Skip to content

CUDA: skip masked out KQ slices in mma FA kernel #14735

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

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
+434 −166
Open

CUDA: skip masked out KQ slices in mma FA kernel #14735

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
164 changes: 104 additions & 60 deletions ggml/src/ggml-cuda/fattn-mma-f16.cuh
View file
Open in desktop

Large diffs are not rendered by default.

35 changes: 30 additions & 5 deletions ggml/src/ggml-cuda/fattn.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,29 +8,53 @@
#include "fattn-wmma-f16.cuh"
#include "fattn.cuh"

template <int DKQ, int DV, int ncols1, int ncols2>
static void ggml_cuda_flash_attn_ext_mma_f16_switch_nstages(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
typedef fattn_mma_f16_config<DKQ, DV> c;
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;

const ggml_tensor * Q = dst->src[0];
const ggml_tensor * mask = dst->src[3];

if (!cp_async_available(cc)) {
ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, ncols1, ncols2, 0>(ctx, dst);
return;
}

if constexpr (c::nstages_max > 1) {
static_assert(c::nstages_max == 2, "bad nstages_max");
if (Q->ne[3] == 1 && mask) {
ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, ncols1, ncols2, 2>(ctx, dst);
return;
}
}

ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, ncols1, ncols2, 1>(ctx, dst);
}

template <int DKQ, int DV, int ncols2>
static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
const ggml_tensor * Q = dst->src[0];

if constexpr (ncols2 <= 8) {
if (Q->ne[1] <= 8/ncols2) {
ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 8/ncols2, ncols2>(ctx, dst);
ggml_cuda_flash_attn_ext_mma_f16_switch_nstages<DKQ, DV, 8/ncols2, ncols2>(ctx, dst);
return;
}
}

if (Q->ne[1] <= 16/ncols2) {
ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
ggml_cuda_flash_attn_ext_mma_f16_switch_nstages<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
return;
}

if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING || Q->ne[1] <= 32/ncols2) {
ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 32/ncols2, ncols2>(ctx, dst);
ggml_cuda_flash_attn_ext_mma_f16_switch_nstages<DKQ, DV, 32/ncols2, ncols2>(ctx, dst);
return;
}

ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 64/ncols2, ncols2>(ctx, dst);
ggml_cuda_flash_attn_ext_mma_f16_switch_nstages<DKQ, DV, 64/ncols2, ncols2>(ctx, dst);
}

template <int DKQ, int DV>
Expand Down Expand Up @@ -325,7 +349,8 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst

const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion;
const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies &&
(Q->ne[3] > 1 || cc < GGML_CUDA_CC_ADA_LOVELACE) && !mma_needs_data_conversion;
const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
if (prec == GGML_PREC_DEFAULT) {
Expand Down
3 changes: 2 additions & 1 deletion ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,4 +2,5 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);
DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16, 0);
DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16, 1);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 1, 8);
DECL_FATTN_MMA_F16_CASE(80, 80, 1, 8);
DECL_FATTN_MMA_F16_CASE(96, 96, 1, 8);
DECL_FATTN_MMA_F16_CASE(112, 112, 1, 8);
DECL_FATTN_MMA_F16_CASE(128, 128, 1, 8);
DECL_FATTN_MMA_F16_CASE(256, 256, 1, 8);
DECL_FATTN_MMA_F16_CASE(64, 64, 1, 8, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 1, 8, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 1, 8, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 1, 8, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 1, 8, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 1, 8, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 1, 8, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 1, 8, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 1, 8, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 1, 8, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 1, 8, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 1, 8, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 1, 8, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 1, 8, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 1, 8, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 1, 8, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 1, 8, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 1, 8, 2);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 16, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 16, 1, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 16, 1, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 16, 1, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 1, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 1, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 1, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 1, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 1, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 1, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 1, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 1, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 1, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 1, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 1, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 1, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 1, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 1, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 1, 2);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 16, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 2);
DECL_FATTN_MMA_F16_CASE(64, 64, 16, 2, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 16, 2, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 16, 2, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 2, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 2, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 2, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 2, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 2, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 2, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 2, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 2, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 2, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 2, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 2, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 2, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 2, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 2, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 2, 2);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 16, 4);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 4);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);
DECL_FATTN_MMA_F16_CASE(64, 64, 16, 4, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 16, 4, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 16, 4, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 4, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 4, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 16, 4, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4, 2);
3 changes: 2 additions & 1 deletion ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,4 +2,5 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);
DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16, 0);
DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16, 1);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 2, 4);
DECL_FATTN_MMA_F16_CASE(80, 80, 2, 4);
DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);
DECL_FATTN_MMA_F16_CASE(64, 64, 2, 4, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 2, 4, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 2, 4, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 2, 4, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 2, 4, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 2, 4, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4, 2);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 2, 8);
DECL_FATTN_MMA_F16_CASE(80, 80, 2, 8);
DECL_FATTN_MMA_F16_CASE(96, 96, 2, 8);
DECL_FATTN_MMA_F16_CASE(112, 112, 2, 8);
DECL_FATTN_MMA_F16_CASE(128, 128, 2, 8);
DECL_FATTN_MMA_F16_CASE(256, 256, 2, 8);
DECL_FATTN_MMA_F16_CASE(64, 64, 2, 8, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 2, 8, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 2, 8, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 2, 8, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 2, 8, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 2, 8, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 2, 8, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 2, 8, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 2, 8, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 2, 8, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 2, 8, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 2, 8, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 2, 8, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 2, 8, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 2, 8, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 2, 8, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 2, 8, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 2, 8, 2);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 32, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 32, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 32, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 32, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 32, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 32, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 32, 1, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 32, 1, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 32, 1, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 32, 1, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 32, 1, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 32, 1, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 32, 1, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 32, 1, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 32, 1, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 32, 1, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 32, 1, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 32, 1, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 32, 1, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 32, 1, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 32, 1, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 32, 1, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 32, 1, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 32, 1, 2);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 32, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 32, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 32, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 32, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 32, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 32, 2);
DECL_FATTN_MMA_F16_CASE(64, 64, 32, 2, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 32, 2, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 32, 2, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 32, 2, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 32, 2, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 32, 2, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 32, 2, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 32, 2, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 32, 2, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 32, 2, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 32, 2, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 32, 2, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 32, 2, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 32, 2, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 32, 2, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 32, 2, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 32, 2, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 32, 2, 2);
3 changes: 2 additions & 1 deletion ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,4 +2,5 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);
DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16, 0);
DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16, 1);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 4, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 2);
DECL_FATTN_MMA_F16_CASE(64, 64, 4, 2, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 4, 2, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 4, 2, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 2, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 2, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 2, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 2, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 2, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 2, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 2, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 2, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 2, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 2, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 2, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 2, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 2, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 2, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 2, 2);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 4, 4);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 4);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);
DECL_FATTN_MMA_F16_CASE(64, 64, 4, 4, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 4, 4, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 4, 4, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 4, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 4, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 4, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4, 2);
24 changes: 18 additions & 6 deletions ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,21 @@

#include "../fattn-mma-f16.cuh"

DECL_FATTN_MMA_F16_CASE(64, 64, 4, 8);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 8);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 8);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 8);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 8);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 8);
DECL_FATTN_MMA_F16_CASE(64, 64, 4, 8, 0);
DECL_FATTN_MMA_F16_CASE(64, 64, 4, 8, 1);
DECL_FATTN_MMA_F16_CASE(64, 64, 4, 8, 2);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 8, 0);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 8, 1);
DECL_FATTN_MMA_F16_CASE(80, 80, 4, 8, 2);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 8, 0);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 8, 1);
DECL_FATTN_MMA_F16_CASE(96, 96, 4, 8, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 8, 0);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 8, 1);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 8, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 8, 0);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 8, 1);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 8, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 8, 0);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 8, 1);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 8, 2);
Loading
Loading