })?.name, - requestTimeoutSeconds: - (entry as { requestTimeoutSeconds?: number })?.requestTimeoutSeconds ?? - DEFAULT_MCP_CONFIG.requestTimeoutSeconds, headers: headers || undefined, useProxy: Boolean((entry as { useProxy?: unknown })?.useProxy) } satisfies MCPServerSettingsEntry; diff --git a/tools/ui/tests/unit/parse-mcp-server-settings.test.ts b/tools/ui/tests/unit/parse-mcp-server-settings.test.ts index f5e0b3a96..8dc7dffe3 100644 --- a/tools/ui/tests/unit/parse-mcp-server-settings.test.ts +++ b/tools/ui/tests/unit/parse-mcp-server-settings.test.ts @@ -1,6 +1,6 @@ import { describe, expect, it, vi } from 'vitest'; import { parseMcpServerSettings } from '$lib/utils/mcp'; -import { DEFAULT_MCP_CONFIG, MCP_SERVER_ID_PREFIX } from '$lib/constants/mcp'; +import { MCP_SERVER_ID_PREFIX } from '$lib/constants/mcp'; /** * Tests for the mcpServers settings parser. @@ -58,24 +58,16 @@ describe('parseMcpServerSettings', () => { expect(parsed[2]?.id).toBe('custom-3'); }); - it('falls back to the configured default requestTimeoutSeconds only for nullish values', () => { - const fallback = DEFAULT_MCP_CONFIG.requestTimeoutSeconds; - + it('does not emit a per-server timeout, the request timeout is a live global setting', () => { + // A stored per-server requestTimeoutSeconds was never editable in + // any UI and froze the global setting at server creation time, + // making the Settings value a no-op for existing servers. The + // parser drops the field so the global applies live everywhere. const parsed = parseMcpServerSettings( - JSON.stringify([ - { id: 'a', url: 'https://a.test' }, - { id: 'b', url: 'https://b.test', requestTimeoutSeconds: undefined }, - { id: 'c', url: 'https://c.test', requestTimeoutSeconds: 0 }, - { id: 'd', url: 'https://d.test', requestTimeoutSeconds: 45 } - ]) + JSON.stringify([{ id: 'a', url: 'https://a.test', requestTimeoutSeconds: 45 }]) ); - // The parser uses ?? for timeout fallback, which only triggers on - // null/undefined. Explicit 0 is preserved at face value. - expect(parsed[0]?.requestTimeoutSeconds).toBe(fallback); - expect(parsed[1]?.requestTimeoutSeconds).toBe(fallback); - expect(parsed[2]?.requestTimeoutSeconds).toBe(0); - expect(parsed[3]?.requestTimeoutSeconds).toBe(45); + expect(parsed[0]).not.toHaveProperty('requestTimeoutSeconds'); }); it('treats whitespace-only headers strings as undefined', () => { @@ -108,6 +100,22 @@ describe('parseMcpServerSettings', () => { expect(parsed[3]?.useProxy).toBe(true); }); + it('keeps disabled entries in the list, enabled is state and never a visibility filter', () => { + // Regression guard for issue #25625: filtering the server list on + // `enabled` hides a toggled-off server from every UI surface with + // no way to re-enable it. Any list derived from this parser must + // contain disabled entries. + const parsed = parseMcpServerSettings( + JSON.stringify([ + { id: 'on', url: 'https://on.test', enabled: true }, + { id: 'off', url: 'https://off.test', enabled: false } + ]) + ); + + expect(parsed.map((entry) => entry.id)).toEqual(['on', 'off']); + expect(parsed[1]?.enabled).toBe(false); + }); + it('preserves input order when mapping entries', () => { const source = [ { id: 'gamma', url: 'https://c.test' }, From 00e79f6fb146b934e7e62aa766a3f729f74b8b2e Mon Sep 17 00:00:00 2001 From: Hongqiang Wang Date: Tue, 14 Jul 2026 08:08:13 -0700 Subject: [PATCH 16/24] opencl: fix a dp4a bug for devices where cl_khr_integer_dot_product is unavailable (#25639) * opencl: do not fail backend init on devices without cl_khr_integer_dot_product * opencl: do not call dp4 kernels when dp is unavailable --------- Co-authored-by: Li He --- ggml/src/ggml-opencl/ggml-opencl.cpp | 126 +++++++++++++++------------ 1 file changed, 68 insertions(+), 58 deletions(-) diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp index f283f6569..7b6917592 100644 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp @@ -532,6 +532,7 @@ struct ggml_backend_opencl_context { bool fp16_support; bool has_vector_subgroup_broadcast; bool has_subgroup_shuffle = false; // cl_khr_subgroup_shuffle or cl_qcom_subgroup_shuffle + bool has_integer_dot = false; // cl_khr_integer_dot_product or cl_qcom_dot_product8 bool has_qcom_subgroup_shuffle = false; // specifically cl_qcom_subgroup_shuffle bool disable_fusion; @@ -834,7 +835,7 @@ struct ggml_backend_opencl_context { cl_kernel kernel_gemv_moe_q5_1_f32_ns, kernel_gemm_moe_q5_1_f32_ns; cl_kernel kernel_gemv_moe_q4_k_f32_ns, kernel_gemm_moe_q4_k_f32_ns, kernel_gemm_moe_q4_k_f32_ns_bin; cl_kernel kernel_gemv_moe_q4_k_f32_ns_wimg = nullptr; // weight-as-texture MoE decode GEMV (opt-in) - cl_kernel kernel_gemm_moe_q4_k_q8_1_dp4a; // dp4a (int8) prefill GEMM variant + cl_kernel kernel_gemm_moe_q4_k_q8_1_dp4a = nullptr; // dp4a (int8) prefill GEMM variant cl_kernel kernel_moe_reorder_quant_a_q8_1; // fused reorder + q8_1 quant for the dp4a GEMM cl_kernel kernel_gemm_moe_q8_1_dp4a_q80 = nullptr; // generic dp4a MoE GEMM (MOE_QT=80), opt-in cl_kernel kernel_moe_expand_scale_q8_0 = nullptr; // q8_0 per-block d -> uniform scale[16] @@ -844,12 +845,12 @@ struct ggml_backend_opencl_context { cl_kernel kernel_moe_expand_scale_q5_K = nullptr; // q5_K 6-bit s[] -> uniform scale[16]/min[8] cl_kernel kernel_gemv_moe_q5_k_f32_ns, kernel_gemm_moe_q5_k_f32_ns; cl_kernel kernel_gemv_moe_q6_k_f32_ns, kernel_gemm_moe_q6_k_f32_ns; - cl_kernel kernel_gemm_moe_q6_k_q8_1_dp4a; // dp4a (int8) q6_K MoE prefill GEMM + cl_kernel kernel_gemm_moe_q6_k_q8_1_dp4a = nullptr; // dp4a (int8) q6_K MoE prefill GEMM cl_kernel kernel_gemv_moe_mxfp4_f32, kernel_gemm_moe_mxfp4_f32; cl_kernel kernel_gemv_moe_mxfp4_f32_ns, kernel_gemm_moe_mxfp4_f32_ns, kernel_gemm_moe_mxfp4_f32_ns_bin; cl_kernel kernel_gemv_moe_mxfp4_f32_ns_wimg = nullptr; // weight-as-texture MoE decode GEMV - cl_kernel kernel_gemm_moe_mxfp4_q8_1_dp4a; // dp4a (int8) mxfp4 MoE prefill GEMM - cl_kernel kernel_gemm_moe_q4_0_q8_1_dp4a; // dp4a (int8) q4_0 MoE prefill GEMM + cl_kernel kernel_gemm_moe_mxfp4_q8_1_dp4a = nullptr; // dp4a (int8) mxfp4 MoE prefill GEMM + cl_kernel kernel_gemm_moe_q4_0_q8_1_dp4a = nullptr; // dp4a (int8) q4_0 MoE prefill GEMM cl_kernel kernel_moe_reorder_b; cl_kernel kernel_moe_histogram, kernel_moe_scan, kernel_moe_fill, kernel_moe_scatter; cl_kernel kernel_moe_combine_f32 = nullptr; // fused router-weight mul + cross-expert sum @@ -1037,10 +1038,10 @@ struct ggml_backend_opencl_context { cl_kernel kernel_gemv_noshuffle_q1_0_f32; cl_kernel kernel_gemv_noshuffle_q4_k_f32; cl_kernel kernel_gemm_noshuffle_q4_k_f32; - cl_kernel kernel_gemm_noshuffle_q4_k_q8_1_dp4a; // dp4a (int8) dense prefill GEMM - cl_kernel kernel_gemm_noshuffle_q4_k_q8_1_dp4a_wimg; // dp4a dense prefill GEMM, weights via texture (X1 opt-in) - cl_kernel kernel_gemm_noshuffle_q5_k_q8_1_dp4a; // dp4a (int8) dense q5_K prefill GEMM - cl_kernel kernel_gemm_noshuffle_q6_k_q8_1_dp4a; // dp4a (int8) dense q6_K prefill GEMM + cl_kernel kernel_gemm_noshuffle_q4_k_q8_1_dp4a = nullptr; // dp4a (int8) dense prefill GEMM + cl_kernel kernel_gemm_noshuffle_q4_k_q8_1_dp4a_wimg = nullptr; // dp4a dense prefill GEMM, weights via texture (X1 opt-in) + cl_kernel kernel_gemm_noshuffle_q5_k_q8_1_dp4a = nullptr; // dp4a (int8) dense q5_K prefill GEMM + cl_kernel kernel_gemm_noshuffle_q6_k_q8_1_dp4a = nullptr; // dp4a (int8) dense q6_K prefill GEMM cl_kernel kernel_quant_a_q8_1; // plain activation q8_1 pre-pass cl_kernel kernel_gemv_noshuffle_q6_K_f32; cl_kernel kernel_gemm_noshuffle_q6_K_f32; @@ -3490,7 +3491,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_noshuffle_q5_0_q8_1_dp4a (dp4a dense q5_0 prefill GEMM) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_noshuffle_q5_0_q8_1_dp4a.cl.h" @@ -3580,7 +3581,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_noshuffle_iq4_nl_q8_1_dp4a (dp4a dense IQ4_NL prefill GEMM) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_noshuffle_iq4_nl_q8_1_dp4a.cl.h" @@ -3595,7 +3596,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_noshuffle_q4_0_q8_1_dp4a (dp4a dense q4_0 prefill GEMM) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_noshuffle_q4_0_q8_1_dp4a.cl.h" @@ -3708,7 +3709,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_noshuffle_q4_k_q8_1_dp4a (dp4a dense prefill GEMM) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_noshuffle_q4_k_q8_1_dp4a.cl.h" @@ -3730,7 +3731,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_noshuffle_q8_0_q8_1_dp4a (dp4a dense q8_0 prefill GEMM) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_noshuffle_q8_0_q8_1_dp4a.cl.h" @@ -3746,7 +3747,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_noshuffle_q5_k_q8_1_dp4a (dp4a dense prefill GEMM for q5_K) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_noshuffle_q5_k_q8_1_dp4a.cl.h" @@ -3761,7 +3762,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_noshuffle_q6_k_q8_1_dp4a (dp4a dense prefill GEMM for q6_K ffn_down/output) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_noshuffle_q6_k_q8_1_dp4a.cl.h" @@ -4091,7 +4092,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_moe_q4_k_q8_1_dp4a (dp4a prefill GEMM) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_moe_q4_k_q8_1_dp4a.cl.h" @@ -4108,7 +4109,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_moe_mxfp4_q8_1_dp4a (dp4a prefill GEMM) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_moe_mxfp4_q8_1_dp4a.cl.h" @@ -4125,7 +4126,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_moe_q4_0_q8_1_dp4a (dp4a prefill GEMM) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_moe_q4_0_q8_1_dp4a.cl.h" @@ -4142,7 +4143,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_moe_q8_1_dp4a (generic dp4a MoE GEMM; MOE_QT=80 -> q8_0 expert variant) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_moe_q8_1_dp4a.cl.h" @@ -4256,7 +4257,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { } // gemm_moe_q6_k_q8_1_dp4a (dp4a q6_K MoE prefill GEMM) - { + if (backend_ctx->has_integer_dot) { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { #include "gemm_moe_q6_k_q8_1_dp4a.cl.h" @@ -5602,6 +5603,8 @@ static void ggml_opencl_print_backend_info(ggml_backend_opencl_device_context * backend_ctx->has_subgroup_shuffle ? "true" : "false"); GGML_LOG_INFO("ggml_opencl: device FP16 support: %s\n", backend_ctx->fp16_support ? "true" : "false"); + GGML_LOG_INFO("ggml_opencl: khr dot product support: %s\n", + backend_ctx->has_integer_dot ? "true" : "false"); GGML_LOG_INFO("ggml_opencl: mem base addr align: %u\n", backend_ctx->alignment); GGML_LOG_INFO("ggml_opencl: global mem size: %zu MB\n", @@ -5810,6 +5813,12 @@ static ggml_backend_opencl_context * ggml_cl_init(ggml_backend_dev_t dev) { strstr(ext_buffer, "cl_khr_subgroup_shuffle") != NULL || backend_ctx->has_qcom_subgroup_shuffle; + // check for cl_khr_integer_dot_product + // cl_qcom_dot_product8 uses signed * unsigned + // while cl_khr_integer_dot_product uses signed * signed -- we stick with khr for now + backend_ctx->has_integer_dot = + strstr(ext_buffer, "cl_khr_integer_dot_product") != NULL; + cl_uint base_align_in_bits; CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &base_align_in_bits, NULL)); GGML_ASSERT(base_align_in_bits % 8u == 0); @@ -15819,18 +15828,14 @@ static void ggml_cl_mul_mat_q4_0_f32_adreno(ggml_backend_t backend, const ggml_t CL_CHECK(clReleaseMemObject(b_sub_buf)); CL_CHECK(clReleaseMemObject(b_img)); } else { - // dp4a (int8) dense prefill GEMM: quant activations to q8_1, then the int8 - // dp4a inner-loop GEMM, in place of the transpose + f16 half-dot kernel. - // q4_0 = d*(q-8); mirrors the IQ4_NL/q8_0 dense dp4a paths (+ the sum term). - // OPT-IN / DEFAULT OFF: correct, but neutral on X2E. q4_0's dequant - // ((q-8)*scale) is already trivial so the f16 GEMM is weight-BW-bound and the - // int8 ALU win has nothing to beat -- same as q5_0 dense (unlike IQ4_NL, whose - // codebook dequant is expensive enough for dp4a to help). Kept for A/B; force - // on with GGML_OPENCL_Q4_0_DENSE_DP4A=1. Needs N>8, K%32==0, M%64==0. + // dp4a (int8) dense prefill GEMM, default off static const char * q4_0_dense_dp4a_env = getenv("GGML_OPENCL_Q4_0_DENSE_DP4A"); - const bool q4_0_dense_dp4a_on = q4_0_dense_dp4a_env + bool q4_0_dense_dp4a_on = q4_0_dense_dp4a_env ? (atoi(q4_0_dense_dp4a_env) != 0) : false; + // dot prod has to be available + q4_0_dense_dp4a_on = backend_ctx->has_integer_dot && q4_0_dense_dp4a_on; + if (q4_0_dense_dp4a_on && backend_ctx->kernel_gemm_noshuffle_q4_0_q8_1_dp4a && N > 8 && (K % 32 == 0) && (M % 64 == 0)) { cl_mem a_sub = nullptr; @@ -16253,27 +16258,16 @@ static void ggml_cl_mul_mat_q5_0_f32_adreno(ggml_backend_t backend, const ggml_t CL_CHECK(clReleaseMemObject(b_sub_buf)); CL_CHECK(clReleaseMemObject(b_img)); } else { - // dp4a (int8) dense q5_0 prefill GEMM. Quantizes the [N,K] activations to - // q8_1 and runs the int8 dot instead of the f16 half-dot. Large-batch - // (ne1>8) only. q5_0 weight = (x-16)*d (x = nibble | hi<<4); x packed as a - // 0..31 byte (dp4a), the -16 centering folded into a single min term - // (d*16) via the q8_1 block sum. Reads the qs/qh/d buffers byte-identically - // to the f16 kernel (greedy byte-identical, MUL_MAT NMSE-OK). - // - // OPT-IN / DEFAULT OFF. Unlike q8_0/q4_K dense, dp4a is not a win for q5_0 on - // X2E: the q5_0 model is bottlenecked elsewhere, so the dense-GEMM int8 win - // has nothing to surface and the q8_1 prepass slightly hurts. Kept correct + - // opt-in for the X1 A/B (different texture-cache dynamic) and the - // weight-texture variant. Env: GGML_OPENCL_Q5_DENSE_DP4A=1. - // Weight-as-texture variant (X1 lever): routes the dominant qs nibble plane - // through an image1d_buffer (qh stays a buffer). Opt-in - // GGML_OPENCL_Q5_DENSE_DP4A_WIMG; when set it also forces the dp4a path on. + // dp4a (int8) dense q5_0 prefill GEMM, default off static const char * q5_dense_dp4a_env = getenv("GGML_OPENCL_Q5_DENSE_DP4A"); static const char * q5_dense_wimg_env = getenv("GGML_OPENCL_Q5_DENSE_DP4A_WIMG"); const bool q5_dense_wimg_on = q5_dense_wimg_env && (atoi(q5_dense_wimg_env) != 0); - const bool q5_dense_dp4a_on = q5_dense_wimg_on + bool q5_dense_dp4a_on = q5_dense_wimg_on ? true : (q5_dense_dp4a_env && (atoi(q5_dense_dp4a_env) != 0)); + // dot prod has to be available + q5_dense_dp4a_on = backend_ctx->has_integer_dot && q5_dense_dp4a_on; + if (q5_dense_dp4a_on && backend_ctx->kernel_gemm_noshuffle_q5_0_q8_1_dp4a && N > 8 && (K % 32 == 0) && (M % 64 == 0)) { cl_mem a_sub = nullptr; @@ -16708,15 +16702,14 @@ static void ggml_cl_mul_mat_iq4_nl_f32_adreno(ggml_backend_t backend, const ggml } else { // dp4a (int8) dense IQ4_NL prefill GEMM. Quantizes the [N,K] activations to // q8_1 and runs the int8 dot instead of the f16 half-dot. Large-batch - // (ne1>8) only. IQ4_NL weight = kvalues[nibble]*d; the codebook value IS the - // int8 (no min term), so this is the q8_0 dense case plus a nibble->int8 LUT - // unpack. Reads the q/d buffers byte-identically to the f16 kernel. No bin - // kernel for IQ4_NL -> baseline is f16, default ON for X2E (like q4_K/q6_K - // dense dp4a). X1 stays on f16. Env: GGML_OPENCL_IQ4NL_DENSE_DP4A. + // (ne1>8) only static const char * iq4nl_dense_dp4a_env = getenv("GGML_OPENCL_IQ4NL_DENSE_DP4A"); - const bool iq4nl_dense_dp4a_on = iq4nl_dense_dp4a_env + bool iq4nl_dense_dp4a_on = iq4nl_dense_dp4a_env ? (atoi(iq4nl_dense_dp4a_env) != 0) : (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E); + // dot prod has to be available + iq4nl_dense_dp4a_on = backend_ctx->has_integer_dot && iq4nl_dense_dp4a_on; + if (iq4nl_dense_dp4a_on && backend_ctx->kernel_gemm_noshuffle_iq4_nl_q8_1_dp4a && N > 8 && (K % 32 == 0) && (M % 64 == 0)) { cl_mem a_sub = nullptr; @@ -16966,13 +16959,15 @@ static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_t static const char * q8_dense_wimg_env = getenv("GGML_OPENCL_Q8_DENSE_DP4A_WIMG"); const bool q8_dense_wimg_on = q8_dense_wimg_env && (atoi(q8_dense_wimg_env) != 0); - const bool q8_bin_loaded = (backend_ctx->kernel_gemm_noshuffle_q8_0_f32_bin != nullptr); + const bool q8_bin_loaded = (backend_ctx->kernel_gemm_noshuffle_q8_0_f32_bin != nullptr); // bin kernel takes precedence - const bool q8_dense_dp4a_on = q8_dense_wimg_on + bool q8_dense_dp4a_on = q8_dense_wimg_on ? true : q8_dense_dp4a_env ? (atoi(q8_dense_dp4a_env) != 0) : (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E && !q8_bin_loaded); + // dot prod has to be available + q8_dense_dp4a_on = backend_ctx->has_integer_dot && q8_dense_dp4a_on; if (q8_dense_dp4a_on && backend_ctx->kernel_gemm_noshuffle_q8_0_q8_1_dp4a && N > 8 && (K % 32 == 0) && (M % 64 == 0)) { @@ -17379,13 +17374,16 @@ static void ggml_cl_mul_mat_q4_k_f32_adreno(ggml_backend_t backend, const ggml_t static const char * q4k_dense_dp4a_env = getenv("GGML_OPENCL_Q4K_DENSE_DP4A"); static const char * q4k_dense_wimg_env = getenv("GGML_OPENCL_Q4K_DENSE_DP4A_WIMG"); - const bool q4k_dense_wimg_on = q4k_dense_wimg_env && (atoi(q4k_dense_wimg_env) != 0); - const bool q4k_dense_dp4a_on = q4k_dense_wimg_on + const bool q4k_dense_wimg_on = q4k_dense_wimg_env && (atoi(q4k_dense_wimg_env) != 0); + bool q4k_dense_dp4a_on = q4k_dense_wimg_on ? true : q4k_dense_dp4a_env ? (atoi(q4k_dense_dp4a_env) != 0) : (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E); + // dp4 has to be available + q4k_dense_dp4a_on = backend_ctx->has_integer_dot && q4k_dense_dp4a_on; + // Min N for the dp4a prefill GEMM, default 9, i.e., ne1 > 8 static const char * q4k_dp4a_minn_env = getenv("GGML_OPENCL_Q4K_DP4A_MINN"); const int q4k_dp4a_minn = q4k_dp4a_minn_env ? atoi(q4k_dp4a_minn_env) : 9; @@ -17608,9 +17606,11 @@ static void ggml_cl_mul_mat_q6_K_f32_adreno(ggml_backend_t backend, const ggml_t // dp4a (int8) dense q6_K prefill GEMM static const char * q6k_dense_dp4a_env = getenv("GGML_OPENCL_Q6K_DENSE_DP4A"); - static const bool q6k_dense_dp4a_on = (q6k_dense_dp4a_env != nullptr) + bool q6k_dense_dp4a_on = (q6k_dense_dp4a_env != nullptr) ? (atoi(q6k_dense_dp4a_env) != 0) : (backend_ctx->adreno_gen != ADRENO_GPU_GEN::X1E); + // dot prod has to be available + q6k_dense_dp4a_on = backend_ctx->has_integer_dot && q6k_dense_dp4a_on; const bool is_output_w_dp4a = strncmp(src0->name, "output", 6) == 0 || strncmp(src0->name, "token_embd", 10) == 0; @@ -17901,9 +17901,11 @@ static void ggml_cl_mul_mat_q5_K_f32_adreno(ggml_backend_t backend, const ggml_t // dp4a (int8) dense q5_K prefill GEMM static const char * q5k_dense_dp4a_env = getenv("GGML_OPENCL_Q5K_DENSE_DP4A"); - const bool q5k_dense_dp4a_on = q5k_dense_dp4a_env + bool q5k_dense_dp4a_on = q5k_dense_dp4a_env ? (atoi(q5k_dense_dp4a_env) != 0) : (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E); + // dot prod has to be available + q5k_dense_dp4a_on = backend_ctx->has_integer_dot && q5k_dense_dp4a_on; if (q5k_dense_dp4a_on && ne1 > 8 && (ne00 % 32 == 0) && (ne01 % 64 == 0)) { const int Mm = ne01, Nn = ne1, Kk = ne00; @@ -20640,6 +20642,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, bool use_moe_dp4a = q4_0_moe_dp4a_env ? (atoi(q4_0_moe_dp4a_env) != 0) : (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E); + // dot prod has to be available + use_moe_dp4a = backend_ctx->has_integer_dot && use_moe_dp4a; // bin kernel takes precedence use_moe_dp4a = use_moe_dp4a && backend_ctx->kernel_gemm_moe_q4_0_f32_ns_bin == nullptr; @@ -21815,6 +21819,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, bool use_moe_dp4a = (q4k_moe_dp4a_env != nullptr) ? (atoi(q4k_moe_dp4a_env) != 0) : (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E || backend_ctx->adreno_gen == ADRENO_GPU_GEN::X1E); + // dot prod has to be available + use_moe_dp4a = backend_ctx->has_integer_dot && use_moe_dp4a; // bin kernel takes precedence use_moe_dp4a = use_moe_dp4a && backend_ctx->kernel_gemm_moe_q4_k_f32_ns_bin == nullptr; @@ -22316,10 +22322,12 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, // dp4a (int8) q6_K MoE prefill GEMM static const char * q6k_moe_dp4a_env = getenv("GGML_OPENCL_Q6K_MOE_DP4A"); - static const bool use_moe_dp4a = (q6k_moe_dp4a_env != nullptr) + bool use_moe_dp4a = (q6k_moe_dp4a_env != nullptr) ? (atoi(q6k_moe_dp4a_env) != 0) : (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E || backend_ctx->adreno_gen == ADRENO_GPU_GEN::X1E); + // dot prod has to be available + use_moe_dp4a = backend_ctx->has_integer_dot && use_moe_dp4a; cl_buffer_region region; region.origin = 0; @@ -22569,6 +22577,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, bool use_moe_dp4a = mxfp4_moe_dp4a_env ? (atoi(mxfp4_moe_dp4a_env) != 0) : (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E); + // dot prod has to be available + use_moe_dp4a = backend_ctx->has_integer_dot && use_moe_dp4a; // bin kernel takes precedence use_moe_dp4a = use_moe_dp4a && backend_ctx->kernel_gemm_moe_mxfp4_f32_ns_bin == nullptr; From dfba90db6392a3f48ea86a09c4ce4fcfcdabd07a Mon Sep 17 00:00:00 2001 From: Emanuil Rusev Date: Tue, 14 Jul 2026 18:12:22 +0300 Subject: [PATCH 17/24] webui: parse effective-parameter sizes (E2B, E4B) as params (#25529) --- tools/ui/src/lib/constants/model-id.ts | 4 +++- tools/ui/tests/unit/model-id-parser.test.ts | 5 +++++ 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/tools/ui/src/lib/constants/model-id.ts b/tools/ui/src/lib/constants/model-id.ts index ab7932240..4108a2132 100644 --- a/tools/ui/src/lib/constants/model-id.ts +++ b/tools/ui/src/lib/constants/model-id.ts @@ -24,8 +24,10 @@ export const MODEL_CUSTOM_QUANTIZATION_PREFIX_RE = /^UD$/i; /** * Matches a parameter-count segment, e.g. `7B`, `1.5b`, `120M`. + * The optional leading `E` covers effective-parameter sizes, e.g. Gemma's + * `E2B`/`E4B` (MatFormer models sized by resident params). */ -export const MODEL_PARAMS_RE = /^\d+(\.\d+)?[BbMmKkTt]$/; +export const MODEL_PARAMS_RE = /^[Ee]?\d+(\.\d+)?[BbMmKkTt]$/; /** * Matches an activated-parameter-count segment, e.g. `A10B`, `a2.4b`. diff --git a/tools/ui/tests/unit/model-id-parser.test.ts b/tools/ui/tests/unit/model-id-parser.test.ts index 3c2937d35..3439d956e 100644 --- a/tools/ui/tests/unit/model-id-parser.test.ts +++ b/tools/ui/tests/unit/model-id-parser.test.ts @@ -36,6 +36,11 @@ describe('parseModelId', () => { expect(parseModelId('model-100b:q4_k_m')).toMatchObject({ params: '100B' }); }); + it('extracts effective parameters correctly', () => { + expect(parseModelId('model-E4B-BF16')).toMatchObject({ params: 'E4B' }); + expect(parseModelId('model-e2b:q4_k_m')).toMatchObject({ params: 'E2B' }); + }); + it('extracts activated parameters correctly', () => { expect(parseModelId('model-100B-A10B-BF16')).toMatchObject({ activatedParams: 'A10B' }); expect(parseModelId('model-100B-A10B:Q4_K_M')).toMatchObject({ activatedParams: 'A10B' }); From 236ab574e05ddfbb663e2bb8eaf1d89821478062 Mon Sep 17 00:00:00 2001 From: Bill Sideris Date: Tue, 14 Jul 2026 18:23:11 +0300 Subject: [PATCH 18/24] ui: Fix spacing in tool-call request (#25634) --- .../ChatMessageActionCardPermissionRequest.svelte | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageActions/ChatMessageActionCard/ChatMessageActionCardPermissionRequest.svelte b/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageActions/ChatMessageActionCard/ChatMessageActionCardPermissionRequest.svelte index 4337bb6a1..7f25c4549 100644 --- a/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageActions/ChatMessageActionCard/ChatMessageActionCardPermissionRequest.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageActions/ChatMessageActionCard/ChatMessageActionCardPermissionRequest.svelte @@ -21,7 +21,7 @@ {#snippet message()} Allow use of {toolName}{#if serverLabel} - from {serverLabel}{/if}? +  from {serverLabel}{/if}? {/snippet} {#snippet actions()} From 6e52db5b72b643cad908458b918ddb4f4b7974d9 Mon Sep 17 00:00:00 2001 From: Xuan-Son Nguyen Date: Tue, 14 Jul 2026 17:23:44 +0200 Subject: [PATCH 19/24] server: add --cors-* options (#25655) * server: add --cors-* options * add special "localhost" value * add tests * fix test * add link to PR --- common/arg.cpp | 50 ++++++++++++++++-- common/common.h | 8 +++ tools/server/server-http.cpp | 33 ++++++++++-- tools/server/server.cpp | 36 +++++++++---- tools/server/tests/unit/test_security.py | 66 +++++++++++++++++++++++- tools/server/tests/utils.py | 5 +- 6 files changed, 177 insertions(+), 21 deletions(-) diff --git a/common/arg.cpp b/common/arg.cpp index 9676adafe..b6fddae00 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -697,7 +697,7 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context } }; - // parse the first time to get -hf option (used for remote preset) + // parse all CLI args now, so that -hf is available below for remote preset resolution parse_cli_args(); postprocess_cpu_params(params.cpuparams, nullptr); @@ -748,6 +748,11 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context params.kv_overrides.back().key[0] = 0; } + if (!params.server_tools.empty() && !params.cors_origins_explicit) { + LOG_WRN("server tools are enabled, using localhost as default CORS origin (change via --cors-origins)\n"); + params.cors_origins = "localhost"; + } + // pad tensor_buft_overrides for llama_params_fit: const size_t ntbo = llama_max_tensor_buft_overrides(); while (params.tensor_buft_overrides.size() < ntbo) { @@ -3047,6 +3052,42 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.public_path = value; } ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH")); + add_opt(common_arg( + {"--cors-origins"}, "ORIGINS", + string_format( + "comma-separated list of allowed origins for CORS (default: %s)\n" + "if set to special value 'localhost', reflect the Origin header only if it is localhost", + params.cors_origins.c_str()), + [](common_params & params, const std::string & value) { + params.cors_origins = value; + params.cors_origins_explicit = true; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CORS_ORIGINS")); + add_opt(common_arg( + {"--cors-methods"}, "METHODS", + string_format("comma-separated list of allowed methods for CORS (default: %s)", params.cors_methods.c_str()), + [](common_params & params, const std::string & value) { + params.cors_methods = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CORS_METHODS")); + add_opt(common_arg( + {"--cors-headers"}, "HEADERS", + string_format("comma-separated list of allowed headers for CORS (default: %s)", params.cors_headers.c_str()), + [](common_params & params, const std::string & value) { + params.cors_headers = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CORS_HEADERS")); + add_opt(common_arg( + {"--cors-credentials"}, + {"--no-cors-credentials"}, + string_format( + "whether to allow credentials for CORS (default: %s)\n" + "note: if this is enabled and --cors-origins is set to * (default), the Origin header will be echoed back, and credentials will always be allowed", + params.cors_credentials ? "enabled" : "disabled"), + [](common_params & params, bool value) { + params.cors_credentials = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CORS_CREDENTIALS")); add_opt(common_arg( {"--api-prefix"}, "PREFIX", string_format("prefix path the server serves from, without the trailing slash (default: %s)", params.api_prefix.c_str()), @@ -3080,7 +3121,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"--tools"}, "TOOL1,TOOL2,...", "experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)\n" "specify \"all\" to enable all tools\n" - "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime", + "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime\n" + "note: for security reasons, this will limit --cors-origins to localhost by default", [](common_params & params, const std::string & value) { params.server_tools = parse_csv_row(value); } @@ -3088,7 +3130,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex add_opt(common_arg( {"-ag", "--agent"}, {"-no-ag", "--no-agent"}, - "whether to enable CORS proxy and all built-in tools - do not enable in untrusted environments (default: disabled)", + "whether to enable CORS proxy and all built-in tools - do not enable in untrusted environments (default: disabled)\n" + "note: for security reasons, this will limit --cors-origins to localhost by default", [](common_params & params, bool value) { if (value) { params.server_tools = {"all"}; @@ -3097,6 +3140,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.server_tools.clear(); params.ui_mcp_proxy = false; } + // note: do not modify cors_origins here, as the options are not evaluated in order (user may explicitly set --cors-origins before --agent) } ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_AGENT")); add_opt(common_arg( diff --git a/common/common.h b/common/common.h index 66760005c..bffc1767a 100644 --- a/common/common.h +++ b/common/common.h @@ -631,6 +631,14 @@ struct common_params { std::string api_prefix = ""; // NOLINT std::string chat_template = ""; // NOLINT bool use_jinja = true; // NOLINT + + // server CORS params + std::string cors_origins = "*"; + std::string cors_methods = "GET, POST, DELETE, OPTIONS"; + std::string cors_headers = "*"; + bool cors_credentials = true; + bool cors_origins_explicit = false; // for --agent option + bool enable_chat_template = true; bool force_pure_content_parser = false; common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; diff --git a/tools/server/server-http.cpp b/tools/server/server-http.cpp index 48f903dfc..24a38452a 100644 --- a/tools/server/server-http.cpp +++ b/tools/server/server-http.cpp @@ -47,6 +47,16 @@ static void log_server_request(const httplib::Request & req, const httplib::Resp SRV_DBG("response: %s\n", res.body.c_str()); } +// returns true if the Origin header value's host is localhost / 127.0.0.1 / ::1 (any port) +static bool origin_is_localhost(const std::string & origin) { + try { + const std::string host = common_http_parse_url(origin).host; + return host == "localhost" || host == "127.0.0.1" || host == "::1"; + } catch (const std::exception &) { + return false; + } +} + // For Google Cloud Platform deployment compatibility struct gcp_params { bool enabled; @@ -266,13 +276,26 @@ bool server_http_context::init(const common_params & params) { }; // register server middlewares - srv->set_pre_routing_handler([middleware_validate_api_key, middleware_server_state](const httplib::Request & req, httplib::Response & res) { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + srv->set_pre_routing_handler([¶ms, middleware_validate_api_key, middleware_server_state](const httplib::Request & req, httplib::Response & res) { + if (params.cors_credentials && params.cors_origins == "*") { + // special case: echo back the Origin header to allow any origin to access the server with credentials + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + } else if (params.cors_origins == "localhost") { + // special case: only reflect the Origin header if it is a localhost origin + std::string origin = req.get_header_value("Origin"); + if (origin_is_localhost(origin)) { + res.set_header("Access-Control-Allow-Origin", origin); + } else { + SRV_WRN("(CORS) skip non-localhost origin: %s\n", origin.c_str()); + } + } else { + res.set_header("Access-Control-Allow-Origin", params.cors_origins); + } // If this is OPTIONS request, skip validation because browsers don't include Authorization header if (req.method == "OPTIONS") { - res.set_header("Access-Control-Allow-Credentials", "true"); - res.set_header("Access-Control-Allow-Methods", "GET, POST"); - res.set_header("Access-Control-Allow-Headers", "*"); + res.set_header("Access-Control-Allow-Credentials", params.cors_credentials ? "true" : "false"); + res.set_header("Access-Control-Allow-Methods", params.cors_methods); + res.set_header("Access-Control-Allow-Headers", params.cors_headers); res.set_content("", "text/html"); // blank response, no data return httplib::Server::HandlerResponse::Handled; // skip further processing } diff --git a/tools/server/server.cpp b/tools/server/server.cpp index c2b21120a..20effbb14 100644 --- a/tools/server/server.cpp +++ b/tools/server/server.cpp @@ -303,14 +303,24 @@ int llama_server(common_params & params, int argc, char ** argv) { return res; }; + if (params.cors_origins == "*" && params.api_keys.empty()) { + SRV_WRN("%s", "-----------------\n"); + SRV_WRN("%s", "CORS is set to allow all origins ('*') and no API key is set\n"); + SRV_WRN("%s", "this can be a security risk (cross-origin attacks)\n"); + SRV_WRN("%s", "more info: https://github.com/ggml-org/llama.cpp/pull/25655\n"); + SRV_WRN("%s", "-----------------\n"); + } + // CORS proxy (EXPERIMENTAL, only used by the Web UI for MCP) + std::vector warn_names; + if (is_router_server) { + warn_names.push_back("router mode"); + } + if (params.ui_mcp_proxy) { - SRV_WRN("%s", "-----------------\n"); - SRV_WRN("%s", "CORS proxy is enabled, do not expose server to untrusted environments\n"); - SRV_WRN("%s", "This feature is EXPERIMENTAL and may be removed or changed in future versions\n"); - SRV_WRN("%s", "-----------------\n"); ctx_http.get ("/cors-proxy", ex_wrapper(proxy_handler_get)); ctx_http.post("/cors-proxy", ex_wrapper(proxy_handler_post)); + warn_names.push_back("MCP proxy (experimental)"); } else { ctx_http.get ("/cors-proxy", ex_wrapper(res_403)); ctx_http.post("/cors-proxy", ex_wrapper(res_403)); @@ -324,17 +334,24 @@ int llama_server(common_params & params, int argc, char ** argv) { SRV_ERR("tools setup failed: %s\n", e.what()); return 1; } - SRV_WRN("%s", "-----------------\n"); - SRV_WRN("%s", "Built-in tools are enabled, do not expose server to untrusted environments\n"); - SRV_WRN("%s", "This feature is EXPERIMENTAL and may be changed in the future\n"); - SRV_WRN("%s", "-----------------\n"); ctx_http.get ("/tools", ex_wrapper(tools.handle_get)); ctx_http.post("/tools", ex_wrapper(tools.handle_post)); + warn_names.push_back("built-in tools (experimental)"); } else { ctx_http.get ("/tools", ex_wrapper(res_403)); ctx_http.post("/tools", ex_wrapper(res_403)); } + if (warn_names.size() > 0) { + SRV_WRN("%s", "-----------------\n"); + SRV_WRN("%s", "the following feature(s) are enabled:\n"); + for (const auto & name : warn_names) { + SRV_WRN(" %s\n", name.c_str()); + } + SRV_WRN("%s", "do not expose the server to untrusted environments\n"); + SRV_WRN("%s", "-----------------\n"); + } + // // Handle downloading model // @@ -452,9 +469,6 @@ int llama_server(common_params & params, int argc, char ** argv) { SRV_INF("listening on %s\n", ctx_http.listening_address.c_str()); if (is_router_server) { - SRV_WRN("%s", "NOTE: router mode is experimental\n"); - SRV_WRN("%s", " it is not recommended to use this mode in untrusted environments\n"); - if (!params.models_preset_hf.empty()) { SRV_WRN( "NOTE: using preset.ini from HF repo '%s'\n", params.models_preset_hf.c_str()); SRV_WRN("%s", " please only use presets that you can trust! Unknown presets may be unsafe\n"); diff --git a/tools/server/tests/unit/test_security.py b/tools/server/tests/unit/test_security.py index a0c3e214a..ac0544575 100644 --- a/tools/server/tests/unit/test_security.py +++ b/tools/server/tests/unit/test_security.py @@ -91,7 +91,7 @@ def test_openai_library_correct_api_key(): ("localhost", "Access-Control-Allow-Origin", "localhost"), ("web.mydomain.fr", "Access-Control-Allow-Origin", "web.mydomain.fr"), ("origin", "Access-Control-Allow-Credentials", "true"), - ("web.mydomain.fr", "Access-Control-Allow-Methods", "GET, POST"), + ("web.mydomain.fr", "Access-Control-Allow-Methods", "GET, POST, DELETE, OPTIONS"), ("web.mydomain.fr", "Access-Control-Allow-Headers", "*"), ]) def test_cors_options(origin: str, cors_header: str, cors_header_value: str): @@ -107,6 +107,70 @@ def test_cors_options(origin: str, cors_header: str, cors_header_value: str): assert res.headers[cors_header] == cors_header_value +@pytest.mark.parametrize("origin", [ + "http://localhost", + "http://localhost:8080", + "http://127.0.0.1", + "http://127.0.0.1:3000", + "http://[::1]", + "http://[::1]:3000", +]) +def test_cors_origins_localhost_reflects(origin: str): + global server + server = ServerPreset.router() + server.cors_origins = "localhost" + server.start() + res = server.make_request("OPTIONS", "/completions", headers={ + "Origin": origin, + "Access-Control-Request-Method": "POST", + "Access-Control-Request-Headers": "Authorization", + }) + assert res.status_code == 200 + assert res.headers["Access-Control-Allow-Origin"] == origin + + +@pytest.mark.parametrize("origin", [ + "http://web.mydomain.fr", + "http://evil.com", + "http://notlocalhost", + "http://localhost.evil.com", +]) +def test_cors_origins_localhost_rejects(origin: str): + global server + server = ServerPreset.router() + server.cors_origins = "localhost" + server.start() + res = server.make_request("OPTIONS", "/completions", headers={ + "Origin": origin, + "Access-Control-Request-Method": "POST", + "Access-Control-Request-Headers": "Authorization", + }) + assert res.status_code == 200 + assert "Access-Control-Allow-Origin" not in res.headers + + +def test_cors_origins_defaults_to_localhost_with_tools_enabled(): + global server + server = ServerPreset.router() + server.server_tools = "all" + server.start() + res = server.make_request("OPTIONS", "/completions", headers={ + "Origin": "http://localhost:8080", + "Access-Control-Request-Method": "POST", + "Access-Control-Request-Headers": "Authorization", + }) + assert res.status_code == 200 + assert res.headers["Access-Control-Allow-Origin"] == "http://localhost:8080" + + res = server.make_request("OPTIONS", "/completions", headers={ + "Origin": "http://evil.com", + "Access-Control-Request-Method": "POST", + "Access-Control-Request-Headers": "Authorization", + }) + assert res.status_code == 200 + assert "Access-Control-Allow-Origin" not in res.headers + + def test_cors_proxy_only_forwards_explicit_proxy_headers(): class CaptureHeadersHandler(BaseHTTPRequestHandler): def do_GET(self): diff --git a/tools/server/tests/utils.py b/tools/server/tests/utils.py index 8c0de384f..f4f0e61e6 100644 --- a/tools/server/tests/utils.py +++ b/tools/server/tests/utils.py @@ -114,6 +114,7 @@ class ServerProcess: backend_sampling: bool = False gcp_compat: bool = False server_tools: str | None = None + cors_origins: str | None = None # session variables process: subprocess.Popen | None = None @@ -170,6 +171,8 @@ class ServerProcess: server_args.extend(["--models-max", self.models_max]) if self.models_preset: server_args.extend(["--models-preset", self.models_preset]) + if self.cors_origins: + server_args.extend(["--cors-origins", self.cors_origins]) if self.n_batch: server_args.extend(["--batch-size", self.n_batch]) if self.n_ubatch: @@ -359,7 +362,7 @@ class ServerProcess: if parse_body: try: result.body = response.json() - except JSONDecodeError: + except (JSONDecodeError, requests.exceptions.JSONDecodeError): result.body = response.text else: result.body = None From bf2c86ddc0685f580595954056c2e77ebabfab4f Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 14 Jul 2026 18:25:52 +0300 Subject: [PATCH 20/24] server : refactor prompt cache state ownership (#25649) * server : clear checkpoints upon prompt clear * server : move the prompt state data to the server_prompt_cache Assisted-by: pi:llama.cpp/Qwen3.6-27B * server : handle batched slot being cleared --- tools/server/server-context.cpp | 71 +++++++++++++++------------------ tools/server/server-task.cpp | 26 ++++++------ tools/server/server-task.h | 53 +++++++++++++----------- 3 files changed, 76 insertions(+), 74 deletions(-) diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp index 5fadc1c8d..7564ad4e9 100644 --- a/tools/server/server-context.cpp +++ b/tools/server/server-context.cpp @@ -220,8 +220,6 @@ struct server_slot { return false; } - GGML_ASSERT(prompt.data.size() == 0); - const size_t cur_size_tgt = llama_state_seq_get_size_ext(ctx_tgt, id, LLAMA_STATE_SEQ_FLAGS_NONE); const size_t cur_size_dft = ctx_dft ? llama_state_seq_get_size_ext(ctx_dft, id, LLAMA_STATE_SEQ_FLAGS_NONE) : 0; @@ -252,11 +250,7 @@ struct server_slot { return res; } - void prompt_clear(bool allow_processing) { - if (!allow_processing) { - GGML_ASSERT(!is_processing()); - } - + void prompt_clear() { SLT_TRC(*this, "clearing prompt with %zu tokens\n", prompt.tokens.size()); common_context_seq_rm(ctx_tgt, id, -1, -1); @@ -264,7 +258,7 @@ struct server_slot { common_context_seq_rm(ctx_dft, id, -1, -1); } - prompt.tokens.clear(); + prompt.clear(); } std::vector lora; @@ -493,7 +487,7 @@ struct server_slot { // do not keep context of the child slots - the parent's context is enough if (task->is_child()) { - prompt_clear(false); + prompt_clear(); } reset(); @@ -1626,7 +1620,7 @@ private: ret->prompt_save(*prompt_cache); if (!ret->prompt_load(*prompt_cache, task.tokens)) { - ret->prompt_clear(false); + ret->prompt_clear(); } prompt_cache->update(); @@ -1658,7 +1652,7 @@ private: if (slot.prompt.n_tokens() > 0) { SRV_WRN("purging slot %d with %zu tokens\n", slot.id, slot.prompt.tokens.size()); - slot.prompt_clear(false); + slot.prompt_clear(); res = true; @@ -1691,7 +1685,7 @@ private: // if lora has changed, check to see if the cache should be cleared if (lora_should_clear_cache(slot.lora, task_loras)) { SLT_TRC(slot, "clearing cache for lora change. %zu loras -> %zu loras\n", slot.lora.size(), task.params.lora.size()); - slot.prompt.tokens.clear(); + slot.prompt.clear(); } else { SLT_TRC(slot, "keeping cache for alora. %zu target loras\n", task_loras.size()); } @@ -2405,7 +2399,7 @@ private: if (params_base.kv_unified) { // [TAG_IDLE_SLOT_CLEAR] - slot.prompt_clear(false); + slot.prompt_clear(); } } } @@ -2573,12 +2567,12 @@ private: size_t token_count = 0; size_t nread = llama_state_seq_load_file(ctx_tgt, filepath.c_str(), slot->id, tokens.data(), tokens.size(), &token_count); if (nread == 0) { - slot->prompt.tokens.clear(); // KV may already been invalidated? + slot->prompt.clear(); // KV may already been invalidated? send_error(task, "Unable to restore slot, no available space in KV cache or invalid slot save file", ERROR_TYPE_INVALID_REQUEST); break; } tokens.resize(token_count); - slot->prompt.tokens.clear(); + slot->prompt.clear(); slot->prompt.tokens.insert(tokens); const int64_t t_end = ggml_time_us(); @@ -2615,7 +2609,7 @@ private: // Erase token cache const size_t n_erased = slot->prompt.tokens.size(); - slot->prompt_clear(false); + slot->prompt_clear(); auto res = std::make_unique(); res->id = task.id; @@ -2775,6 +2769,27 @@ private: abort_all_slots("pre_decode() failed: " + std::string(e.what())); } + GGML_ASSERT(batch.slot_batched || batch.size() == 0); + + if (batch.slot_batched) { + auto & slot_batched = batch.slot_batched; + auto & alora_scale = batch.alora_scale; + auto & alora_disabled_id = batch.alora_disabled_id; + + // TODO @ngxson : alora handling is too messy, need to refactor it to be more clear and maintainable + // apply lora, only need to do it once per batch + common_set_adapter_lora(ctx_tgt, slot_batched->lora); + + // if the lora is temporarily disabled for an alora, re-enable it + // for next time + if (alora_scale > 0.0f) { + SRV_DBG("re-enabling alora with scale %f\n", alora_scale); + slot_batched->lora[alora_disabled_id].scale = alora_scale; + } + + llama_set_embeddings(ctx_tgt, slot_batched->need_embd()); + } + llama_batch batch_view; int32_t off_next = 0; int32_t n_batch = llama_n_batch(ctx_tgt); @@ -2814,7 +2829,6 @@ private: abort_all_slots("post_decode() failed: " + std::string(e.what())); break; // stop any further processing } - } } @@ -2880,7 +2894,7 @@ private: new_tokens.resize(slot.prompt.tokens.size() - n_discard); - slot.prompt.tokens.clear(); + slot.prompt.clear(); slot.prompt.tokens.insert(new_tokens); } @@ -3556,25 +3570,6 @@ private: bool decode(int32_t & n_batch, int32_t off, llama_batch & batch_view) { SRV_DBG("n_batch (effective) = %d, off = %d\n", n_batch, off); - auto & slot_batched = batch.slot_batched; - auto & alora_scale = batch.alora_scale; - auto & alora_disabled_id = batch.alora_disabled_id; - - // TODO @ngxson : alora handling is too messy, need to refactor it to be more clear and maintainable - if (slot_batched) { - // apply lora, only need to do it once per batch - common_set_adapter_lora(ctx_tgt, slot_batched->lora); - - // if the lora is temporarily disabled for an alora, re-enable it - // for next time - if (alora_scale > 0.0f) { - SRV_DBG("re-enabling alora with scale %f\n", alora_scale); - slot_batched->lora[alora_disabled_id].scale = alora_scale; - } - - llama_set_embeddings(ctx_tgt, slot_batched->need_embd()); - } - if (batch.size() == 0) { SRV_WRN("%s", "no tokens to decode\n"); @@ -3622,7 +3617,7 @@ private: // note: it's complicated to keep track of how much of the current batch has been // processed before the error occurred, so we simply clear the entire context - slot.prompt_clear(false); + slot.prompt_clear(); } } diff --git a/tools/server/server-task.cpp b/tools/server/server-task.cpp index 8d611e520..f01780100 100644 --- a/tools/server/server-task.cpp +++ b/tools/server/server-task.cpp @@ -1646,16 +1646,16 @@ size_t server_prompt_cache::n_tokens() const { size_t res = 0; for (const auto & state : states) { - res += state.n_tokens(); + res += state.prompt.n_tokens(); } return res; } -server_prompt * server_prompt_cache::alloc(const server_prompt & prompt, size_t state_size_tgt, size_t state_size_dft) { +server_prompt_cache_state * server_prompt_cache::alloc(const server_prompt & prompt, size_t state_size_tgt, size_t state_size_dft) { // first check if the current state is contained fully in the cache for (auto it = states.begin(); it != states.end(); ++it) { - const int cur_lcp_len = it->tokens.get_common_prefix(prompt.tokens); + const int cur_lcp_len = it->prompt.tokens.get_common_prefix(prompt.tokens); if (cur_lcp_len == (int) prompt.tokens.size()) { SRV_TRC("%s", " - prompt is already in the cache, skipping\n"); @@ -1680,9 +1680,9 @@ server_prompt * server_prompt_cache::alloc(const server_prompt & prompt, size_t // remove any cached prompts that are fully contained in the current prompt for (auto it = states.begin(); it != states.end();) { - const int len = it->tokens.get_common_prefix(prompt.tokens); + const int len = it->prompt.tokens.get_common_prefix(prompt.tokens); - if (len == (int) it->tokens.size()) { + if (len == (int) it->prompt.tokens.size()) { SRV_TRC(" - removing obsolete cached prompt with length %d\n", len); it = states.erase(it); @@ -1721,12 +1721,14 @@ server_prompt * server_prompt_cache::alloc(const server_prompt & prompt, size_t } states.push_back({ - /*.tokens =*/ prompt.tokens.clone(), - /*.data =*/ { + /*.prompt =*/ { + /*.tokens =*/ prompt.tokens.clone(), + /*.checkpoints =*/ prompt.checkpoints, + }, + /*.data =*/ { /*.main =*/ std::move(state_data_tgt), /*.drft =*/ std::move(state_data_dft), }, - /*.checkpoints =*/ prompt.checkpoints, }); return &states.back(); @@ -1744,9 +1746,9 @@ bool server_prompt_cache::load(server_prompt & prompt, const server_tokens & tok // find the most similar cached prompt, that would also preserve the most context for (auto it = states.begin(); it != states.end(); ++it) { - const int lcp_cur = it->tokens.get_common_prefix(tokens_new); + const int lcp_cur = it->prompt.tokens.get_common_prefix(tokens_new); - const float f_keep_cur = float(lcp_cur) / it->tokens.size(); + const float f_keep_cur = float(lcp_cur) / it->prompt.tokens.size(); const float sim_cur = float(lcp_cur) / tokens_new.size(); // don't trash large prompts @@ -1799,7 +1801,7 @@ bool server_prompt_cache::load(server_prompt & prompt, const server_tokens & tok } } - prompt = std::move(*it_best); + prompt = std::move(it_best->prompt); states.erase(it_best); } @@ -1836,6 +1838,6 @@ void server_prompt_cache::update() { for (const auto & state : states) { SRV_TRC(" - prompt %p: %7d tokens, checkpoints: %2zu, %9.3f MiB\n", - (const void *)&state, state.n_tokens(), state.checkpoints.size(), state.size() / (1024.0 * 1024.0)); + (const void *)&state, state.prompt.n_tokens(), state.prompt.checkpoints.size(), state.size() / (1024.0 * 1024.0)); } } diff --git a/tools/server/server-task.h b/tools/server/server-task.h index dc6b2dac1..c3eea2ecb 100644 --- a/tools/server/server-task.h +++ b/tools/server/server-task.h @@ -584,32 +584,14 @@ struct server_task_result_apply_lora : server_task_result { virtual json to_json() override; }; -struct server_prompt_data { - std::vector main; - std::vector drft; - - size_t size() const { - return main.size() + drft.size(); - } -}; - struct server_prompt { server_tokens tokens; - server_prompt_data data; - std::list checkpoints; - size_t size() const { - size_t res = 0; - - res += data.size(); - - for (const auto & ckpt : checkpoints) { - res += ckpt.size(); - } - - return res; + void clear() { + tokens.clear(); + checkpoints.clear(); } int n_tokens() const { @@ -619,19 +601,42 @@ struct server_prompt { server_prompt clone() const { return server_prompt { tokens.clone(), - data, checkpoints, }; } }; +struct server_prompt_data { + std::vector main; + std::vector drft; + + size_t size() const { + return main.size() + drft.size(); + } +}; + +struct server_prompt_cache_state { + server_prompt prompt; + server_prompt_data data; + + size_t size() const { + size_t res = data.size(); + + for (const auto & ckpt : prompt.checkpoints) { + res += ckpt.size(); + } + + return res; + } +}; + struct server_prompt_cache { server_prompt_cache(int32_t limit_size_mib, size_t limit_tokens) { this->limit_size = 1024ull*1024ull*(limit_size_mib < 0 ? 0 : limit_size_mib); this->limit_tokens = limit_tokens; } - std::list states; + std::list states; // in bytes, 0 = no limit size_t limit_size = 0; @@ -643,7 +648,7 @@ struct server_prompt_cache { size_t n_tokens() const; - server_prompt * alloc(const server_prompt & prompt, size_t state_size_main, size_t state_size_drft); + server_prompt_cache_state * alloc(const server_prompt & prompt, size_t state_size_main, size_t state_size_drft); bool load(server_prompt & prompt, const server_tokens & tokens_new, llama_context * ctx_main, llama_context * ctx_drft, int32_t id_slot); From c71854292f7c367cc3b35939f88121d81945472f Mon Sep 17 00:00:00 2001 From: Chyan <163109379+chyan8@users.noreply.github.com> Date: Wed, 15 Jul 2026 03:27:09 +0800 Subject: [PATCH 21/24] hexagon: fix hmx-queue signal enum-narrowing problem (#25677) --- ggml/src/ggml-hexagon/htp/hmx-queue.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml/src/ggml-hexagon/htp/hmx-queue.c b/ggml/src/ggml-hexagon/htp/hmx-queue.c index 5f6a5e206..3add542ba 100644 --- a/ggml/src/ggml-hexagon/htp/hmx-queue.c +++ b/ggml/src/ggml-hexagon/htp/hmx-queue.c @@ -38,7 +38,7 @@ static inline void hmx_queue_process(struct hmx_queue *q, bool* killed) { if (!d->done) { FARF(HIGH, "hmx-queue-process: ir %u func %p data %p", ir, d->func, d->data); - enum hmx_queue_signal sig = (enum hmx_queue_signal) (unsigned int) d->func; + uintptr_t sig = (uintptr_t) d->func; switch (sig) { case HMX_QUEUE_NOOP: /* noop */; break; case HMX_QUEUE_KILL: *killed = true; break; From a4ce2595c55eaed284f509fab398f4ddafe5a3a4 Mon Sep 17 00:00:00 2001 From: Hongqiang Wang Date: Tue, 14 Jul 2026 12:27:56 -0700 Subject: [PATCH 22/24] opencl: avoid the vec path in GEMV for unaligned row stride (#25671) The f16 GEMV kernels take a vectorized path for ne00 >= 128 that casts the row pointers to half4 or float4. When the row stride is not aligned, the wide load becomes misaligned. On devices that require natural alignment for vector loads, the kernel reads garbage. This is the case Intel GPUs and the kernels produce incorrect results there. Adreno happpens to be byte addressable and the kernels happen to work. --- ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl | 9 ++++++++- ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl | 9 ++++++++- ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl | 9 ++++++++- 3 files changed, 24 insertions(+), 3 deletions(-) diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl index 9393b5494..b4b03eb11 100644 --- a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +++ b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl @@ -64,7 +64,14 @@ kernel void kernel_mul_mat_f16_f16( global half * x = (global half *) (src0 + offset_src0); - if (ne00 < 128) { + // The vector path below casts the row pointers to half4, which must be 8-byte aligned. + // A row address is r0*nb01 + ..., and a permuted or strided src leaves nb01/nb11 + // unconstrained -- an odd ne00, say, gives a row that is only 2-byte aligned. Every + // src1 row this work-item walks is src1_base + r1*nb11, so require both. + const ulong src1_base = (ulong) (src1 + (i12)*nb12 + (i13)*nb13); + const bool row_aligned = (((ulong) x) & 7) == 0 && (src1_base & 7) == 0 && (nb11 & 7) == 0; + + if (ne00 < 128 || !row_aligned) { for (int row = 0; row < N_F16_F16; ++row) { int r1 = rb + row; if (r1 >= ne11) { diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl index e52d3c6d4..8f3ed9c7b 100644 --- a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +++ b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl @@ -64,7 +64,14 @@ kernel void kernel_mul_mat_f16_f32( global half * x = (global half *) (src0 + offset_src0); - if (ne00 < 128) { + // The vector path below casts the row pointers to half4/float4, which must be 8- and + // 16-byte aligned. A row address is r0*nb01 + ..., and a permuted or strided src leaves + // nb01/nb11 unconstrained -- an odd ne00, say, gives a row that is only 2-byte aligned. + // Every src1 row this work-item walks is src1_base + r1*nb11, so require both. + const ulong src1_base = (ulong) (src1 + (i12)*nb12 + (i13)*nb13); + const bool row_aligned = (((ulong) x) & 7) == 0 && (src1_base & 15) == 0 && (nb11 & 15) == 0; + + if (ne00 < 128 || !row_aligned) { for (int row = 0; row < N_F16_F32; ++row) { int r1 = rb + row; if (r1 >= ne11) { diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl index 28d30212c..eca45615e 100644 --- a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +++ b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl @@ -64,8 +64,15 @@ kernel void kernel_mul_mat_f16_f32_1row( global half * x = (global half *) (src0 + offset_src0); global float * y = (global float *) (src1 + offset_src1); + // The vector path below casts the row pointers to half4/float4, which must be 8- and + // 16-byte aligned. A row address is r0*nb01 + ..., and a permuted or strided src leaves + // nb01/nb11 unconstrained -- an odd ne00, say, gives a row that is only 2-byte aligned. + // Take the vector path only when the rows this work-item touches are actually aligned; + // the scalar loop has no such requirement. + const bool row_aligned = (((ulong) x) & 7) == 0 && (((ulong) y) & 15) == 0; + float sumf = 0; - if (ne00 < 128) { + if (ne00 < 128 || !row_aligned) { for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) { sumf += (float) x[i] * (float) y[i]; } From 00fa7cb284cbf133fc426733bd64238a3588a33e Mon Sep 17 00:00:00 2001 From: Hongqiang Wang Date: Tue, 14 Jul 2026 13:46:54 -0700 Subject: [PATCH 23/24] opencl: handle OOB write in noshuffle GEMV kernels (odd ne01) (#25640) --- ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl | 7 ++++++- ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl | 6 +++++- ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32.cl | 7 ++++++- .../ggml-opencl/kernels/gemv_noshuffle_q4_0_f32_spec.cl | 6 +++++- ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl | 7 ++++++- ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl | 7 ++++++- ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl | 7 ++++++- ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl | 7 ++++++- ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl | 7 ++++++- ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl | 7 ++++++- ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl | 6 +++++- 11 files changed, 63 insertions(+), 11 deletions(-) diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl index 9386bf25a..1f832cb25 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl @@ -296,7 +296,12 @@ kernel void kernel_gemv_noshuffle_iq4_nl_f32( // 2 outputs per fiber in wave 0 if (groupId == 0) { dst = (global float*)((global char*)dst + offsetd); - vstore2(totalSum, 0, &(dst[gid * 2])); + // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64, + // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01 + // and would overrun dst into the adjacent tensor. No-op / byte-identical when + // ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding). + if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0; + if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1; } } diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl index e83c5d068..9efede294 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl @@ -116,6 +116,10 @@ __kernel void kernel_gemv_noshuffle_q1_0_f32( if (groupId == 0) { dst = (global float*)((global char*)dst + offsetd); - dst[gid] = totalSum; + // Guard the output row. The x-grid is padded to CEIL_DIV(M,wavesize)*wavesize, + // so when ne01 is not a multiple of the wave size the tail work-items run past + // row ne01 and would overrun dst into the adjacent tensor. No-op / byte-identical + // when ne01 is wave-aligned (no padding). + if (gid < M) dst[gid] = totalSum; } } diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32.cl index 106832069..8de0de1cc 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32.cl @@ -268,7 +268,12 @@ __kernel void kernel_gemv_noshuffle_q4_0_f32( // 2 outputs per fiber in wave 0 if (groupId == 0) { dst = (global float*)((global char*)dst + offsetd); - vstore2(totalSum, 0, &(dst[gid * 2])); + // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64, + // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01 + // and would overrun dst into the adjacent tensor. No-op / byte-identical when + // ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding). + if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0; + if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1; } } diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32_spec.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32_spec.cl index 571a375da..0dca20f71 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32_spec.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32_spec.cl @@ -262,7 +262,11 @@ __kernel void kernel_gemv_noshuffle_q4_0_f32( // 2 outputs per fiber in wave 0 if (groupId == 0) { dst = (global float*)((global char*)dst + offsetd); - vstore2(totalSum, 0, &(dst[gid * 2])); + // Guard the two output rows against the padded x-grid tail overrunning dst. + // The current shape specializations are all ne01 % 128 == 0 (no padding), so + // this is a no-op / byte-identical today; keep it in lockstep with the base kernel. + if (gid * 2 + 0 < ne01) dst[gid * 2 + 0] = totalSum.s0; + if (gid * 2 + 1 < ne01) dst[gid * 2 + 1] = totalSum.s1; } } diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl index fdc147245..5fa312780 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl @@ -277,7 +277,12 @@ kernel void kernel_gemv_noshuffle_q4_1_f32( // 2 outputs per fiber in wave 0 if (groupId == 0) { dst = (global float*)((global char*)dst + offsetd); - vstore2(totalSum, 0, &(dst[gid * 2])); + // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64, + // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01 + // and would overrun dst into the adjacent tensor. No-op / byte-identical when + // ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding). + if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0; + if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1; } } diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl index dd1e2b55c..2eb20e2f7 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl @@ -312,7 +312,12 @@ kernel void kernel_gemv_noshuffle_q4_k_f32( // 2 outputs per fiber in wave 0 if (groupId == 0) { dst = (global float*)((global char*)dst + offsetd); - vstore2(totalSum, 0, &(dst[gid * 2])); + // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64, + // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01 + // and would overrun dst into the adjacent tensor. No-op / byte-identical when + // ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding). + if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0; + if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1; } } diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl index c228f717a..7dbf5a3bb 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl @@ -285,7 +285,12 @@ __kernel void kernel_gemv_noshuffle_q5_0_f32( // 2 outputs per fiber in wave 0 if (groupId == 0) { dst = (global float*)((global char*)dst + offsetd); - vstore2(totalSum, 0, &(dst[gid * 2])); + // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64, + // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01 + // and would overrun dst into the adjacent tensor. No-op / byte-identical when + // ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding). + if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0; + if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1; } } diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl index daf1308ea..ba0e2a711 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl @@ -288,7 +288,12 @@ __kernel void kernel_gemv_noshuffle_q5_1_f32( // 2 outputs per fiber in wave 0 if (groupId == 0) { dst = (global float*)((global char*)dst + offsetd); - vstore2(totalSum, 0, &(dst[gid * 2])); + // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64, + // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01 + // and would overrun dst into the adjacent tensor. No-op / byte-identical when + // ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding). + if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0; + if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1; } } diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl index c40db1666..446f46533 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl @@ -321,6 +321,11 @@ kernel void kernel_gemv_noshuffle_q5_k_f32( // 2 outputs per fiber in wave 0 if (groupId == 0) { dst = (global float*)((global char*)dst + offsetd); - vstore2(totalSum, 0, &(dst[gid * 2])); + // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64, + // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01 + // and would overrun dst into the adjacent tensor. No-op / byte-identical when + // ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding). + if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0; + if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1; } } diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl index 6f89cf968..51682eceb 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl @@ -288,6 +288,11 @@ kernel void kernel_gemv_noshuffle_q6_K_f32( if (grp == 0) { dst = (global float*)((global char*)dst + offsetd); - vstore2(total_sum, 0, &(dst[gid * 2])); + // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64, + // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01 + // and would overrun dst into the adjacent tensor (garbage downstream). + // No-op / byte-identical when ne01 % 128 == 0 (no padding). + if (gid * 2 + 0 < ne01) dst[gid * 2 + 0] = total_sum.s0; + if (gid * 2 + 1 < ne01) dst[gid * 2 + 1] = total_sum.s1; } } diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl index f5c6fb3e8..09bae2d55 100644 --- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl +++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl @@ -190,6 +190,10 @@ __kernel void kernel_gemv_noshuffle_q8_0_f32( // 1 outputs per fiber in wave 0 if (groupId == 0) { dst = (global float*)((global char*)dst + offsetd); - dst[gid] = totalSum; + // Guard the output row. The x-grid is padded to CEIL_DIV(M,wavesize)*wavesize, + // so when ne01 is not a multiple of the wave size the tail work-items run past + // row ne01 and would overrun dst into the adjacent tensor. No-op / byte-identical + // when ne01 is wave-aligned (no padding). + if (gid < M) dst[gid] = totalSum; } } From 12127defda4f41b7679cb2477a4b0d65ee6a0c8f Mon Sep 17 00:00:00 2001 From: Hongqiang Wang Date: Tue, 14 Jul 2026 19:53:56 -0700 Subject: [PATCH 24/24] opencl: do not use `clCreateBufferWithProperties` when targeting CL 2.x (#25673) --- ggml/src/ggml-opencl/ggml-opencl.cpp | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp index 7b6917592..b14ea8133 100644 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp @@ -10525,10 +10525,16 @@ static ggml_backend_buffer_t ggml_backend_opencl_buffer_type_alloc_buffer(ggml_b cl_int err; cl_mem mem = clCreateBuffer(backend_ctx->context, CL_MEM_READ_WRITE, size, NULL, &err); +#if GGML_OPENCL_TARGET_VERSION >= 300 + // clCreateBufferWithProperties and cl_mem_properties are OpenCL 3.0. Drivers older than + // that do not export the symbol, so a build targeting them fails to link. The large + // buffer extension is only ever enabled on drivers that are well past 3.0, so this path + // is dead there anyway. if (err != CL_SUCCESS && backend_ctx->adreno_use_large_buffer) { cl_mem_properties props[] = { 0x41A6 /* CL_LARGE_BUFFER_QCOM */, 1, 0 }; mem = clCreateBufferWithProperties(backend_ctx->context, props, CL_MEM_READ_WRITE, size, NULL, &err); } +#endif if (err != CL_SUCCESS) { GGML_LOG_INFO("%s: failed to allocate %.2f MiB\n", __func__, size / 1024.0 / 1024.0);