vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 09:04:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

try fix oldpc cuda broken without flash attn since upstream pr14361 between 1.94 and 1.95 (+1 squashed commits)

Browse source 
Squashed commits:

[940f0c639] try fix oldpc cuda broken without flash attn since upstream pr14361 between 1.94 and 1.95
This commit is contained in:
Concedo 2025-08-10 00:09:16 +08:00
parent 4c1faf61b2
commit 487d509b44
2 changed files with 8 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

7
ggml/src/ggml-cuda/ggml-cuda.cu
View file

@ -2061,6 +2061,13 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
bool use_batched_cublas_bf16 = src0->type == GGML_TYPE_BF16 && bf16_mma_hardware_available(cc);
bool use_batched_cublas_f32 = src0->type == GGML_TYPE_F32;
if(ggml_cuda_highest_compiled_arch(cc) <= GGML_CUDA_CC_TURING)
{
//kcpp: https://github.com/ggml-org/llama.cpp/pull/14361 broke oldpc mode without this.
use_batched_cublas_bf16 = false;
use_batched_cublas_f32 = false;
}
if (!split && use_mul_mat_vec_f) {
// the custom F16 vector kernel can be used over batched cuBLAS GEMM
// but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)

2
koboldcpp.py
View file

@ -63,7 +63,7 @@ dry_seq_break_max = 128
extra_images_max = 4
# global vars
KcppVersion = "1.97.3"
KcppVersion = "1.97.4"
showdebug = True
kcpp_instance = None #global running instance
global_memory = {"tunnel_url": "", "restart_target":"", "input_to_exit":False, "load_complete":False, "restart_override_config_target":""}