Skip to content

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

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Sep 27, 2025
Merged

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

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8ee101c
CUDA: mul_mat_id for mmf for bs <= 64 for f16 and bs <= 32 for f32
am17an Sep 11, 2025
6aef41d
Review: refactor if statement
am17an Sep 27, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2031,15 +2031,15 @@ 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);
}
} else {
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);
}
Expand Down Expand Up @@ -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;
}
Expand Down
14 changes: 12 additions & 2 deletions ggml/src/ggml-cuda/mmf.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,7 +84,7 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
}
}

bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, const int src1_ncols) {
bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, const int src1_ncols, bool mul_mat_id) {

if (ggml_is_quantized(type)) {
return false;
Expand All @@ -96,7 +96,17 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
if (src0_ne[1] % MMF_ROWS_PER_BLOCK != 0) {
return false;
}
if (src1_ncols > 16) {

if (mul_mat_id) {
if (type == GGML_TYPE_F32 && src1_ncols > 32) {
return false;
}
if ((type == GGML_TYPE_F16 || type == GGML_TYPE_BF16) && src1_ncols > 64) {
return false;
}
}

if (!mul_mat_id && src1_ncols > 16) {
return false;
}

Expand Down
Loading
Loading