CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-12 10:47:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: mul_mat_id for mmf for bs <= 64 for f16 and bs <= 32 for f32 (#16277)

Browse Source 
* CUDA: mul_mat_id for mmf for bs <= 64 for f16 and bs <= 32 for f32

This commit adds mul_mat_id support for ncols_dst >= 16. It does this by
packing ncols_dst tiles into the blockDim.y.

My tests on a RTX 3090 show that this is faster than the cuBLAS fallback
for f16 till bs=64, and for f32 till bs=32

* Review: refactor if statement
This commit is contained in:
Aman Gupta
2025-09-28 00:49:32 +08:00
committed by GitHub
parent 75a3a6c2cd
commit c0bfc57af4
4 changed files with 92 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -2031,7 +2031,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
const int cc = ggml_cuda_info().devices[id].cc;
const int warp_size = ggml_cuda_info().devices[id].warp_size;
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
use_mul_mat_f = use_mul_mat_f && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1]);
use_mul_mat_f = use_mul_mat_f && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc);
}
@@ -2039,7 +2039,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
const int cc = ggml_cuda_info().devices[ctx.device].cc;
const int warp_size = ggml_cuda_info().devices[ctx.device].warp_size;
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
use_mul_mat_f = use_mul_mat_f && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1]);
use_mul_mat_f = use_mul_mat_f && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc);
}
@@ -2111,7 +2111,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
return;
}
if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src1->ne[2])) {
if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src1->ne[2], /*mul_mat_id=*/true)) {
ggml_cuda_mul_mat_f(ctx, src0, src1, ids, dst);
return;
}