vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

CUDA: refactor and optimize IQ MMVQ (#8215)

Browse source 
* CUDA: refactor and optimize IQ MMVQ

* uint -> uint32_t

* __dp4a -> ggml_cuda_dp4a

* remove MIN_CC_DP4A checks

* change default

* try CI fix
This commit is contained in:
Johannes Gäßler 2024-07-01 20:39:06 +02:00 • committed by GitHub
parent dae57a1ebc
commit cb5fad4c6c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 406 additions and 487 deletions
Download patch file Download diff file Expand all files Collapse all files

12
ggml/src/ggml-cuda.cu
View file

@ -1882,6 +1882,11 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
bool use_mul_mat_q = ggml_is_quantized(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
// if mmvq is available it's a better choice than dmmv:
#ifndef GGML_CUDA_FORCE_DMMV
use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
#endif // GGML_CUDA_FORCE_DMMV
bool any_gpus_with_slow_fp16 = false;
if (split) {
@ -1894,22 +1899,15 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
}
const int cc = ggml_cuda_info().devices[id].cc;
use_mul_mat_vec_q = use_mul_mat_vec_q && cc >= MIN_CC_DP4A;
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
}
} else {
const int cc = ggml_cuda_info().devices[ctx.device].cc;
use_mul_mat_vec_q = use_mul_mat_vec_q && cc >= MIN_CC_DP4A;
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
}
// if mmvq is available it's a better choice than dmmv:
#ifndef GGML_CUDA_FORCE_DMMV
use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
#endif // GGML_CUDA_FORCE_DMMV
// debug helpers
//printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
//printf(" %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);