From 4cbafad4f09c82f1b32c76b714302f3937be1df7 Mon Sep 17 00:00:00 2001 From: Xuan-Son Nguyen Date: Wed, 24 Dec 2025 23:07:08 +0100 Subject: [PATCH 01/17] model: support MiMo-V2-Flash (#18328) * mimov2: convert ok * rename mimov2 --> mimo2 * fix conversion * runnable not incorrect * use sink * add_sliding_window_pattern * add swa and per-layer n_head_kv * correct params * somewhat working * correct gating func * nits * mimo2: wire RMS eps + MoE bias + converter guards * add co-author Co-authored-by: Aaryan-Kapoor * use add_rope_freq_base_swa --------- Co-authored-by: Aaryan Kapoor Co-authored-by: Aaryan-Kapoor --- convert_hf_to_gguf.py | 84 ++++++++++++++++++++++ gguf-py/gguf/constants.py | 24 ++++++- gguf-py/gguf/tensor_mapping.py | 1 + src/CMakeLists.txt | 1 + src/llama-arch.cpp | 22 ++++++ src/llama-arch.h | 1 + src/llama-hparams.h | 7 +- src/llama-model.cpp | 60 ++++++++++++++++ src/llama-model.h | 1 + src/models/mimo2-iswa.cpp | 123 +++++++++++++++++++++++++++++++++ src/models/models.h | 4 ++ 11 files changed, 324 insertions(+), 4 deletions(-) create mode 100644 src/models/mimo2-iswa.cpp diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index c638e3398..69abb7367 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -7362,6 +7362,90 @@ class MiniMaxM2Model(TextModel): return super().modify_tensors(data_torch, name, bid) +@ModelBase.register("MiMoV2FlashForCausalLM") +class MimoV2Model(TextModel): + model_arch = gguf.MODEL_ARCH.MIMO2 + + def set_gguf_parameters(self): + super().set_gguf_parameters() + + assert self.hparams["swa_head_dim"] == self.hparams["head_dim"] + assert self.hparams["swa_num_attention_heads"] == self.hparams["num_attention_heads"] + assert self.hparams["swa_v_head_dim"] == self.hparams["v_head_dim"] + assert self.hparams["topk_method"] == "noaux_tc" + + n_head_kv = self.hparams["num_key_value_heads"] + n_head_kv_swa = self.hparams["swa_num_key_value_heads"] + n_head_kv_arr = [n_head_kv_swa if use_swa == 1 else n_head_kv for use_swa in self.hparams["hybrid_layer_pattern"]] + self.gguf_writer.add_head_count_kv(n_head_kv_arr) + + self.gguf_writer.add_sliding_window(self.hparams["sliding_window"]) + self.gguf_writer.add_sliding_window_pattern(self.hparams["hybrid_layer_pattern"]) + self.gguf_writer.add_rope_freq_base_swa(self.hparams["swa_rope_theta"]) + self.gguf_writer.add_value_length(self.hparams["v_head_dim"]) + self.gguf_writer.add_expert_count(self.hparams["n_routed_experts"]) + self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"]) + + rope_dim = int(self.hparams["head_dim"] * self.hparams["partial_rotary_factor"]) + self.gguf_writer.add_rope_dimension_count(rope_dim) + + self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("layernorm_epsilon", 1e-5)) + + _experts: list[dict[str, Tensor]] | None = None + + def modify_tensors(self, data_torch, name, bid): + if name.endswith("e_score_correction_bias"): + name = name.replace("e_score_correction_bias", "e_score_correction.bias") + + if "attention_sink" in name and not name.endswith(".weight"): + name += ".weight" + + # TODO: mimo v2 does not indicate the number of next-token-prediction layers, therefore we cannot do the same way as GLM4_MOE + if "model.mtp." in name: + return [] + + # process the experts separately + if name.find("mlp.experts") != -1: + n_experts = self.hparams["n_routed_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: + tensors: list[tuple[str, Tensor]] = [] + + # merge the experts into a single 3d tensor + for w_name in ["gate_proj", "up_proj", "down_proj"]: + datas: list[Tensor] = [] + + for xid in range(n_experts): + ename_to_retrieve = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight" + datas.append(self._experts[bid][ename_to_retrieve]) + del self._experts[bid][ename_to_retrieve] + + data_torch = torch.stack(datas, dim=0) + merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight" + new_name = self.map_tensor_name(merged_name) + tensors.append((new_name, data_torch)) + + return tensors + else: + return [] + return [(self.map_tensor_name(name), data_torch)] + + def prepare_tensors(self): + super().prepare_tensors() + + if self._experts is not None: + # flatten `list[dict[str, Tensor]]` into `list[str]` + experts = [k for d in self._experts for k in d.keys()] + if len(experts) > 0: + raise ValueError(f"Unprocessed experts: {experts}") + + @ModelBase.register("PanguEmbeddedForCausalLM") class PanguEmbeddedModel(TextModel): model_arch = gguf.MODEL_ARCH.PANGU_EMBED diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index baff8547a..27578daaf 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -449,6 +449,7 @@ class MODEL_ARCH(IntEnum): RND1 = auto() PANGU_EMBED = auto() MISTRAL3 = auto() + MIMO2 = auto() LLAMA_EMBED = auto() @@ -845,6 +846,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.RND1: "rnd1", MODEL_ARCH.PANGU_EMBED: "pangu-embedded", MODEL_ARCH.MISTRAL3: "mistral3", + MODEL_ARCH.MIMO2: "mimo2", MODEL_ARCH.LLAMA_EMBED: "llama-embed", } @@ -3198,6 +3200,26 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN_EXP, MODEL_TENSOR.FFN_UP_EXP, ], + MODEL_ARCH.MIMO2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_SINKS, + 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_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_EXP_PROBS_B, + ], MODEL_ARCH.LLAMA_EMBED: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, @@ -3217,7 +3239,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_GATE_EXP, MODEL_TENSOR.FFN_DOWN_EXP, MODEL_TENSOR.FFN_UP_EXP, - ] + ], # TODO } diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 276720fcd..1690d991f 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -320,6 +320,7 @@ class TensorNameMap: MODEL_TENSOR.ATTN_SINKS: ( "model.layers.{bid}.self_attn.sinks", # openai-moe + "model.layers.{bid}.self_attn.attention_sink_bias", # mimov2 ), MODEL_TENSOR.ATTN_GATE: ( diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index 4ca897491..1e155534b 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -88,6 +88,7 @@ add_library(llama models/llama-iswa.cpp models/llama.cpp models/mamba.cpp + models/mimo2-iswa.cpp models/minicpm3.cpp models/minimax-m2.cpp models/modern-bert.cpp diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 73420d3c9..75013d8d3 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -115,6 +115,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_RND1, "rnd1" }, { LLM_ARCH_PANGU_EMBED, "pangu-embedded" }, { LLM_ARCH_MISTRAL3, "mistral3" }, + { LLM_ARCH_MIMO2, "mimo2" }, { LLM_ARCH_LLAMA_EMBED, "llama-embed" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -2190,6 +2191,27 @@ static std::set llm_get_tensor_names(llm_arch arch) { LLM_TENSOR_VISEXP_FFN_DOWN, LLM_TENSOR_VISEXP_FFN_UP, }; + case LLM_ARCH_MIMO2: + return { + LLM_TENSOR_TOKEN_EMBD, + LLM_TENSOR_OUTPUT_NORM, + LLM_TENSOR_OUTPUT, + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_Q, + LLM_TENSOR_ATTN_K, + LLM_TENSOR_ATTN_V, + LLM_TENSOR_ATTN_SINKS, + LLM_TENSOR_ATTN_OUT, + LLM_TENSOR_FFN_NORM, + LLM_TENSOR_FFN_GATE, + LLM_TENSOR_FFN_DOWN, + LLM_TENSOR_FFN_UP, + LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_GATE_EXPS, + LLM_TENSOR_FFN_DOWN_EXPS, + LLM_TENSOR_FFN_UP_EXPS, + LLM_TENSOR_FFN_EXP_PROBS_B, + }; case LLM_ARCH_GPTJ: case LLM_ARCH_UNKNOWN: return { diff --git a/src/llama-arch.h b/src/llama-arch.h index 433ee4bc1..27bdedc83 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -119,6 +119,7 @@ enum llm_arch { LLM_ARCH_RND1, LLM_ARCH_PANGU_EMBED, LLM_ARCH_MISTRAL3, + LLM_ARCH_MIMO2, LLM_ARCH_LLAMA_EMBED, LLM_ARCH_UNKNOWN, }; diff --git a/src/llama-hparams.h b/src/llama-hparams.h index f6e95b5d2..42def73f0 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -123,10 +123,11 @@ struct llama_hparams { llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE; // the size of the sliding window (0 - no SWA) uint32_t n_swa = 0; - // if swa_layers[il] == true, then layer il is SWA - // if swa_layers[il] == false, then layer il is dense (i.e. non-SWA) + // if swa_layers[il] == 1, then layer il is SWA + // if swa_layers[il] == 0, then layer il is dense (i.e. non-SWA) // by default, all layers are dense - std::array swa_layers; + // note: using uint32_t type for compatibility reason + std::array swa_layers; // for State Space Models uint32_t ssm_d_conv = 0; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 9fada915d..69075742c 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -130,6 +130,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_230B_A10B: return "230B.A10B"; case LLM_TYPE_235B_A22B: return "235B.A22B"; case LLM_TYPE_300B_A47B: return "300B.A47B"; + case LLM_TYPE_310B_A15B: return "310B.A15B"; case LLM_TYPE_355B_A32B: return "355B.A32B"; case LLM_TYPE_E2B: return "E2B"; case LLM_TYPE_E4B: return "E4B"; @@ -2339,6 +2340,22 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_MIMO2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + hparams.swa_type = LLAMA_SWA_TYPE_STANDARD; + + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); + ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa); + ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer); + + switch (hparams.n_layer) { + case 48: type = LLM_TYPE_310B_A15B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; default: throw std::runtime_error("unsupported model architecture"); } @@ -6648,6 +6665,44 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { hparams.n_ff_shexp, n_embd }, 0); } } break; + case LLM_ARCH_MIMO2: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + 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}, 0); + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i); + uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i); + uint32_t n_head = hparams.n_head(i); + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_v * n_head, n_embd }, 0); + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, TENSOR_NOT_REQUIRED); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + // non-MoE branch + 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 branch + int64_t n_ff_exp = hparams.n_ff_exp; + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, 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); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -7710,6 +7765,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { { llm = std::make_unique(*this, params); } break; + case LLM_ARCH_MIMO2: + { + llm = std::make_unique(*this, params); + } break; default: GGML_ABORT("fatal error"); } @@ -7940,6 +7999,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_PANGU_EMBED: case LLM_ARCH_AFMOE: case LLM_ARCH_QWEN3NEXT: + case LLM_ARCH_MIMO2: return LLAMA_ROPE_TYPE_NEOX; case LLM_ARCH_QWEN2VL: diff --git a/src/llama-model.h b/src/llama-model.h index 7f560d462..9c00eec75 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -123,6 +123,7 @@ enum llm_type { LLM_TYPE_230B_A10B, // Minimax M2 LLM_TYPE_235B_A22B, LLM_TYPE_300B_A47B, // Ernie MoE big + LLM_TYPE_310B_A15B, // /MiMo-V2-Flash LLM_TYPE_355B_A32B, // GLM-4.5 LLM_TYPE_E2B, LLM_TYPE_E4B, diff --git a/src/models/mimo2-iswa.cpp b/src/models/mimo2-iswa.cpp new file mode 100644 index 000000000..edc87cc9f --- /dev/null +++ b/src/models/mimo2-iswa.cpp @@ -0,0 +1,123 @@ + +#include "models.h" + +llm_build_mimo2_iswa::llm_build_mimo2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + 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_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + uint32_t n_head_l = hparams.n_head(il); + uint32_t n_head_kv_l = hparams.n_head_kv(il); + const float freq_base_l = model.get_rope_freq_base(cparams, il); + const float freq_scale_l = model.get_rope_freq_scale(cparams, il); + + cur = inpL; + + // self_attention + { + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + ggml_tensor * sinks = model.layers[il].attn_sinks; + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, sinks, nullptr, 1.0f/sqrtf(float(n_embd_head_k)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + 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, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + if (model.layers[il].ffn_gate_inp == nullptr) { + // dense branch + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = 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, true, false, + 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID, il); + cb(cur, "ffn_moe_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + 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 fca505b30..dd0e286ed 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -316,6 +316,10 @@ struct llm_build_mamba : public llm_graph_context_mamba { llm_build_mamba(const llama_model & model, const llm_graph_params & params); }; +struct llm_build_mimo2_iswa : public llm_graph_context { + llm_build_mimo2_iswa(const llama_model & model, const llm_graph_params & params); +}; + struct llm_build_minicpm3 : public llm_graph_context { llm_build_minicpm3(const llama_model & model, const llm_graph_params & params); }; From f5acfb2ffa1a0bc9d01a0a40f4d7c17421d7d570 Mon Sep 17 00:00:00 2001 From: Xuan-Son Nguyen Date: Wed, 24 Dec 2025 23:47:49 +0100 Subject: [PATCH 02/17] server: (router) add stop-timeout option (#18350) * server: (router) add stop-timeout option * also allow stop while loading * add docs * unload_lru: also wait for unload to complete --- common/arg.cpp | 12 +-- common/arg.h | 1 + tools/server/README.md | 4 +- tools/server/server-models.cpp | 172 ++++++++++++++++++++++----------- tools/server/server-models.h | 8 +- 5 files changed, 133 insertions(+), 64 deletions(-) diff --git a/common/arg.cpp b/common/arg.cpp index fded0bd26..774f8731a 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -3518,15 +3518,15 @@ void common_params_add_preset_options(std::vector & args) { [](common_params &, const std::string &) { /* unused */ } ).set_env(COMMON_ARG_PRESET_LOAD_ON_STARTUP).set_preset_only()); + args.push_back(common_arg( + {"stop-timeout"}, "SECONDS", + "in server router mode, force-kill model instance after this many seconds of graceful shutdown", + [](common_params &, int) { /* unused */ } + ).set_env(COMMON_ARG_PRESET_STOP_TIMEOUT).set_preset_only()); + // args.push_back(common_arg( // {"pin"}, // "in server router mode, do not unload this model if models_max is exceeded", // [](common_params &) { /* unused */ } // ).set_preset_only()); - - // args.push_back(common_arg( - // {"unload-idle-seconds"}, "SECONDS", - // "in server router mode, unload models idle for more than this many seconds", - // [](common_params &, int) { /* unused */ } - // ).set_preset_only()); } diff --git a/common/arg.h b/common/arg.h index f5111c658..a1b6a14e6 100644 --- a/common/arg.h +++ b/common/arg.h @@ -10,6 +10,7 @@ // pseudo-env variable to identify preset-only arguments #define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP" +#define COMMON_ARG_PRESET_STOP_TIMEOUT "__PRESET_STOP_TIMEOUT" // // CLI argument parsing diff --git a/tools/server/README.md b/tools/server/README.md index 1ae5eae4c..7d2f6f798 100644 --- a/tools/server/README.md +++ b/tools/server/README.md @@ -1486,6 +1486,7 @@ The precedence rule for preset options is as follows: We also offer additional options that are exclusive to presets (these aren't treated as command-line arguments): - `load-on-startup` (boolean): Controls whether the model loads automatically when the server starts +- `stop-timeout` (int, seconds): After requested unload, wait for this many seconds before forcing termination (default: 10) ### Routing requests @@ -1574,8 +1575,7 @@ Payload: ```json { - "model": "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M", - "extra_args": ["-n", "128", "--top-k", "4"] + "model": "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M" } ``` diff --git a/tools/server/server-models.cpp b/tools/server/server-models.cpp index 08a0da5c8..cb7e70455 100644 --- a/tools/server/server-models.cpp +++ b/tools/server/server-models.cpp @@ -34,6 +34,8 @@ #include #endif +#define DEFAULT_STOP_TIMEOUT 10 // seconds + #define CMD_ROUTER_TO_CHILD_EXIT "cmd_router_to_child:exit" #define CMD_CHILD_TO_ROUTER_READY "cmd_child_to_router:ready" @@ -203,13 +205,14 @@ void server_models::load_models() { // convert presets to server_model_meta and add to mapping for (const auto & preset : final_presets) { server_model_meta meta{ - /* preset */ preset.second, - /* name */ preset.first, - /* port */ 0, - /* status */ SERVER_MODEL_STATUS_UNLOADED, - /* last_used */ 0, - /* args */ std::vector(), - /* exit_code */ 0 + /* preset */ preset.second, + /* name */ preset.first, + /* port */ 0, + /* status */ SERVER_MODEL_STATUS_UNLOADED, + /* last_used */ 0, + /* args */ std::vector(), + /* exit_code */ 0, + /* stop_timeout */ DEFAULT_STOP_TIMEOUT, }; add_model(std::move(meta)); } @@ -227,6 +230,20 @@ void server_models::load_models() { } } + // handle custom stop-timeout option + for (auto & [name, inst] : mapping) { + std::string val; + if (inst.meta.preset.get_option(COMMON_ARG_PRESET_STOP_TIMEOUT, val)) { + try { + inst.meta.stop_timeout = std::stoi(val); + } catch (...) { + SRV_WRN("invalid stop-timeout value '%s' for model '%s', using default %d seconds\n", + val.c_str(), name.c_str(), DEFAULT_STOP_TIMEOUT); + inst.meta.stop_timeout = DEFAULT_STOP_TIMEOUT; + } + } + } + // load any autoload models std::vector models_to_load; for (const auto & [name, inst] : mapping) { @@ -362,7 +379,7 @@ void server_models::unload_lru() { int64_t lru_last_used = ggml_time_ms(); size_t count_active = 0; { - std::lock_guard lk(mutex); + std::unique_lock lk(mutex); for (const auto & m : mapping) { if (m.second.meta.is_active()) { count_active++; @@ -376,6 +393,13 @@ void server_models::unload_lru() { if (!lru_model_name.empty() && count_active >= (size_t)base_params.models_max) { SRV_INF("models_max limit reached, removing LRU name=%s\n", lru_model_name.c_str()); unload(lru_model_name); + // wait for unload to complete + { + std::unique_lock lk(mutex); + cv.wait(lk, [this, &lru_model_name]() { + return mapping[lru_model_name].meta.status == SERVER_MODEL_STATUS_UNLOADED; + }); + } } } @@ -436,38 +460,83 @@ void server_models::load(const std::string & name) { // start a thread to manage the child process // captured variables are guaranteed to be destroyed only after the thread is joined - inst.th = std::thread([this, name, child_proc = inst.subproc, port = inst.meta.port]() { - // read stdout/stderr and forward to main server log - bool state_received = false; // true if child state received - FILE * p_stdout_stderr = subprocess_stdout(child_proc.get()); - if (p_stdout_stderr) { - char buffer[4096]; - while (fgets(buffer, sizeof(buffer), p_stdout_stderr) != nullptr) { - LOG("[%5d] %s", port, buffer); - if (!state_received && std::strstr(buffer, CMD_CHILD_TO_ROUTER_READY) != nullptr) { - // child process is ready - this->update_status(name, SERVER_MODEL_STATUS_LOADED); - state_received = true; + inst.th = std::thread([this, name, child_proc = inst.subproc, port = inst.meta.port, stop_timeout = inst.meta.stop_timeout]() { + FILE * stdin_file = subprocess_stdin(child_proc.get()); + FILE * stdout_file = subprocess_stdout(child_proc.get()); // combined stdout/stderr + + std::thread log_thread([&]() { + // read stdout/stderr and forward to main server log + // also handle status report from child process + bool state_received = false; // true if child state received + if (stdout_file) { + char buffer[4096]; + while (fgets(buffer, sizeof(buffer), stdout_file) != nullptr) { + LOG("[%5d] %s", port, buffer); + if (!state_received && std::strstr(buffer, CMD_CHILD_TO_ROUTER_READY) != nullptr) { + // child process is ready + this->update_status(name, SERVER_MODEL_STATUS_LOADED, 0); + state_received = true; + } } + } else { + SRV_ERR("failed to get stdout/stderr of child process for name=%s\n", name.c_str()); } - } else { - SRV_ERR("failed to get stdout/stderr of child process for name=%s\n", name.c_str()); - } + }); + + std::thread stopping_thread([&]() { + // thread to monitor stopping signal + auto is_stopping = [this, &name]() { + return this->stopping_models.find(name) != this->stopping_models.end(); + }; + { + std::unique_lock lk(this->mutex); + this->cv_stop.wait(lk, is_stopping); + } + SRV_INF("stopping model instance name=%s\n", name.c_str()); + // send interrupt to child process + fprintf(stdin_file, "%s\n", CMD_ROUTER_TO_CHILD_EXIT); + fflush(stdin_file); + // wait to stop gracefully or timeout + int64_t start_time = ggml_time_ms(); + while (true) { + std::unique_lock lk(this->mutex); + if (!is_stopping()) { + return; // already stopped + } + int64_t elapsed = ggml_time_ms() - start_time; + if (elapsed >= stop_timeout * 1000) { + // timeout, force kill + SRV_WRN("force-killing model instance name=%s after %d seconds timeout\n", name.c_str(), stop_timeout); + subprocess_terminate(child_proc.get()); + return; + } + this->cv_stop.wait_for(lk, std::chrono::seconds(1)); + } + }); + // we reach here when the child process exits + // note: we cannot join() prior to this point because it will close stdin_file + if (log_thread.joinable()) { + log_thread.join(); + } + + // stop the timeout monitoring thread + { + std::lock_guard lk(this->mutex); + stopping_models.erase(name); + cv_stop.notify_all(); + } + if (stopping_thread.joinable()) { + stopping_thread.join(); + } + + // get the exit code int exit_code = 0; subprocess_join(child_proc.get(), &exit_code); subprocess_destroy(child_proc.get()); - // update PID and status - { - std::lock_guard lk(mutex); - auto it = mapping.find(name); - if (it != mapping.end()) { - auto & meta = it->second.meta; - meta.exit_code = exit_code; - meta.status = SERVER_MODEL_STATUS_UNLOADED; - } - cv.notify_all(); - } + + // update status and exit code + this->update_status(name, SERVER_MODEL_STATUS_UNLOADED, exit_code); SRV_INF("instance name=%s exited with status %d\n", name.c_str(), exit_code); }); @@ -488,22 +557,14 @@ void server_models::load(const std::string & name) { cv.notify_all(); } -static void interrupt_subprocess(FILE * stdin_file) { - // because subprocess.h does not provide a way to send SIGINT, - // we will send a command to the child process to exit gracefully - if (stdin_file) { - fprintf(stdin_file, "%s\n", CMD_ROUTER_TO_CHILD_EXIT); - fflush(stdin_file); - } -} - void server_models::unload(const std::string & name) { std::lock_guard lk(mutex); auto it = mapping.find(name); if (it != mapping.end()) { if (it->second.meta.is_active()) { SRV_INF("unloading model instance name=%s\n", name.c_str()); - interrupt_subprocess(it->second.stdin_file); + stopping_models.insert(name); + cv_stop.notify_all(); // status change will be handled by the managing thread } else { SRV_WRN("model instance name=%s is not loaded\n", name.c_str()); @@ -518,7 +579,8 @@ void server_models::unload_all() { for (auto & [name, inst] : mapping) { if (inst.meta.is_active()) { SRV_INF("unloading model instance name=%s\n", name.c_str()); - interrupt_subprocess(inst.stdin_file); + stopping_models.insert(name); + cv_stop.notify_all(); // status change will be handled by the managing thread } // moving the thread to join list to avoid deadlock @@ -532,16 +594,15 @@ void server_models::unload_all() { } } -void server_models::update_status(const std::string & name, server_model_status status) { - // for now, we only allow updating to LOADED status - if (status != SERVER_MODEL_STATUS_LOADED) { - throw std::runtime_error("invalid status value"); - } - auto meta = get_meta(name); - if (meta.has_value()) { - meta->status = status; - update_meta(name, meta.value()); +void server_models::update_status(const std::string & name, server_model_status status, int exit_code) { + std::unique_lock lk(mutex); + auto it = mapping.find(name); + if (it != mapping.end()) { + auto & meta = it->second.meta; + meta.status = status; + meta.exit_code = exit_code; } + cv.notify_all(); } void server_models::wait_until_loaded(const std::string & name) { @@ -568,6 +629,7 @@ bool server_models::ensure_model_loaded(const std::string & name) { load(name); } + // for loading state SRV_INF("waiting until model name=%s is fully loaded...\n", name.c_str()); wait_until_loaded(name); @@ -795,7 +857,7 @@ void server_models_routes::init_routes() { res_err(res, format_error_response("model is not found", ERROR_TYPE_INVALID_REQUEST)); return res; } - if (model->status != SERVER_MODEL_STATUS_LOADED) { + if (!model->is_active()) { res_err(res, format_error_response("model is not loaded", ERROR_TYPE_INVALID_REQUEST)); return res; } diff --git a/tools/server/server-models.h b/tools/server/server-models.h index 3e1868c27..7e3353753 100644 --- a/tools/server/server-models.h +++ b/tools/server/server-models.h @@ -9,6 +9,7 @@ #include #include #include +#include /** * state diagram: @@ -56,6 +57,7 @@ struct server_model_meta { int64_t last_used = 0; // for LRU unloading std::vector args; // args passed to the model instance, will be populated by render_args() int exit_code = 0; // exit code of the model instance process (only valid if status == FAILED) + int stop_timeout = 0; // seconds to wait before force-killing the model instance during shutdown bool is_active() const { return status == SERVER_MODEL_STATUS_LOADED || status == SERVER_MODEL_STATUS_LOADING; @@ -83,6 +85,10 @@ private: std::condition_variable cv; std::map mapping; + // for stopping models + std::condition_variable cv_stop; + std::set stopping_models; + common_preset_context ctx_preset; common_params base_params; @@ -119,7 +125,7 @@ public: void unload_all(); // update the status of a model instance (thread-safe) - void update_status(const std::string & name, server_model_status status); + void update_status(const std::string & name, server_model_status status, int exit_code); // wait until the model instance is fully loaded (thread-safe) // return when the model is loaded or failed to load From c54bba869da97feeaf788cac499679c9a9ff9fea Mon Sep 17 00:00:00 2001 From: Aadeshveer Singh <24b0926@iitb.ac.in> Date: Thu, 25 Dec 2025 09:41:13 +0530 Subject: [PATCH 03/17] ggml : optimize cuda cumsum fallback kernel (#18343) --- ggml/src/ggml-cuda/cumsum.cu | 53 ++++++++++++++++++++++++++---------- 1 file changed, 38 insertions(+), 15 deletions(-) diff --git a/ggml/src/ggml-cuda/cumsum.cu b/ggml/src/ggml-cuda/cumsum.cu index d2f2def8b..0f72e33bb 100644 --- a/ggml/src/ggml-cuda/cumsum.cu +++ b/ggml/src/ggml-cuda/cumsum.cu @@ -69,7 +69,7 @@ static __global__ void cumsum_cub_kernel( #endif // GGML_CUDA_USE_CUB } -// Fallback kernel implementation (original) +// Fallback kernel implementation template static __global__ void cumsum_kernel( const T * src, T * dst, @@ -86,10 +86,10 @@ static __global__ void cumsum_kernel( const int warps_per_block = blockDim.x / warp_size; extern __shared__ float smem[]; - float * s_vals = smem; - float * s_warp_sums = smem + blockDim.x; - float * s_carry = smem + blockDim.x + warps_per_block; - float * s_chunk_total = s_carry + 1; + float * s_vals = smem; + float * s_warp_sums = smem + blockDim.x; + float * s_carry = smem + blockDim.x + warps_per_block; + float * s_chunk_total = s_carry + 1; // Initialize carry if (tid == 0) { @@ -107,21 +107,39 @@ static __global__ void cumsum_kernel( const T * src_row = src + i1 * s01 + i2 * s02 + i3 * s03; T * dst_row = dst + i1 * s1 + i2 * s2 + i3 * s3; - for (int64_t start = 0; start < ne00; start += blockDim.x) { - int64_t idx = start + tid; - float val = (idx < ne00) ? ggml_cuda_cast(src_row[idx]) : 0.0f; + // register blocking: process 4 elements per thread to hide latency + // and reduce synchronization overhead + constexpr int num_unroll = 4; + T temp[num_unroll]; - // 1. Warp inclusive scan + for (int64_t i = 0; i < ne00; i += num_unroll * blockDim.x) { + int64_t idx = i + tid * num_unroll; + + // thread local sequential scan + temp[0] = (idx < ne00 ? src_row[idx] : T(0)); +#pragma unroll + for (int64_t j = 1; j < num_unroll; j++) { + temp[j] = temp[j - 1]; + if (idx + j < ne00) { + temp[j] += src_row[idx + j]; + } else { + temp[j] += 0; + } + } + + // last emenent is sum of all values assigned to thread + float val = (idx < ne00) ? ggml_cuda_cast(temp[num_unroll - 1]) : 0.0f; + + // Warp inclusive scan val = warp_prefix_inclusive_sum(val); s_vals[tid] = val; - // Store warp total if (lane == warp_size - 1) { s_warp_sums[warp] = val; } __syncthreads(); - // 2. Exclusive scan of warp sums (warp 0 only) + // Exclusive scan of warp sums (warp 0 only) if (warp == 0) { float w = (tid < warps_per_block) ? s_warp_sums[tid] : 0.0f; float inc = warp_prefix_inclusive_sum(w); @@ -134,12 +152,17 @@ static __global__ void cumsum_kernel( } __syncthreads(); + // write back results float carry = *s_carry; - float final_val = s_vals[tid] + s_warp_sums[warp] + carry; - if (idx < ne00) { - dst_row[idx] = ggml_cuda_cast(final_val); + // calculate sum offset for this thread + float final_val_offset = s_vals[tid] + s_warp_sums[warp] + carry - temp[num_unroll - 1]; + +#pragma unroll + for (int32_t j = 0; j < num_unroll; j++) { + if (idx + j < ne00) { + dst_row[idx + j] = temp[j] + ggml_cuda_cast(final_val_offset); + } } - __syncthreads(); // Update carry for next chunk if (tid == 0) { From e68c19b0fdbb18d7a19217194c2795897e9c683c Mon Sep 17 00:00:00 2001 From: Penglin Cai <1402538448@qq.com> Date: Thu, 25 Dec 2025 16:46:09 +0800 Subject: [PATCH 04/17] CANN: Add support for CONV_TRANSPOSE_1D when kernel size > 255 (#17934) * CONV_TRANSPOSE_1D kernel_size>255 * remove condition check * fix the bug of type conversion * removing trailing whitespaces * fix: return true in the switch case --- ggml/src/ggml-cann/aclnn_ops.cpp | 150 ++++++++++++++++++++++++++++--- ggml/src/ggml-cann/aclnn_ops.h | 1 + ggml/src/ggml-cann/ggml-cann.cpp | 3 +- 3 files changed, 139 insertions(+), 15 deletions(-) diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp index dff72a277..3688abdd5 100644 --- a/ggml/src/ggml-cann/aclnn_ops.cpp +++ b/ggml/src/ggml-cann/aclnn_ops.cpp @@ -2990,32 +2990,156 @@ void ggml_cann_argmax(ggml_backend_cann_context & ctx, ggml_tensor * dst) { GGML_CANN_CALL_ACLNN_OP(ctx, ArgMax, acl_src.get(), 3, false, acl_dst.get()); } -void ggml_cann_conv_transpose_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst) { +void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst){ ggml_tensor * src0 = dst->src[0]; ggml_tensor * src1 = dst->src[1]; // stride - int64_t s0 = ((const int32_t *) (dst->op_params))[0]; + int64_t s0 = ((const int32_t*)(dst->op_params))[0]; - acl_tensor_ptr acl_input = ggml_cann_create_tensor(src1, src1->ne, src1->nb, 3, ACL_FORMAT_NCL); + acl_tensor_ptr acl_input = ggml_cann_create_tensor(src1, src1->ne, src1->nb, 3, ACL_FORMAT_NCL); acl_tensor_ptr acl_weight = ggml_cann_create_tensor(src0, src0->ne, src0->nb, 3, ACL_FORMAT_NCL); - acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst, dst->ne, dst->nb, 3, ACL_FORMAT_NCL); + acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst, dst->ne, dst->nb, 3, ACL_FORMAT_NCL); + + // get base information of input and kernel + int64_t input_len = *(src1->ne); + int64_t dst_len = *(dst->ne); + int64_t kernel_size = *(src0->ne); + + // set the max kernel size for each conv + int64_t max_kernel_size = 255; + + // compute the partition of kernel + int64_t part_num = 1; + part_num = (kernel_size + max_kernel_size - 1) / max_kernel_size; int64_t strideVal[1]; - strideVal[0] = s0; - acl_int_array_ptr stride = ggml_cann_create_int_array(strideVal, 1); - int64_t paddingVal[] = { 0 }; - acl_int_array_ptr padding = ggml_cann_create_int_array(paddingVal, 1); - int64_t dilationVal[] = { 1 }; - acl_int_array_ptr dilation = ggml_cann_create_int_array(dilationVal, 1); - int8_t cubeMathType = 0; + strideVal[0] = s0; + acl_int_array_ptr stride = ggml_cann_create_int_array(strideVal, 1); + int64_t paddingVal[] = {0}; + acl_int_array_ptr padding = ggml_cann_create_int_array(paddingVal, 1); + int64_t dilationVal[] = {1}; + acl_int_array_ptr dilation = ggml_cann_create_int_array(dilationVal, 1); + bool transposed = true; + int64_t groups = 1; + int8_t cubeMathType = 0; #ifdef ASCEND_310P cubeMathType = 1; #endif - GGML_CANN_CALL_ACLNN_OP(ctx, Convolution, acl_input.get(), acl_weight.get(), nullptr, stride.get(), padding.get(), - dilation.get(), true, padding.get(), 1, acl_dst.get(), cubeMathType); + auto weight_type = ggml_cann_type_mapping(src0->type); + auto dst_type = ggml_cann_type_mapping(dst->type); + + // slice the kernel to make each conv available + int64_t slice_dim = -1; + int64_t slice_start = 0; + int64_t slice_end = max_kernel_size; + int64_t slice_step = 1; + int64_t interval = max_kernel_size; + + int64_t left_pad_len = dilationVal[0] * (max_kernel_size - 1) + 1 - 2 * paddingVal[0]; + int64_t right_pad_len = 0; + + acl_scalar_ptr alpha = nullptr; + float alphaValue = 1.0; + alpha = ggml_cann_create_scalar(&alphaValue, aclDataType::ACL_FLOAT); + + // set zero to destination + GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, acl_dst.get()); + + for(int k = 0; k < part_num; k++){ + + // create part kernel tensor and slice from big kernel + slice_start = max_kernel_size * k; + if(k == part_num - 1){ + slice_end = kernel_size; + interval = kernel_size - max_kernel_size * k; + }else{ + slice_end = max_kernel_size * (k+1); + } + + int64_t part_ne[4]; + for(int i = 0; i < 4; i++) { + part_ne[i] = *(src0->ne + i); + } + part_ne[0] = interval; + + size_t part_nb[4]; + part_nb[0] = sizeof(weight_type); + for (int i = 1; i < 4; i++) { + part_nb[i] = part_nb[i - 1] * part_ne[i - 1]; + } + + ggml_cann_pool_alloc part_kernel_allocator; + part_kernel_allocator.alloc(ctx.pool(), part_nb[3]); + void* part_kernel_buf = part_kernel_allocator.get(); + + acl_tensor_ptr part_kernel = ggml_cann_create_tensor(part_kernel_buf, weight_type, + ggml_element_size(src0), part_ne, part_nb, 3, ACL_FORMAT_NCL); + + GGML_CANN_CALL_ACLNN_OP(ctx, Slice, acl_weight.get(), slice_dim, slice_start, slice_end, slice_step, part_kernel.get()); + + // create the part conv result tensor + int64_t part_dst_ne[4]; + for(int i = 0; i < 4; i++){ + part_dst_ne[i] = *(dst->ne + i); + } + part_dst_ne[0] = (input_len - 1) * strideVal[0] - 2 * paddingVal[0] + dilationVal[0] * (part_ne[0] - 1) + 1; + + size_t part_dst_nb[4]; + part_dst_nb[0] = sizeof(weight_type); + for (int i = 1; i < 4; i++) { + part_dst_nb[i] = part_dst_nb[i - 1] * part_dst_ne[i - 1]; + } + ggml_cann_pool_alloc part_dst_allocator; + part_dst_allocator.alloc(ctx.pool(), part_dst_nb[3]); + void* part_dst_buf = part_dst_allocator.get(); + + acl_tensor_ptr acl_part_dst = ggml_cann_create_tensor(part_dst_buf, dst_type, ggml_element_size(dst), + part_dst_ne, part_dst_nb, 3, ACL_FORMAT_NCL); + GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, acl_part_dst.get()); + + // compute part conv transpose 1d + GGML_CANN_CALL_ACLNN_OP(ctx, Convolution, acl_input.get(), part_kernel.get(), nullptr, stride.get(), + padding.get(), dilation.get(), transposed, padding.get(), groups, acl_part_dst.get(), cubeMathType); + + // compute the position of part result in final result + int64_t global_start = slice_start; + int64_t global_end = std::min((input_len - 1) * strideVal[0] + slice_end, dst_len); + + left_pad_len = global_start; + right_pad_len = dst_len - global_end; + + std::vector padDataVal = {left_pad_len,right_pad_len}; + acl_int_array_ptr padData = ggml_cann_create_int_array(padDataVal.data(), 2); + + acl_scalar_ptr pad_value = nullptr; + float pad_valueVal = 0.0; + pad_value = ggml_cann_create_scalar(&pad_valueVal, aclDataType::ACL_FLOAT); + + int64_t conv_result_ne[4]; + for(int i = 0; i < 4; i++){ + conv_result_ne[i] = *(dst->ne + i); + } + + size_t conv_result_nb[4]; + conv_result_nb[0] = sizeof(weight_type); + for (int i = 1; i < 4; i++) { + conv_result_nb[i] = conv_result_nb[i - 1] * conv_result_ne[i - 1]; + } + + ggml_cann_pool_alloc conv_result_allocator; + conv_result_allocator.alloc(ctx.pool(), conv_result_nb[3]); + void* conv_result_buf = conv_result_allocator.get(); + + acl_tensor_ptr conv_result = ggml_cann_create_tensor(conv_result_buf, dst_type, ggml_element_size(dst), + conv_result_ne, conv_result_nb, 3, ACL_FORMAT_NCL); + + GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, conv_result.get()); + GGML_CANN_CALL_ACLNN_OP(ctx, ConstantPadNd, acl_part_dst.get(), padData.get(), pad_value.get(), conv_result.get()); + GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdd, acl_dst.get(), conv_result.get(), alpha.get()); + } } void ggml_cann_elu(ggml_backend_cann_context & ctx, ggml_tensor * dst) { diff --git a/ggml/src/ggml-cann/aclnn_ops.h b/ggml/src/ggml-cann/aclnn_ops.h index 1ebbc769c..6ec464028 100644 --- a/ggml/src/ggml-cann/aclnn_ops.h +++ b/ggml/src/ggml-cann/aclnn_ops.h @@ -47,6 +47,7 @@ #include #include #include +#include #include #include diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp index 402e86d70..e90759f98 100644 --- a/ggml/src/ggml-cann/ggml-cann.cpp +++ b/ggml/src/ggml-cann/ggml-cann.cpp @@ -2424,8 +2424,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten } } case GGML_OP_CONV_TRANSPOSE_1D: - // TODO: ((weightL - 1) * dilationW - padLeft)=1336 should not be larger than 255. - return (op->src[0]->ne[0] - 1) <= 255; + return true; case GGML_OP_SCALE: float bias; memcpy(&bias, (const float *) (op->op_params) + 1, sizeof(float)); From b0fb0f0aee2c931cbc08e95aef68f615c6de90a2 Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Thu, 25 Dec 2025 22:12:11 +0800 Subject: [PATCH 05/17] ggml-cuda: fix blackwell native builds (#18361) * ggml-cuda: fix blackwell native builds Replace 12x in native architectures by 12xa * replace for GGML_NATIVE=OFF too * only replace for native * remove 120f-virtual for default compilation --------- Co-authored-by: Aman Gupta --- ggml/src/ggml-cuda/CMakeLists.txt | 25 +++++++++++++++++++++---- 1 file changed, 21 insertions(+), 4 deletions(-) diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt index f1412e8b1..f3c3ebff7 100644 --- a/ggml/src/ggml-cuda/CMakeLists.txt +++ b/ggml/src/ggml-cuda/CMakeLists.txt @@ -35,16 +35,33 @@ if (CUDAToolkit_FOUND) if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8") list(APPEND CMAKE_CUDA_ARCHITECTURES 89-real) endif() - - if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8") - list(APPEND CMAKE_CUDA_ARCHITECTURES 120f-virtual) - endif() endif() endif() message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}") enable_language(CUDA) + # Replace any 12x-real architectures with 12x{a}-real. FP4 ptx instructions are not available in just 12x + if (GGML_NATIVE) + set(PROCESSED_ARCHITECTURES "") + foreach(ARCH ${CMAKE_CUDA_ARCHITECTURES_NATIVE}) + if(ARCH MATCHES "^12[0-9]$") + string(REGEX REPLACE "^(12[0-9]).*$" "\\1" BASE_ARCH ${ARCH}) + message(STATUS "Replacing ${ARCH} with ${BASE_ARCH}a-real") + list(APPEND PROCESSED_ARCHITECTURES "${BASE_ARCH}a-real") + else() + list(APPEND PROCESSED_ARCHITECTURES ${ARCH}) + endif() + endforeach() + set(CMAKE_CUDA_ARCHITECTURES ${PROCESSED_ARCHITECTURES}) + else() + foreach(ARCH ${CMAKE_CUDA_ARCHITECTURES}) + if(ARCH MATCHES "^12[0-9]$") + message(FATAL_ERROR "Compute capability ${ARCH} used, use ${ARCH}a or ${ARCH}f for Blackwell specific optimizations") + endif() + endforeach() + endif() + file(GLOB GGML_HEADERS_CUDA "*.cuh") list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h") From 83b3b1c271c78bd77664120431aa8c354d68daac Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Thu, 25 Dec 2025 23:55:38 +0800 Subject: [PATCH 06/17] cuda: optimize cumsum cub path (#18362) * cuda: optimize cumsum cub path * remove heavy perf test --- ggml/src/ggml-cuda/cumsum.cu | 51 ++++++++++++++++++++++-------------- 1 file changed, 31 insertions(+), 20 deletions(-) diff --git a/ggml/src/ggml-cuda/cumsum.cu b/ggml/src/ggml-cuda/cumsum.cu index 0f72e33bb..e82171f9c 100644 --- a/ggml/src/ggml-cuda/cumsum.cu +++ b/ggml/src/ggml-cuda/cumsum.cu @@ -5,7 +5,7 @@ #include "ggml.h" #ifdef GGML_CUDA_USE_CUB -# include +# include #endif // GGML_CUDA_USE_CUB template @@ -16,12 +16,14 @@ static __global__ void cumsum_cub_kernel( const int64_t s01, const int64_t s02, const int64_t s03, const int64_t s1, const int64_t s2, const int64_t s3) { #ifdef GGML_CUDA_USE_CUB - using BlockScan = cub::BlockScan; + using BlockScanT = cub::BlockScan; - __shared__ typename BlockScan::TempStorage temp_storage; - __shared__ T block_carry; // carry from previous tile + __shared__ typename BlockScanT::TempStorage temp_storage; + __shared__ T block_carry; const int tid = threadIdx.x; + constexpr int UNROLL_FACTOR = 4; + constexpr int TILE_SIZE = BLOCK_SIZE * UNROLL_FACTOR; const int64_t i1 = blockIdx.x; const int64_t i2 = blockIdx.y; @@ -39,29 +41,38 @@ static __global__ void cumsum_cub_kernel( } __syncthreads(); - for (int64_t start = 0; start < ne00; start += BLOCK_SIZE) { - int64_t idx = start + tid; - T x = (idx < ne00) ? src_row[idx] : T(0); + for (int64_t start = 0; start < ne00; start += TILE_SIZE) { + T items[UNROLL_FACTOR]; + T thread_sum = T(0); - T inclusive; - T block_total; - BlockScan(temp_storage).InclusiveSum(x, inclusive, block_total); - - __syncthreads(); - - T final_val = inclusive + block_carry; - - // store result - if (idx < ne00) { - dst_row[idx] = final_val; +#pragma unroll + for (int i = 0; i < UNROLL_FACTOR; i++) { + int64_t idx = start + tid * UNROLL_FACTOR + i; + T val = (idx < ne00) ? src_row[idx] : T(0); + thread_sum += val; + items[i] = thread_sum; } + // Block-wide scan on thread sums + T thread_prefix; + T block_total; + BlockScanT(temp_storage).InclusiveSum(thread_sum, thread_prefix, block_total); __syncthreads(); + // Add offset to each item and store + T thread_offset = thread_prefix - thread_sum + block_carry; + #pragma unroll + for (int i = 0; i < UNROLL_FACTOR; i++) { + int64_t idx = start + tid * UNROLL_FACTOR + i; + if (idx < ne00) { + dst_row[idx] = items[i] + thread_offset; + } + } + + // Update carry for next tile if (tid == 0) { block_carry += block_total; } - __syncthreads(); } #else @@ -200,7 +211,7 @@ static void cumsum_cuda( const int warps_per_block = block_size / warp_size; const size_t shmem_size = (block_size + warps_per_block + 2) * sizeof(float); - if (use_cub) { + if (use_cub && ne00 >= 1024) { cumsum_cub_kernel<<>>( src, dst, ne00, ne01, ne02, ne03, From 85c40c9b02941ebf1add1469af75f1796d513ef4 Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Fri, 26 Dec 2025 01:35:14 +0800 Subject: [PATCH 07/17] ggml-cuda: fix regex for arch list (#18371) * ggml-cuda: fix regex for arch list * make regex exact --- ggml/src/ggml-cuda/CMakeLists.txt | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt index f3c3ebff7..c0f8bcaa3 100644 --- a/ggml/src/ggml-cuda/CMakeLists.txt +++ b/ggml/src/ggml-cuda/CMakeLists.txt @@ -44,8 +44,13 @@ if (CUDAToolkit_FOUND) # Replace any 12x-real architectures with 12x{a}-real. FP4 ptx instructions are not available in just 12x if (GGML_NATIVE) set(PROCESSED_ARCHITECTURES "") - foreach(ARCH ${CMAKE_CUDA_ARCHITECTURES_NATIVE}) - if(ARCH MATCHES "^12[0-9]$") + if (CMAKE_CUDA_ARCHITECTURES_NATIVE) + set(ARCH_LIST ${CMAKE_CUDA_ARCHITECTURES_NATIVE}) + else() + set(ARCH_LIST ${CMAKE_CUDA_ARCHITECTURES}) + endif() + foreach(ARCH ${ARCH_LIST}) + if (ARCH MATCHES "^12[0-9](-real|-virtual)?$") string(REGEX REPLACE "^(12[0-9]).*$" "\\1" BASE_ARCH ${ARCH}) message(STATUS "Replacing ${ARCH} with ${BASE_ARCH}a-real") list(APPEND PROCESSED_ARCHITECTURES "${BASE_ARCH}a-real") From b07cda687c6e5667b574c5e33eabad58a8dc7b12 Mon Sep 17 00:00:00 2001 From: 0Marble <85058989+0Marble@users.noreply.github.com> Date: Fri, 26 Dec 2025 09:12:04 +0800 Subject: [PATCH 08/17] CANN: implement the SSM_CONV operator (#17737) * CANN: implement SSM_CONV operator Co-authored-by: Aleksei Lobanov, Co-authored-by: Sujin Kang, * CANN: remove custom error limit for SSM_CONV * CANN: merge SSM_CONV tensor shape/strides into one line --------- Co-authored-by: Sujin Kang, --- ggml/src/ggml-cann/aclnn_ops.cpp | 103 +++++++++++++++++++++++++++++++ ggml/src/ggml-cann/aclnn_ops.h | 2 + ggml/src/ggml-cann/ggml-cann.cpp | 4 ++ 3 files changed, 109 insertions(+) diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp index 3688abdd5..2180a06fd 100644 --- a/ggml/src/ggml-cann/aclnn_ops.cpp +++ b/ggml/src/ggml-cann/aclnn_ops.cpp @@ -3702,3 +3702,106 @@ void ggml_cann_out_prod(ggml_backend_cann_context & ctx, ggml_tensor * dst) { break; } } + +void ggml_cann_ssm_conv(ggml_backend_cann_context & ctx, ggml_tensor * dst) { + ggml_tensor * src0 = dst->src[0]; // conv_x + ggml_tensor * src1 = dst->src[1]; // conv1d.weight + + // This op is currently defined only for F32 in ggml_cpu + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + // Shapes follow ggml_compute_forward_ssm_conv_f32 + const int64_t nc = src1->ne[0]; // d_conv + const int64_t ncs = src0->ne[0]; // d_conv - 1 + n_t + const int64_t nr = src0->ne[1]; // d_inner + const int64_t n_s = src0->ne[2]; // n_seqs + + const int64_t n_t = dst->ne[1]; // tokens per sequence + + GGML_ASSERT(dst->ne[0] == nr); // dst: {d_inner, n_t, n_s} + GGML_ASSERT(src1->ne[1] == nr); // weight: {d_conv, d_inner} + GGML_ASSERT(ncs == nc - 1 + n_t); // conv_x: {d_conv - 1 + n_t, d_inner, n_s} + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(src1->nb[0] == sizeof(float)); + + // --- Build CANN tensors --- + + // 1) Input: conv_x as NCL + // + // src0->ne = { ncs, nr, n_s, 1 } // {L_in, C, N} + // Passing ACL_FORMAT_NCL here means: + // reversed dims -> [N, C, L_in] = [n_s, nr, ncs] + acl_tensor_ptr acl_x = ggml_cann_create_tensor(src0, src0->ne, src0->nb, 3, ACL_FORMAT_NCL); + + // 2) Weights: depthwise conv kernel, view src1 as {K, 1, C} + // + // src1 original: ne = { nc, nr, 1, 1 } // [K, C, 1, 1] + // we want a view: ne_w = { nc, 1, nr } // [K, 1, C] + // so that reversed dims -> [C, 1, K] which matches + // [out_channels, in_channels/groups, kernel_size] + int64_t w_ne[GGML_MAX_DIMS] = { nc, 1, nr, 1 }; // [K, 1 input ch. per group, C groups] + // Layout: src1 data is [K, C] with + // offset(k, c) = k*nb0 + c*nb1 + // We want offset_w(k, 0, c) = k*nb0 + c*nb1, + // so we can reuse nb0 and nb1, and set nb2 = nb1. + size_t w_nb[GGML_MAX_DIMS] = { src1->nb[0], src1->nb[1], src1->nb[1], src1->nb[3] }; // same as src1 + + acl_tensor_ptr acl_w = ggml_cann_create_tensor( + src1->data, ggml_cann_type_mapping(src1->type), ggml_type_size(src1->type), w_ne, w_nb, 3, ACL_FORMAT_NCL); + + // 3) Output: dst is { d_inner, n_t, n_s } (CLN) + // + // We need an NCL view of the same buffer: + // desired NCL logical shape: { L_out = n_t, C = nr, N = n_s } + // + // Original CLN layout: + // dst->ne = { nr, n_t, n_s } + // dst->nb[0] = sizeof(float) + // dst->nb[1] = nr * sizeof(float) + // dst->nb[2] = nr * n_t * sizeof(float) + // + // We want offset_new(L, C, N) = offset_orig(C, L, N). + // Choose: + // nb_y[0] = nr * sizeof(float); // step in L + // nb_y[1] = sizeof(float); // step in C + // nb_y[2] = nr * n_t * sizeof(float); // step in N + int64_t y_ne[GGML_MAX_DIMS] = { n_t, nr, n_s, 1 }; // [L_out, C, N] + size_t y_nb[GGML_MAX_DIMS] = { dst->ne[0] * sizeof(float), sizeof(float), dst->ne[0] * dst->ne[1] * sizeof(float), dst->nb[3] }; // [nr, 1, nr * n_t] + + acl_tensor_ptr acl_y = ggml_cann_create_tensor( + dst->data, ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type), y_ne, y_nb, 3, ACL_FORMAT_NCL); + + // --- Conv1d parameters: depthwise, stride 1, no padding ("valid") --- + int64_t strideVal[1] = { 1 }; + int64_t paddingVal[1] = { 0 }; + int64_t dilationVal[1] = { 1 }; + + acl_int_array_ptr stride = ggml_cann_create_int_array(strideVal, 1); + acl_int_array_ptr padding = ggml_cann_create_int_array(paddingVal, 1); + acl_int_array_ptr dilation = ggml_cann_create_int_array(dilationVal, 1); + + const bool transposed = false; + const int64_t groups = nr; // depthwise: one group per inner dim + int8_t cubeMathType = 0; + +#ifdef ASCEND_310P + cubeMathType = 1; +#endif + + GGML_CANN_CALL_ACLNN_OP(ctx, + Convolution, + acl_x.get(), // input: N, C, L_in = ncs + acl_w.get(), // weight: [C, 1, K] with groups=nr + nullptr, // bias + stride.get(), + padding.get(), + dilation.get(), + transposed, + padding.get(), // output padding (unused for non-transposed) + groups, + acl_y.get(), + cubeMathType); +} + diff --git a/ggml/src/ggml-cann/aclnn_ops.h b/ggml/src/ggml-cann/aclnn_ops.h index 6ec464028..a6ea016c5 100644 --- a/ggml/src/ggml-cann/aclnn_ops.h +++ b/ggml/src/ggml-cann/aclnn_ops.h @@ -1033,6 +1033,8 @@ void ggml_cann_op_unary(std::function Date: Thu, 25 Dec 2025 21:34:30 -0500 Subject: [PATCH 09/17] docs: added note for pre SYCL Intel hardware (#18016) Specify that it's for pre sycl hardware --- docs/backend/OPENCL.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/backend/OPENCL.md b/docs/backend/OPENCL.md index e52baffdf..ce6c7b560 100644 --- a/docs/backend/OPENCL.md +++ b/docs/backend/OPENCL.md @@ -17,7 +17,7 @@ OpenCL (Open Computing Language) is an open, royalty-free standard for cross-pla ### Llama.cpp + OpenCL -The llama.cpp OpenCL backend is designed to enable llama.cpp on **Qualcomm Adreno GPU** firstly via OpenCL. Thanks to the portabilty of OpenCL, the OpenCL backend can also run on certain Intel GPUs although the performance is not optimal. +The llama.cpp OpenCL backend is designed to enable llama.cpp on **Qualcomm Adreno GPU** firstly via OpenCL. Thanks to the portabilty of OpenCL, the OpenCL backend can also run on certain Intel GPUs such as those that do not have [SYCL](/docs/backend/SYCL.md) support although the performance is not optimal. ## OS From 4893cc07bba09525d6a1720d0686ed09b5a9b1c8 Mon Sep 17 00:00:00 2001 From: o7si <32285332+o7si@users.noreply.github.com> Date: Fri, 26 Dec 2025 23:35:29 +0800 Subject: [PATCH 10/17] server : fix crash when seq_rm fails for hybrid/recurrent models (#18391) * server : fix crash when seq_rm fails for hybrid/recurrent models * server : add allow_processing param to clear_slot --- tools/server/server-context.cpp | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp index 94825dc86..1abbf6d6d 100644 --- a/tools/server/server-context.cpp +++ b/tools/server/server-context.cpp @@ -1007,8 +1007,10 @@ private: return ret; } - void clear_slot(server_slot & slot) const { - GGML_ASSERT(!slot.is_processing()); + void clear_slot(server_slot & slot, bool allow_processing = false) const { + if (!allow_processing) { + GGML_ASSERT(!slot.is_processing()); + } SLT_WRN(slot, "clearing slot with %zu tokens\n", slot.prompt.tokens.size()); @@ -2336,7 +2338,7 @@ private: if (!llama_memory_seq_rm(llama_get_memory(ctx), slot.id, p0, -1)) { SLT_WRN(slot, "failed to truncate tokens with position >= %d - clearing the memory\n", p0); - clear_slot(slot); + clear_slot(slot, /*allow_processing=*/true); // there is no common part left slot.n_prompt_tokens_cache = 0; From 10dc500bdb043e16d7bb4e16bc4ee8480fad5262 Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Fri, 26 Dec 2025 09:53:46 -0600 Subject: [PATCH 11/17] vulkan: handle rope with large number of rows (#18306) --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 17 ++++++++++++++--- .../ggml-vulkan/vulkan-shaders/rope_multi.comp | 5 ++++- .../ggml-vulkan/vulkan-shaders/rope_neox.comp | 5 ++++- .../ggml-vulkan/vulkan-shaders/rope_norm.comp | 5 ++++- .../ggml-vulkan/vulkan-shaders/rope_params.glsl | 1 + .../ggml-vulkan/vulkan-shaders/rope_vision.comp | 5 ++++- tests/test-backend-ops.cpp | 3 +++ 7 files changed, 34 insertions(+), 7 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 1459b2608..e7ce518fb 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -1192,6 +1192,7 @@ struct vk_op_diag_mask_push_constants { struct vk_op_rope_push_constants { uint32_t rope_mode; uint32_t ncols; + uint32_t nrows; uint32_t n_dims; float freq_scale; uint32_t p_delta_rows; @@ -9090,10 +9091,20 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co elements = { num_groups * (uint32_t)src0->ne[3], 1, 1 }; } break; case GGML_OP_DIAG_MASK_INF: - case GGML_OP_ROPE: - case GGML_OP_ROPE_BACK: elements = { (uint32_t)ggml_nrows(src0), (uint32_t)ne00, 1 }; break; + case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: + { + uint32_t nrows = (uint32_t)ggml_nrows(src0); + uint32_t z = 1; + if (nrows > ctx->device->properties.limits.maxComputeWorkGroupCount[0]) { + z = CEIL_DIV(nrows, 32768); + nrows = 32768; + } + elements = { nrows, (uint32_t)ne00, z }; + + } break; case GGML_OP_GET_ROWS: elements = { (uint32_t)ne00, (uint32_t)ne10, (uint32_t)(ne11 * ne12) }; elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]); @@ -10021,7 +10032,7 @@ static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor * uint32_t nb02 = src0->nb[2] / ggml_type_size(src0->type); vk_op_rope_push_constants rope { - (uint32_t)mode, (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1], + (uint32_t)mode, (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale, has_ff, (uint32_t)src0->ne[2], nb01, nb02, { sections[0], sections[1], sections[2], sections[3] }, is_imrope, backprop, set_rows_stride, diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp index 7c1fb1cd2..f7587468a 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp @@ -6,6 +6,9 @@ void main() { const uint i0 = 2*gl_GlobalInvocationID.y; // i1 is actually i2*nb2+i1, but the rows are contiguous - const uint i1 = gl_GlobalInvocationID.x; + const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z; + if (i1 >= pc.nrows) { + return; + } rope_multi(i0, i1, pc); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp index 68f00c180..acb8ed781 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp @@ -6,6 +6,9 @@ void main() { const uint i0 = 2*gl_GlobalInvocationID.y; // i1 is actually i2*nb2+i1, but the rows are contiguous - const uint i1 = gl_GlobalInvocationID.x; + const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z; + if (i1 >= pc.nrows) { + return; + } rope_neox(i0, i1, pc); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp index 28a939ec6..0033cdb22 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp @@ -6,6 +6,9 @@ void main() { const uint i0 = 2*gl_GlobalInvocationID.y; // i1 is actually i2*nb2+i1, but the rows are contiguous - const uint i1 = gl_GlobalInvocationID.x; + const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z; + if (i1 >= pc.nrows) { + return; + } rope_norm(i0, i1, pc); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl index 82f39cee3..939cf3c51 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl @@ -6,6 +6,7 @@ struct rope_params { uint rope_mode; uint ncols; + uint nrows; uint n_dims; float freq_scale; uint p_delta_rows; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp index ea1e0fdb4..d93800b5e 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp @@ -6,6 +6,9 @@ void main() { const uint i0 = 2*gl_GlobalInvocationID.y; // i1 is actually i2*nb2+i1, but the rows are contiguous - const uint i1 = gl_GlobalInvocationID.x; + const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z; + if (i1 >= pc.nrows) { + return; + } rope_vision(i0, i1, pc); } diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 6b65f6e1c..a801455f1 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -7775,6 +7775,7 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_rope(type, {128, 40, 2, 1}, 128, GGML_ROPE_TYPE_NORMAL, 512, fs, ef, af, ff, v, fw)); // llama 13B test_cases.emplace_back(new test_rope(type, {128, 52, 2, 1}, 128, GGML_ROPE_TYPE_NORMAL, 512, fs, ef, af, ff, v, fw)); // llama 30B test_cases.emplace_back(new test_rope(type, {128, 64, 2, 1}, 128, GGML_ROPE_TYPE_NORMAL, 512, fs, ef, af, ff, v, fw)); // llama 65B + test_cases.emplace_back(new test_rope(type, {16, 16, 8192, 1}, 16, GGML_ROPE_TYPE_NORMAL, 512, fs, ef, af, ff, v, fw)); } if (all) { @@ -7789,6 +7790,7 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_rope(type, { 80, 32, 2, 1}, 20, GGML_ROPE_TYPE_NEOX, 512, fs, ef, af, ff, v, fw)); // neox (stablelm) test_cases.emplace_back(new test_rope(type, { 80, 32, 2, 1}, 32, GGML_ROPE_TYPE_NEOX, 512, fs, ef, af, ff, v, fw)); // neox (phi-2) test_cases.emplace_back(new test_rope(type, { 80, 32, 4, 1}, 32, GGML_ROPE_TYPE_NEOX, 512, fs, ef, af, ff, v, fw)); // neox (phi-2) + test_cases.emplace_back(new test_rope(type, { 16, 16, 8192, 1}, 16, GGML_ROPE_TYPE_NEOX, 512, fs, ef, af, ff, v, fw)); } if (all) { @@ -7802,6 +7804,7 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_rope(type, {128, 28, 2, 1}, 32, GGML_ROPE_TYPE_IMROPE, 512, fs, ef, af, ff, v, fw)); test_cases.emplace_back(new test_rope(type, { 80, 16, 2, 1}, 80, GGML_ROPE_TYPE_VISION, 512, fs, ef, af, ff, v, fw)); // rope_multi,m-rope (qwen2vl ViT) test_cases.emplace_back(new test_rope(type, {128, 16, 2, 1}, 128, GGML_ROPE_TYPE_IMROPE, 512, fs, ef, af, ff, v, fw)); // rope_multi,m-rope (qwen3vl) + test_cases.emplace_back(new test_rope(type, {16, 16, 8192, 1}, 16, GGML_ROPE_TYPE_IMROPE, 512, fs, ef, af, ff, v, fw)); } test_cases.emplace_back(new test_rope(type, { 64, 128, 2, 1}, 64, GGML_ROPE_TYPE_NEOX, 512, fs, ef, af, ff, v, fw)); // neox (falcon 40B) From b96b82fc852b83ea6f58ffcdc264308f60f79211 Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Fri, 26 Dec 2025 10:00:57 -0600 Subject: [PATCH 12/17] vulkan: Support UPSCALE w/antialias (#18327) --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 14 +++-- .../ggml-vulkan/vulkan-shaders/upscale.comp | 54 +++++++++++++++++++ tests/test-backend-ops.cpp | 32 ++++++----- 3 files changed, 84 insertions(+), 16 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index e7ce518fb..7ae2e3835 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -651,7 +651,7 @@ struct vk_device_struct { vk_pipeline pipeline_add_id_f32; vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32; - vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32; + vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32, pipeline_upscale_bilinear_antialias_f32; vk_pipeline pipeline_scale_f32; vk_pipeline pipeline_sqr_f32; vk_pipeline pipeline_sqrt_f32; @@ -3956,6 +3956,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_upscale_nearest_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_NEAREST}, 1); ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR}, 1); ggml_vk_create_pipeline(device, device->pipeline_upscale_bicubic_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BICUBIC}, 1); + ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_antialias_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ANTIALIAS}, 1); ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); @@ -8433,7 +8434,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return nullptr; case GGML_OP_UPSCALE: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - ggml_scale_mode mode = (ggml_scale_mode)(ggml_get_op_params_i32(dst, 0) & 0xFF); + uint32_t mode = (ggml_get_op_params_i32(dst, 0) & (0xFF | GGML_SCALE_FLAG_ANTIALIAS)); switch (mode) { case GGML_SCALE_MODE_NEAREST: return ctx->device->pipeline_upscale_nearest_f32; @@ -8441,6 +8442,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_upscale_bilinear_f32; case GGML_SCALE_MODE_BICUBIC: return ctx->device->pipeline_upscale_bicubic_f32; + case GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ANTIALIAS: + return ctx->device->pipeline_upscale_bilinear_antialias_f32; default: return nullptr; } @@ -14341,7 +14344,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm } return true; case GGML_OP_UPSCALE: - return op->src[0]->type == GGML_TYPE_F32 && !(op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS); + if (op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS) { + if ((op->op_params[0] & 0xFF) != GGML_SCALE_MODE_BILINEAR) { + return false; + } + } + return op->src[0]->type == GGML_TYPE_F32; case GGML_OP_ACC: return op->src[0]->type == GGML_TYPE_F32; case GGML_OP_CONCAT: diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp b/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp index 037ab0c78..f7d12a8dd 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp @@ -21,6 +21,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; #define NEAREST 0 #define BILINEAR 1 #define BICUBIC 2 +#define BILINEAR_ANTIALIAS 513 layout (constant_id = 0) const uint scale_mode = 0; @@ -62,6 +63,56 @@ float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) { return fetch_bilinear(c0, c1, d, i12, i13); } +float triangle_filter(float x) { + return max(1.0f - abs(x), 0.0f); +} + +float interpolate_bilinear_antialias(uint i10, uint i11, uint i12, uint i13) { + const float support1 = max(1.0f, 1.0f / p.sf1); + const float invscale1 = 1.0f / support1; + const float support0 = max(1.0f, 1.0f / p.sf0); + const float invscale0 = 1.0f / support0; + + const uint i02 = uint(i12 / p.sf2); + const uint i03 = uint(i13 / p.sf3); + + const float y = (float(i11) + p.pixel_offset) / p.sf1; + const float x = (float(i10) + p.pixel_offset) / p.sf0; + + // the range of source pixels that contribute + const int x_min = max(int(x - support0 + p.pixel_offset), 0); + const int x_max = min(int(x + support0 + p.pixel_offset), int(p.ne00)); + const int y_min = max(int(y - support1 + p.pixel_offset), 0); + const int y_max = min(int(y + support1 + p.pixel_offset), int(p.ne01)); + + // bilinear filter with antialiasing + float val = 0.0f; + float total_weight = 0.0f; + + for (int sy = y_min; sy < y_max; sy++) { + const float weight_y = triangle_filter((sy - y + p.pixel_offset) * invscale1); + + for (int sx = x_min; sx < x_max; sx++) { + const float weight_x = triangle_filter((sx - x + p.pixel_offset) * invscale0); + const float weight = weight_x * weight_y; + + if (weight <= 0.0f) { + continue; + } + + const float pixel = data_a[p.a_offset + i03 * p.nb03 + i02 * p.nb02 + sy * p.nb01 + sx * p.nb00]; + val += pixel * weight; + total_weight += weight; + } + } + + if (total_weight > 0.0f) { + val /= total_weight; + } + + return val; +} + // Bicubic interpolation with alpha = -0.75 // https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm const vec4 bcoeffs1 = vec4( 1.25, -2.25, 0.0, 1.0); @@ -118,6 +169,9 @@ void main() { case BICUBIC: result = interpolate_bicubic(i10, i11, i12, i13); break; + case BILINEAR_ANTIALIAS: + result = interpolate_bilinear_antialias(i10, i11, i12, i13); + break; } data_d[p.d_offset + idx] = D_TYPE(result); diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index a801455f1..0b981b178 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -402,12 +402,20 @@ static std::string var_to_str(ggml_op_pool pool) { } static std::string var_to_str(ggml_scale_mode mode) { - switch (mode) { - case GGML_SCALE_MODE_NEAREST: return "nearest"; - case GGML_SCALE_MODE_BILINEAR: return "bilinear"; - case GGML_SCALE_MODE_BICUBIC: return "bicubic"; - default: return std::to_string(mode); + std::string str; + switch (mode & 0xFF) { + case GGML_SCALE_MODE_NEAREST: str = "nearest"; break; + case GGML_SCALE_MODE_BILINEAR: str = "bilinear"; break; + case GGML_SCALE_MODE_BICUBIC: str = "bicubic"; break; + default: str = std::to_string(mode); break; } + if (mode & GGML_SCALE_FLAG_ALIGN_CORNERS) { + str += "|align_corners"; + } + if (mode & GGML_SCALE_FLAG_ANTIALIAS) { + str += "|antialias"; + } + return str; } #define VAR_TO_STR(x) (#x "=" + var_to_str(x)) @@ -5535,18 +5543,16 @@ struct test_interpolate : public test_case { const ggml_type type; const std::array ne; const std::array ne_tgt; - const uint32_t mode = GGML_SCALE_MODE_NEAREST; + const ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST; std::string vars() override { - ggml_scale_mode mode = (ggml_scale_mode)(this->mode & 0xFF); - std::string flags = (this->mode & GGML_SCALE_FLAG_ALIGN_CORNERS) ? "align_corners" : "none"; - return VARS_TO_STR5(type, ne, ne_tgt, mode, flags); + return VARS_TO_STR4(type, ne, ne_tgt, mode); } test_interpolate(ggml_type type = GGML_TYPE_F32, std::array ne = {2, 5, 7, 11}, std::array ne_tgt = {5, 7, 11, 13}, - uint32_t mode = GGML_SCALE_MODE_NEAREST) + ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST) : type(type), ne(ne), ne_tgt(ne_tgt), mode(mode) {} ggml_tensor * build_graph(ggml_context * ctx) override { @@ -7883,9 +7889,9 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {5, 7, 11, 13}, {2, 5, 7, 11}, mode)); } for (ggml_scale_mode mode : {GGML_SCALE_MODE_BILINEAR, GGML_SCALE_MODE_BICUBIC}) { - test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {2, 5, 7, 11}, {5, 7, 11, 13}, mode | GGML_SCALE_FLAG_ALIGN_CORNERS)); - test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {1, 4, 3, 2}, {2, 8, 3, 2}, mode | GGML_SCALE_FLAG_ALIGN_CORNERS)); - test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {4, 1, 3, 2}, {1, 1, 3, 2}, mode | GGML_SCALE_FLAG_ALIGN_CORNERS)); + test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {2, 5, 7, 11}, {5, 7, 11, 13}, (ggml_scale_mode)(mode | GGML_SCALE_FLAG_ALIGN_CORNERS))); + test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {1, 4, 3, 2}, {2, 8, 3, 2}, (ggml_scale_mode)(mode | GGML_SCALE_FLAG_ALIGN_CORNERS))); + test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {4, 1, 3, 2}, {1, 1, 3, 2}, (ggml_scale_mode)(mode | GGML_SCALE_FLAG_ALIGN_CORNERS))); } test_cases.emplace_back(new test_sum()); From cb999704fb5c5dc1763f2e692887229c809e14d7 Mon Sep 17 00:00:00 2001 From: Eve <139727413+netrunnereve@users.noreply.github.com> Date: Fri, 26 Dec 2025 17:12:11 +0000 Subject: [PATCH 13/17] vulkan: small dequantization improvements (#18380) * iq4_xs * quants --- .../vulkan-shaders/dequant_funcs.glsl | 8 +---- .../vulkan-shaders/mul_mm_funcs.glsl | 35 ++++++++++--------- .../src/ggml-vulkan/vulkan-shaders/types.glsl | 28 +++++++++++---- 3 files changed, 42 insertions(+), 29 deletions(-) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl index 70ee542d9..376944f1e 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl @@ -401,13 +401,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) { const uint sl = (data_a[a_offset + ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF; const uint sh = (data_a[a_offset + ib].scales_h >> (2 * ib32)) & 3; const uint qshift = (iqs & 16) >> 2; - u8vec4 qs = u8vec4( - data_a[a_offset + ib].qs[iq + 0], - data_a[a_offset + ib].qs[iq + 1], - data_a[a_offset + ib].qs[iq + 2], - data_a[a_offset + ib].qs[iq + 3] - ); - qs = (qs >> qshift) & uint8_t(0xF); + const u8vec4 qs = unpack8((data_a_packed32[a_offset + ib].qs[iq/4] >> qshift) & 0x0F0F0F0F); const float dl = float(int(sl | (sh << 4)) - 32); return dl * vec4( diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl index 58ede0440..1a3531761 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl @@ -159,14 +159,16 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin const uint is = iqs / 8; // 0..15 const uint halfsplit = ((iqs % 64) / 16); // 0,1,2,3 const uint qsshift = halfsplit * 2; // 0,2,4,6 - const uint m = 1 << (4 * n + halfsplit); // 1,2,4,8,16,32,64,128 const int8_t us = int8_t(((data_a[ib].scales[is % 8] >> (4 * int(is / 8))) & 0xF) | (((data_a[ib].scales[8 + (is % 4)] >> (2 * int(is / 4))) & 3) << 4)); const float dl = float(data_a[ib].d) * float(us - 32); - buf_a[buf_idx] = FLOAT_TYPE_VEC2(dl * float(int8_t((data_a[ib].qs[qsi ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi ] & m) != 0) ? 0 : 4)), - dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4))); + const vec2 qs = vec2(unpack8((uint(data_a_packed16[ib].qs[qsi / 2]) >> qsshift) & 0x0303).xy); + const vec2 hm = vec2(unpack8(((uint(data_a_packed16[ib].hmask[hmi / 2]) >> (4 * n + halfsplit)) & 0x0101 ^ 0x0101) << 2).xy); + + buf_a[buf_idx] = FLOAT_TYPE_VEC2(dl * (qs.x - hm.x), + dl * (qs.y - hm.y)); #elif defined(DATA_A_Q4_K) const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row; const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2; @@ -198,8 +200,10 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin const float d = loadd.x * sc; const float m = -loadd.y * mbyte; - buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, float((data_a[ib].qs[qsi ] >> (b * 4)) & 0xF), m), - fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m)); + const vec2 q = vec2(unpack8((uint(data_a_packed16[ib].qs[qsi / 2]) >> (b * 4)) & 0x0F0F).xy); + + buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, q.x, m), + fma(d, q.y, m)); #elif defined(DATA_A_Q5_K) const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row; const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2; @@ -213,8 +217,6 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin const uint qsi = n * 32 + (iqs % 16) * 2; // 0,2,4..126 const uint qhi = (iqs % 16) * 2; // 0,2,4..30 - const uint8_t hm = uint8_t(1 << (iqs / 16)); - const vec2 loadd = vec2(data_a[ib].dm); const uint scidx0 = (is < 4) ? is : (is + 4); @@ -234,8 +236,12 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin const float d = loadd.x * sc; const float m = -loadd.y * mbyte; - buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, float((data_a[ib].qs[qsi ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi ] & hm) != 0 ? 16 : 0), m), - fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m)); + const uint qs = (uint(data_a_packed16[ib].qs[qsi / 2]) >> (b * 4)) & 0x0F0F; + const uint qh = ((uint(data_a_packed16[ib].qh[qhi / 2]) >> (iqs / 16)) & 0x0101) << 4; + const vec2 q = vec2(unpack8(qs | qh).xy); + + buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, q.x, m), + fma(d, q.y, m)); #elif defined(DATA_A_Q6_K) const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row; const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2; @@ -394,11 +400,9 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin const float d = float(data_a[ib].d); const uint qs = data_a[ib].qs[iqs]; - const uint signs = pack32(u8vec4( - data_a[ib].qs[is+0], - data_a[ib].qs[is+1], - data_a[ib].qs[is+2], - data_a[ib].qs[is+3] + const uint signs = pack32(u16vec2( + data_a_packed16[ib].qs[is/2], + data_a_packed16[ib].qs[is/2+1] )); const float db = d * 0.5 * (0.5 + (signs >> 28)); const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7); @@ -443,8 +447,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin const uint sl = (data_a[ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF; const uint sh = ((data_a[ib].scales_h) >> (2 * ib32)) & 3; const uint qshift = (idx & 8) >> 1; - u8vec2 qs = u8vec2(data_a[ib].qs[iq], data_a[ib].qs[iq + 1]); - qs = (qs >> qshift) & uint8_t(0xF); + u8vec2 qs = unpack8((uint(data_a_packed16[ib].qs[iq/2]) >> qshift) & 0x0F0F).xy; const float d = float(data_a[ib].d); const vec2 v = d * float(int(sl | (sh << 4)) - 32) * vec2(kvalues_iq4nl[qs.x], kvalues_iq4nl[qs.y]); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl index 02578c77c..402a2a839 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl @@ -172,16 +172,12 @@ struct block_q8_0 float16_t d; int8_t qs[32]; }; + struct block_q8_0_packed16 { float16_t d; int16_t qs[32/2]; }; -struct block_q8_0_packed32 -{ - float16_t d; - int32_t qs[32/4]; -}; #if defined(DATA_A_Q8_0) #define QUANT_K QUANT_K_Q8_0 @@ -189,7 +185,6 @@ struct block_q8_0_packed32 #define QUANT_AUXF 1 #define A_TYPE block_q8_0 #define A_TYPE_PACKED16 block_q8_0_packed16 -#define A_TYPE_PACKED32 block_q8_0_packed32 #define DATA_A_QUANT_LEGACY #endif @@ -201,11 +196,13 @@ struct block_q8_1 f16vec2 ds; int8_t qs[32]; }; + struct block_q8_1_packed16 { f16vec2 ds; int16_t qs[16]; }; + struct block_q8_1_packed32 { f16vec2 ds; @@ -218,6 +215,7 @@ struct block_q8_1_x4 f16vec2 ds[4]; int32_t qs[32]; }; + struct block_q8_1_x4_packed128 { f16vec2 ds[4]; @@ -1346,10 +1344,28 @@ struct block_iq4_xs uint8_t qs[QUANT_K_IQ4_XS/2]; }; +struct block_iq4_xs_packed16 +{ + float16_t d; + uint16_t scales_h; + uint16_t scales_l[QUANT_K_IQ4_XS/128]; + uint16_t qs[QUANT_K_IQ4_XS/4]; +}; + +struct block_iq4_xs_packed32 +{ + float16_t d; + uint16_t scales_h; + uint32_t scales_l; + uint32_t qs[QUANT_K_IQ4_XS/8]; +}; + #if defined(DATA_A_IQ4_XS) #define QUANT_K QUANT_K_IQ4_XS #define QUANT_R QUANT_R_IQ4_XS #define A_TYPE block_iq4_xs +#define A_TYPE_PACKED16 block_iq4_xs_packed16 +#define A_TYPE_PACKED32 block_iq4_xs_packed32 #endif #define QUANT_K_IQ4_NL 32 From 9bf20d8ac3dc05dc23e6890144816b4becc85f9d Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Fri, 26 Dec 2025 11:15:02 -0600 Subject: [PATCH 14/17] vulkan: Use BK=32 for coopmat2 mul_mat_id (#18332) --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 7ae2e3835..35a1f19fe 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -1565,7 +1565,7 @@ class vk_perf_logger { total_op_times += time; } std::cerr << t.first << ": " << t.second.size() << " x " << (total_op_times / t.second.size() / 1000.0) - << " us"; + << " us = " << (total_op_times / 1000.0) << " us"; // If we have as many flops entries as timing entries for the op, then compute and log the flops/S. auto it = flops.find(t.first); @@ -2830,9 +2830,9 @@ static void ggml_vk_load_shaders(vk_device& device) { s_mmq_wg_denoms_k = { 32, 64, 1 }; // spec constants and tile sizes for quant matmul_id - l_warptile_mmqid = { 256, 128, 128, 16, 1, device->subgroup_size }; - m_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size }; - s_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size }; + l_warptile_mmqid = { 256, 128, 128, 32, 1, device->subgroup_size }; + m_warptile_mmqid = { 256, 128, 64, 32, 0, device->subgroup_size }; + s_warptile_mmqid = { 256, 128, 64, 32, 0, device->subgroup_size }; l_mmqid_wg_denoms = { 128, 128, 1 }; m_mmqid_wg_denoms = { 128, 64, 1 }; s_mmqid_wg_denoms = { 128, 64, 1 }; From 7ac8902133da6eb390c4d8368a7d252279123942 Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Fri, 26 Dec 2025 11:15:50 -0600 Subject: [PATCH 15/17] vulkan: optimize decodeFuncB in coopmat2 mul_mat_id shader (#18349) * vulkan: Use BK=32 for coopmat2 mul_mat_id * vulkan: optimize decodeFuncB in coopmat2 mul_mat_id shader Disable robustness, remove the OOB check in decodeFuncB, and initialize the row_ids to zero to avoid OOB access. Don't slice/offset the B matrix to ic * BN, only to adjust the coord back down to the range [0, BN) in decodeFuncB. Instead just slice with a row offset of zero and remove the '& (BN - 1)'. This allows the compiler to common some of the shared memory loads. --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 12 +++---- .../vulkan-shaders/mul_mm_cm2.comp | 31 +++++++++++-------- 2 files changed, 24 insertions(+), 19 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 35a1f19fe..c043368e7 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -3073,12 +3073,12 @@ static void ggml_vk_load_shaders(vk_device& device) { // Create 6 variants, {s,m,l}x{unaligned,aligned} #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, true); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, true); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, true); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, true); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, true); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, true); \ // Create 2 variants, {f16,f32} accumulator #define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp index 2e04baa44..293b25e9b 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp @@ -107,11 +107,7 @@ B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const i { const uint row_i = blockCoords[0]; - if (row_i >= _ne1) { - return B_TYPE(0.0); - } - - const u16vec4 row_idx = row_ids[row_i & (BN - 1)]; + const u16vec4 row_idx = row_ids[row_i]; B_TYPE ret = data_b[row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + blockCoords[1]]; return ret; @@ -194,12 +190,21 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) { #endif void main() { + const uint tid = gl_LocalInvocationIndex; +#ifdef MUL_MAT_ID + // initialize to row 0 so we don't need to bounds check + if (tid < BN) { + row_ids[tid] = u16vec4(0); + } +#if !defined(NEEDS_INIT_IQ_SHMEM) + barrier(); +#endif +#endif + #ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); #endif - const uint tid = gl_LocalInvocationIndex; - #ifdef MUL_MAT_ID const uint expert_idx = gl_GlobalInvocationID.z; #else @@ -482,7 +487,7 @@ void main() { coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB); sum = coopMatMulAdd(mat_a, mat_b, sum); } else { @@ -490,7 +495,7 @@ void main() { coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA); - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB); sum = coopMatMulAdd(mat_a, mat_b, sum); } @@ -526,7 +531,7 @@ void main() { coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB); sum = coopMatMulAdd(mat_a, mat_b, sum); } else { @@ -534,7 +539,7 @@ void main() { coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA); - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB); sum = coopMatMulAdd(mat_a, mat_b, sum); } @@ -571,7 +576,7 @@ void main() { coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); #ifdef MUL_MAT_ID - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose, decodeFuncB); #else coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose); #endif @@ -583,7 +588,7 @@ void main() { coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA); #ifdef MUL_MAT_ID - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose, decodeFuncB); #else coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose); #endif From c9ced4910ba0c8d95565950ac11bcc54fee309cd Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Fri, 26 Dec 2025 16:12:58 -0600 Subject: [PATCH 16/17] vulkan: preprocess mul_mat_id experts and discard workgroups more quickly (#18352) Run a preprocess to count how many times each expert is used, and use this to quickly discard workgroups that aren't needed. --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 377 ++++++++++-------- .../vulkan-shaders/count_experts.comp | 51 +++ .../ggml-vulkan/vulkan-shaders/mul_mm.comp | 14 +- .../vulkan-shaders/mul_mm_cm2.comp | 16 +- .../vulkan-shaders/mul_mm_id_funcs.glsl | 4 +- .../ggml-vulkan/vulkan-shaders/mul_mmq.comp | 14 +- .../vulkan-shaders/vulkan-shaders-gen.cpp | 2 + 7 files changed, 294 insertions(+), 184 deletions(-) create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index c043368e7..493ee9c9a 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -763,6 +763,7 @@ struct vk_device_struct { std::map pipeline_flash_attn_f32_f16[GGML_TYPE_COUNT]; vk_pipeline pipeline_flash_attn_split_k_reduce; + vk_pipeline pipeline_count_experts; // [2] is for whether to take n_experts from spec constant (0) or push constant (1) vk_pipeline pipeline_topk_moe[num_topk_moe_pipelines][TOPK_MOE_COUNT][2]; @@ -1004,6 +1005,14 @@ struct vk_op_push_constants { float param4; }; +struct vk_op_count_experts_push_constants { + uint32_t ne00; + uint32_t ne01; + uint32_t nb00; + uint32_t nb01; + uint32_t a_offset; +}; + struct vk_op_glu_push_constants { uint32_t N; uint32_t ne00; @@ -3068,6 +3077,8 @@ static void ggml_vk_load_shaders(vk_device& device) { #endif #undef CREATE_FA + const int mul_mat_id_param_count = 5; + #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) if (device->coopmat2) { @@ -3114,32 +3125,32 @@ static void ggml_vk_load_shaders(vk_device& device) { GGML_ASSERT(device->subgroup_ballot); - CREATE_MM2(pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) + CREATE_MM2(pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 5) #if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) if (device->coopmat_bf16_support) { - CREATE_MM(pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 5) } #endif - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) + CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5) #undef CREATE_MM #undef CREATE_MM2 } else @@ -3228,35 +3239,35 @@ static void ggml_vk_load_shaders(vk_device& device) { GGML_ASSERT(device->subgroup_ballot); - CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id); #if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) if (device->coopmat_bf16_support) { - CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id); } #endif - CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); + CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id); #undef CREATE_MM2 #undef CREATE_MM } else @@ -3341,91 +3352,91 @@ static void ggml_vk_load_shaders(vk_device& device) { #endif if (device->subgroup_ballot && device->subgroup_require_full_support && subgroup_min_size_16) { - CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16); - CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16); - CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16); - CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16); + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); - CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) if (device->integer_dot_product) { - CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); + CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); - CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); + CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); - CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16); - CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16); - CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16); - CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16); - CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16); + CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); + CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); + CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); + CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); + CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); } #endif } else { - CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id, 0); + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); - CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_iq1_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_iq1_m_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_iq2_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_iq2_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_iq3_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_iq4_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_mxfp4_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); + CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_iq1_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_iq1_m_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_iq2_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_iq2_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_iq3_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_iq4_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_mxfp4_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) if (device->integer_dot_product) { - CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_0], matmul_id_q4_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_1], matmul_id_q4_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_0], matmul_id_q5_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_1], matmul_id_q5_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q8_0], matmul_id_q8_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0); + CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_0], matmul_id_q4_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_1], matmul_id_q4_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_0], matmul_id_q5_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_1], matmul_id_q5_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q8_0], matmul_id_q8_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); - CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_MXFP4], matmul_id_mxfp4_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0); + CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_MXFP4], matmul_id_mxfp4_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); - CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q2_K], matmul_id_q2_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q3_K], matmul_id_q3_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_K], matmul_id_q4_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_K], matmul_id_q5_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q6_K], matmul_id_q6_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0); + CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q2_K], matmul_id_q2_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q3_K], matmul_id_q3_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_K], matmul_id_q4_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_K], matmul_id_q5_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q6_K], matmul_id_q6_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); } #endif } @@ -3502,57 +3513,57 @@ static void ggml_vk_load_shaders(vk_device& device) { #endif if (device->subgroup_ballot && device->subgroup_require_full_support && subgroup_min_size_16) { - CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16); - CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_subgroup_f16, , wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16); - CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_subgroup_f16_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16); - CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16); + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); + CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_subgroup_f16, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); + CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_subgroup_f16_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16); - CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_subgroup_q4_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_subgroup_q4_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_subgroup_q5_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_subgroup_q5_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_subgroup_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_subgroup_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_subgroup_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_subgroup_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_subgroup_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_subgroup_q6_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f32acc, matmul_id_subgroup_iq1_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f32acc, matmul_id_subgroup_iq1_m_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_subgroup_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_subgroup_iq2_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_subgroup_iq2_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_subgroup_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_subgroup_iq3_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_subgroup_iq4_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_subgroup_iq4_nl_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); - CREATE_MM(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f32acc, matmul_id_subgroup_mxfp4_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_subgroup_q4_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_subgroup_q4_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_subgroup_q5_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_subgroup_q5_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_subgroup_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_subgroup_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_subgroup_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_subgroup_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_subgroup_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_subgroup_q6_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f32acc, matmul_id_subgroup_iq1_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f32acc, matmul_id_subgroup_iq1_m_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_subgroup_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_subgroup_iq2_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_subgroup_iq2_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_subgroup_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_subgroup_iq3_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_subgroup_iq4_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_subgroup_iq4_nl_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); + CREATE_MM(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f32acc, matmul_id_subgroup_mxfp4_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size); } else { - CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id, 0); + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); - CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f32acc, matmul_id_iq1_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f32acc, matmul_id_iq1_m_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_iq2_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_iq2_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); - CREATE_MM(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f32acc, matmul_id_mxfp4_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0); + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f32acc, matmul_id_iq1_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f32acc, matmul_id_iq1_m_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_iq2_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_iq2_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); + CREATE_MM(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f32acc, matmul_id_mxfp4_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); } } // reusing CREATE_MM from the fp32 path @@ -3571,7 +3582,7 @@ static void ggml_vk_load_shaders(vk_device& device) { s_wg_denoms = { 32, 32, 1 }; CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, , 0); - CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id, 0); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); } #undef CREATE_MM @@ -4128,6 +4139,8 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_count_equal_i32, "count_equal_i32", count_equal_i32_len, count_equal_i32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(device, device->pipeline_count_experts, "count_experts", count_experts_len, count_experts_data, "main", 2, sizeof(vk_op_count_experts_push_constants), {1, 1, 1}, {}, 1, true); + for (auto &s : device->pipeline_solve_tri_f32) { const vk_solve_tri_pipeline_state &state = s.first; @@ -6525,18 +6538,18 @@ static uint32_t ggml_vk_guess_matmul_id_pipeline_align(ggml_backend_vk_context * static void ggml_vk_matmul_id( ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline, - vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids, + vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids, const vk_subbuffer & expert_count_buf, uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d, uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11, uint32_t padded_n) { - VK_LOG_DEBUG("ggml_vk_matmul_id(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), ids: (" << ids.buffer->buffer << ", " << ids.offset << ", " << ids.size << "), " << + VK_LOG_DEBUG("ggml_vk_matmul_id(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), ids: (" << ids.buffer->buffer << ", " << ids.offset << ", " << ids.size << "), expert_count: (" << expert_count_buf.buffer->buffer << ", " << expert_count_buf.offset << ", " << expert_count_buf.size << "), " << "m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", " << "batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", " << "n_as: " << n_as << ", nei0: " << nei0 << ", nei1: " << nei1 << ", nbi1: " << nbi1 << ", ne11: " << ne11 << ")"); const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, nei0, nei1, nbi1, ne11, padded_n }; - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, pc, { m, nei1, n_as }); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids, expert_count_buf }, pc, { m, nei1, n_as }); } static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) { @@ -7519,6 +7532,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t nei0 = ids->ne[0]; const uint64_t nei1 = ids->ne[1]; + const uint32_t nbi0 = ids->nb[0]; const uint32_t nbi1 = ids->nb[1]; const uint32_t nbi2 = ids->nb[2]; @@ -7626,6 +7640,9 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& if (quantize_y) { to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1); } + vk_pipeline count_experts = ctx->device->pipeline_count_experts; + + uint32_t expert_count_size = sizeof(uint32_t) * n_as; { if ( @@ -7641,6 +7658,10 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& ctx->prealloc_size_y = y_sz; ggml_vk_preallocate_buffers(ctx, subctx); } + if (ctx->prealloc_size_split_k < expert_count_size) { + ctx->prealloc_size_split_k = expert_count_size; + ggml_vk_preallocate_buffers(ctx, subctx); + } // Request descriptor sets ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1); @@ -7653,6 +7674,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& if (quantize_y) { ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1); } + ggml_pipeline_request_descriptor_sets(ctx, count_experts, 1); } vk_buffer d_D = dst_buf_ctx->dev_buffer; @@ -7702,6 +7724,20 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& ggml_vk_sync_buffers(ctx, subctx); } } + // Count how many times each expert is used + vk_subbuffer expert_count_buf = ggml_vk_subbuffer(ctx, ctx->prealloc_split_k, 0); + if (ctx->prealloc_split_k_need_sync) { + ggml_vk_sync_buffers(ctx, subctx); + } + { + const std::vector pc = { (uint32_t)nei0, + (uint32_t)nei1, + (uint32_t)(nbi0 / ggml_type_size(ids->type)), + (uint32_t)(nbi1 / ggml_type_size(ids->type)), + (uint32_t)(get_misalign_bytes(ctx, ids) / ggml_type_size(ids->type)) }; + ggml_vk_dispatch_pipeline(ctx, subctx, count_experts, + { vk_subbuffer{ d_ids, ids_buf_offset, ids_sz }, expert_count_buf }, pc, { (uint32_t)n_as, 1, 1}); + } if (x_non_contig) { ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, ggml_vk_subbuffer(ctx, d_Qx, qx_buf_offset), ggml_vk_subbuffer(ctx, d_X, 0)); @@ -7709,7 +7745,6 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const std::vector pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) }; ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_X, 0, x_sz } }, pc, { (uint32_t)x_ne, 1, 1}); - ggml_vk_sync_buffers(ctx, subctx); } if (y_non_contig) { if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() || @@ -7733,6 +7768,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& ctx->prealloc_y_last_tensor_used = src1; } } + ggml_vk_sync_buffers(ctx, subctx); uint32_t stride_batch_x = ne00*ne01; uint32_t stride_batch_y = ne10*ne11; @@ -7749,7 +7785,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& ggml_vk_matmul_id( ctx, subctx, pipeline, { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, - { d_D, d_buf_offset, d_sz }, { d_ids, ids_buf_offset, ids_sz }, + { d_D, d_buf_offset, d_sz }, { d_ids, ids_buf_offset, ids_sz }, expert_count_buf, ne01, ne21, ne10, ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21, n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n @@ -7761,6 +7797,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& if (y_non_contig || quantize_y) { ctx->prealloc_y_need_sync = true; } + ctx->prealloc_split_k_need_sync = true; } static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp b/ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp new file mode 100644 index 000000000..ffc860869 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp @@ -0,0 +1,51 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable + +#include "types.glsl" + +layout (push_constant) uniform parameter +{ + uint32_t ne00; + uint32_t ne01; + uint32_t nb00; + uint32_t nb01; + uint32_t a_offset; +} p; + +#define BLOCK_SIZE 256 + +layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {uint data_a[];}; +layout (binding = 1) writeonly buffer D {uint data_d[];}; + +shared uint vals[BLOCK_SIZE]; + +void main() { + const uint expert_id = gl_WorkGroupID.x; + const uint num_elements = p.ne00 * p.ne01; + const uint tid = gl_LocalInvocationID.x; + + uint count = 0; + for (uint idx = tid; idx < num_elements; idx += BLOCK_SIZE) { + const uint i01 = idx / p.ne00; + const uint i00 = idx % p.ne00; + const uint a = data_a[p.a_offset + i01 * p.nb01 + i00 * p.nb00]; + + count += uint(a == expert_id); + } + + vals[tid] = count; + barrier(); + [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) { + vals[tid] += vals[tid + s]; + } + barrier(); + } + + if (tid == 0) { + data_d[expert_id] = vals[0]; + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp index 5c5251da3..c0c00d28f 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp @@ -68,6 +68,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; #ifdef MUL_MAT_ID layout (binding = 3) readonly buffer IDS {int data_ids[];}; +layout (binding = 4) readonly buffer Counts {int data_expert_count[];}; #endif layout (push_constant) uniform parameter @@ -135,13 +136,19 @@ shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS]; #include "mul_mm_funcs.glsl" void main() { + const uint ic = gl_WorkGroupID.y; + +#ifdef MUL_MAT_ID + const uint expert_idx = gl_GlobalInvocationID.z; + if (ic * BN >= data_expert_count[expert_idx]) { + return; + } +#endif #ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); #endif -#ifdef MUL_MAT_ID - const uint expert_idx = gl_GlobalInvocationID.z; -#else +#ifndef MUL_MAT_ID const uint batch_idx = gl_GlobalInvocationID.z; const uint i13 = batch_idx / p.ne12; @@ -156,7 +163,6 @@ void main() { const uint blocks_m = (p.M + BM - 1) / BM; const uint ir = gl_WorkGroupID.x % blocks_m; const uint ik = gl_WorkGroupID.x / blocks_m; - const uint ic = gl_WorkGroupID.y; const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER); const uint WSUBM = WM / WMITER; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp index 293b25e9b..d0d1d8ef7 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp @@ -92,6 +92,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; #ifdef MUL_MAT_ID layout (binding = 3) readonly buffer IDS {int data_ids[];}; +layout (binding = 4) readonly buffer Counts {int data_expert_count[];}; shared u16vec4 row_ids[BN]; @@ -134,6 +135,8 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) { uint ids[16]; uint iter = 0; + uint expert_count = data_expert_count[expert_idx]; + for (uint j = 0; j < num_elements; j += BLOCK_SIZE) { // prefetch up to 16 elements if (iter == 0) { @@ -181,7 +184,7 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) { } _ne1 += total; iter &= 15; - if (_ne1 >= (ic + 1) * BN) { + if (_ne1 >= (ic + 1) * BN || _ne1 == expert_count) { break; } } @@ -191,7 +194,13 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) { void main() { const uint tid = gl_LocalInvocationIndex; + const uint ic = gl_WorkGroupID.y; + #ifdef MUL_MAT_ID + const uint expert_idx = gl_GlobalInvocationID.z; + if (ic * BN >= data_expert_count[expert_idx]) { + return; + } // initialize to row 0 so we don't need to bounds check if (tid < BN) { row_ids[tid] = u16vec4(0); @@ -205,9 +214,7 @@ void main() { init_iq_shmem(gl_WorkGroupSize); #endif -#ifdef MUL_MAT_ID - const uint expert_idx = gl_GlobalInvocationID.z; -#else +#ifndef MUL_MAT_ID const uint batch_idx = gl_GlobalInvocationID.z; const uint i13 = batch_idx / p.ne12; @@ -222,7 +229,6 @@ void main() { const uint blocks_m = (p.M + BM - 1) / BM; const uint ir = gl_WorkGroupID.x % blocks_m; const uint ik = gl_WorkGroupID.x / blocks_m; - const uint ic = gl_WorkGroupID.y; #ifdef MUL_MAT_ID if (bitCount(p.nei0) == 1) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl index 1d0e84ac9..743004ff8 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl @@ -13,6 +13,8 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) { uint ids[16]; uint iter = 0; + uint expert_count = data_expert_count[expert_idx]; + for (uint j = 0; j < num_elements; j += BLOCK_SIZE) { // prefetch up to 16 elements if (iter == 0) { @@ -60,7 +62,7 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) { } _ne1 += total; iter &= 15; - if (_ne1 >= (ic + 1) * BN) { + if (_ne1 >= (ic + 1) * BN || _ne1 == expert_count) { break; } } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp index dc8b3df47..cd36e270a 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp @@ -35,6 +35,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; #ifdef MUL_MAT_ID layout (binding = 3) readonly buffer IDS {int data_ids[];}; +layout (binding = 4) readonly buffer Counts {int data_expert_count[];}; #endif layout (push_constant) uniform parameter @@ -104,13 +105,19 @@ block_b_cache cache_b; #include "mul_mmq_funcs.glsl" void main() { + const uint ic = gl_WorkGroupID.y; + +#ifdef MUL_MAT_ID + const uint expert_idx = gl_GlobalInvocationID.z; + if (ic * BN >= data_expert_count[expert_idx]) { + return; + } +#endif #ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); #endif -#ifdef MUL_MAT_ID - const uint expert_idx = gl_GlobalInvocationID.z; -#else +#ifndef MUL_MAT_ID const uint batch_idx = gl_GlobalInvocationID.z; const uint i13 = batch_idx / p.ne12; @@ -125,7 +132,6 @@ void main() { const uint blocks_m = (p.M + BM - 1) / BM; const uint ir = gl_WorkGroupID.x % blocks_m; const uint ik = gl_WorkGroupID.x / blocks_m; - const uint ic = gl_WorkGroupID.y; const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER); const uint WSUBM = WM / WMITER; 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 e237a8e10..4a8337837 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -945,6 +945,8 @@ void process_shaders() { string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}})); string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("count_experts", "count_experts.comp", merge_maps(base_dict, {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}})); + for (std::string dim_str : {"", "_3d"}) { for (bool bda : {false, true}) { std::string bda_str = bda ? "_bda" : ""; From 9045c9afe57513eef8d16e329042ea2d7afd9a95 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Sat, 27 Dec 2025 09:56:04 +0100 Subject: [PATCH 17/17] llama-fit-params: fix Gemma 3 calculation (#18372) --- src/llama.cpp | 45 +++++++++++++++++++++++++-------------------- 1 file changed, 25 insertions(+), 20 deletions(-) diff --git a/src/llama.cpp b/src/llama.cpp index 1e18637e3..3428b1bd3 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -181,12 +181,11 @@ static void llama_params_fit_impl( } } - int64_t sum_total = 0; + int64_t sum_free = 0; int64_t sum_projected_free = 0; int64_t min_projected_free = INT64_MAX; int64_t sum_projected_used = 0; int64_t sum_projected_model = 0; - int64_t sum_projected_ctx = 0; if (nd > 1) { LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__); @@ -197,12 +196,11 @@ static void llama_params_fit_impl( const int64_t projected_used = dmd.mb.total(); const int64_t projected_free = dmd.free - projected_used; - sum_total += dmd.total; + sum_free += dmd.free; sum_projected_used += projected_used; sum_projected_free += projected_free; min_projected_free = std::min(min_projected_free, projected_free); sum_projected_model += dmd.mb.model; - sum_projected_ctx += dmd.mb.context; if (nd > 1) { LLAMA_LOG_INFO("%s: - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " %s\n", @@ -210,10 +208,9 @@ static void llama_params_fit_impl( projected_free >= 0 ? "surplus" : "deficit"); } } - assert(sum_total >= 0 && sum_projected_used >= 0 && sum_projected_ctx >= 0); - assert(sum_projected_used >= sum_projected_ctx); + assert(sum_free >= 0 && sum_projected_used >= 0); LLAMA_LOG_INFO("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n", - __func__, sum_projected_used/MiB, sum_total/MiB); + __func__, sum_projected_used/MiB, sum_free/MiB); if (min_projected_free >= margin) { if (nd == 1) { LLAMA_LOG_INFO("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n", @@ -236,9 +233,7 @@ static void llama_params_fit_impl( __func__, margin/MiB, -global_surplus/MiB); if (cparams->n_ctx == 0) { if (hp_nct > n_ctx_min) { - const int64_t bytes_per_ctx = sum_projected_ctx / hp_nct; - - int64_t memory_reduction = -global_surplus; + int64_t sum_used_target = sum_free - nd*margin_s; if (nd > 1) { // for multiple devices we need to be more conservative in terms of how much context we think can fit: // - for dense models only whole layers can be assigned to devices @@ -246,24 +241,34 @@ static void llama_params_fit_impl( // - on average we expect a waste of 0.5 layers/tensors per device // - use slightly more than the expected average for nd devices to be safe const int64_t model_per_layer = sum_projected_model / std::min(uint32_t(mparams->n_gpu_layers), hp_ngl); - memory_reduction += (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6); + sum_used_target -= (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6); } - uint32_t ctx_reduction = std::min(uint32_t((memory_reduction + bytes_per_ctx - 1) / bytes_per_ctx), hp_nct - n_ctx_min); - cparams->n_ctx = hp_nct - ctx_reduction; - cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend + int64_t sum_projected_used_min_ctx = 0; + cparams->n_ctx = n_ctx_min; + const dmds_t dmds_min_ctx = llama_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level); + for (const auto & dmd : dmds_min_ctx) { + sum_projected_used_min_ctx += dmd.mb.total(); + } + if (sum_used_target > sum_projected_used_min_ctx) { + // linear interpolation between minimum and maximum context size: + cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx) + / (sum_projected_used - sum_projected_used_min_ctx); + cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend - ctx_reduction = hp_nct - cparams->n_ctx; - memory_reduction = ctx_reduction * bytes_per_ctx; - global_surplus += memory_reduction; - LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n", - __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB); - if (global_surplus >= 0) { + const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min); + const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx; + LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n", + __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB); if (nd == 1) { LLAMA_LOG_INFO("%s: entire model can be fit by reducing context\n", __func__); return; } LLAMA_LOG_INFO("%s: entire model should be fit across devices by reducing context\n", __func__); + } else { + const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx; + LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n", + __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB); } } else { LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",