-
Notifications
You must be signed in to change notification settings - Fork 1.3k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Split bwd into more .cu files to speed up compilation
- Loading branch information
Showing
89 changed files
with
304 additions
and
153 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 128, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim128<cutlass::bfloat16_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 128>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim128<cutlass::bfloat16_t>(params, stream); | ||
void run_mha_bwd_<cutlass::bfloat16_t, 128, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim128<cutlass::bfloat16_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 128, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim128<cutlass::half_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 128>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim128<cutlass::half_t>(params, stream); | ||
void run_mha_bwd_<cutlass::half_t, 128, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim128<cutlass::half_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 160, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim160<cutlass::bfloat16_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 160>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim160<cutlass::bfloat16_t>(params, stream); | ||
void run_mha_bwd_<cutlass::bfloat16_t, 160, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim160<cutlass::bfloat16_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 160, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim160<cutlass::half_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 160>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim160<cutlass::half_t>(params, stream); | ||
void run_mha_bwd_<cutlass::half_t, 160, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim160<cutlass::half_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 192, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim192<cutlass::bfloat16_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 192>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim192<cutlass::bfloat16_t>(params, stream); | ||
void run_mha_bwd_<cutlass::bfloat16_t, 192, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim192<cutlass::bfloat16_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 192, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim192<cutlass::half_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 192>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim192<cutlass::half_t>(params, stream); | ||
void run_mha_bwd_<cutlass::half_t, 192, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim192<cutlass::half_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 256, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim256<cutlass::bfloat16_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 256>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim256<cutlass::bfloat16_t>(params, stream); | ||
void run_mha_bwd_<cutlass::bfloat16_t, 256, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim256<cutlass::bfloat16_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 256, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim256<cutlass::half_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 256>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim256<cutlass::half_t>(params, stream); | ||
void run_mha_bwd_<cutlass::half_t, 256, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim256<cutlass::half_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 32, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim32<cutlass::bfloat16_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 32>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim32<cutlass::bfloat16_t>(params, stream); | ||
void run_mha_bwd_<cutlass::bfloat16_t, 32, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim32<cutlass::bfloat16_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 32, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim32<cutlass::half_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 32>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim32<cutlass::half_t>(params, stream); | ||
void run_mha_bwd_<cutlass::half_t, 32, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim32<cutlass::half_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 64, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim64<cutlass::bfloat16_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 64>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim64<cutlass::bfloat16_t>(params, stream); | ||
void run_mha_bwd_<cutlass::bfloat16_t, 64, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim64<cutlass::bfloat16_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 64, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim64<cutlass::half_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 64>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim64<cutlass::half_t>(params, stream); | ||
void run_mha_bwd_<cutlass::half_t, 64, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim64<cutlass::half_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 96, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim96<cutlass::bfloat16_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::bfloat16_t, 96>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim96<cutlass::bfloat16_t>(params, stream); | ||
void run_mha_bwd_<cutlass::bfloat16_t, 96, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim96<cutlass::bfloat16_t, false>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,10 @@ | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 96, true>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim96<cutlass::half_t, true>(params, stream); | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,10 +1,10 @@ | ||
// Copyright (c) 2023, Tri Dao. | ||
// Copyright (c) 2024, Tri Dao. | ||
// Splitting the different head dimensions to different files to speed up compilation. | ||
// This file is auto-generated. See "generate_kernels.py" | ||
|
||
#include "flash_bwd_launch_template.h" | ||
|
||
template<> | ||
void run_mha_bwd_<cutlass::half_t, 96>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim96<cutlass::half_t>(params, stream); | ||
void run_mha_bwd_<cutlass::half_t, 96, false>(Flash_bwd_params ¶ms, cudaStream_t stream) { | ||
run_mha_bwd_hdim96<cutlass::half_t, false>(params, stream); | ||
} |
Oops, something went wrong.