Use generate_kernels.py script from Driss Guessous

csrc/flash_attn/src/flash_bwd_hdim128_bf16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
-//     using elem_type = cutlass::bfloat16_t;
-//     if (params.h == params.h_k) {
-//         run_flash_bwd<Flash_bwd_kernel_traits<128, 64, 128, 8, 2, 4, 2, false, false, elem_type>>(params, stream, configure);
-//     } else {
-//         run_flash_bwd_seqq_parallel<Flash_bwd_kernel_traits<128, 128, 64, 8, 4, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     }
-// }
-
 template<>
 void run_mha_bwd_<cutlass::bfloat16_t, 128>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim128<cutlass::bfloat16_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim128_fp16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
-//     using elem_type = cutlass::half_t;
-//     if (params.h == params.h_k) {
-//     // run_flash_bwd<Flash_bwd_kernel_traits<128, 32, 128, 8, 2, 2, 2, false, false, elem_type>>(params, stream, configure);
-//     // This is faster, in the case of sequence-parallel bwd (where we need fewer registers).
-//     // Out of these three, the 2nd one is slightly faster (2% faster than the first). Idk why.
-//     // run_flash_bwd<Flash_bwd_kernel_traits<128, 64, 128, 8, 2, 2, 2, false, false, elem_type>>(params, stream, configure);
-//         run_flash_bwd<Flash_bwd_kernel_traits<128, 64, 128, 8, 2, 4, 2, false, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<128, 64, 128, 8, 2, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<128, 128, 64, 8, 4, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     } else {
-//         run_flash_bwd_seqq_parallel<Flash_bwd_kernel_traits<128, 128, 64, 8, 4, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     }
-// }
-
 template<>
 void run_mha_bwd_<cutlass::half_t, 128>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim128<cutlass::half_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim160_bf16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim160<cutlass::bfloat16_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim160_fp16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim160<cutlass::half_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim192_bf16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim192<cutlass::bfloat16_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim192_fp16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim192<cutlass::half_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim224_bf16_sm80.cu

 // Copyright (c) 2023, 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, 224>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim224<cutlass::bfloat16_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim224_fp16_sm80.cu

 // Copyright (c) 2023, 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, 224>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim224<cutlass::half_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim256_bf16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim256<cutlass::bfloat16_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim256_fp16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim256<cutlass::half_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim32_bf16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
-//     using elem_type = cutlass::bfloat16_t;
-//     run_flash_bwd<Flash_bwd_kernel_traits<32, 128, 128, 8, 4, 4, 4, false, false, elem_type>>(params, stream, configure);
-// }
-
 template<>
 void run_mha_bwd_<cutlass::bfloat16_t, 32>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim32<cutlass::bfloat16_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim32_fp16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
-//     using elem_type = cutlass::half_t;
-//     run_flash_bwd<Flash_bwd_kernel_traits<32, 128, 128, 8, 4, 4, 4, false, false, elem_type>>(params, stream, configure);
-// }
-
 template<>
 void run_mha_bwd_<cutlass::half_t, 32>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim32<cutlass::half_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim64_bf16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
-//     using elem_type = cutlass::bfloat16_t;
-//     run_flash_bwd<Flash_bwd_kernel_traits<64, 128, 128, 8, 4, 4, 4, false, false, elem_type>>(params, stream, configure);
-// }
-
 template<>
 void run_mha_bwd_<cutlass::bfloat16_t, 64>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim64<cutlass::bfloat16_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim64_fp16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
-//     using elem_type = cutlass::half_t;
-//     // Changing AtomLayoutMdQ from 2 to 4 takes the same time
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 64, 128, 8, 2, 4, 2, false, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 64, 128, 8, 2, 4, 2, true, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 128, 128, 8, 2, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     // This is slightly faster. We want to split M more so we need fewer registers to store LSE.
-//     run_flash_bwd<Flash_bwd_kernel_traits<64, 128, 128, 8, 4, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 128, 64, 8, 4, 2, 4, true, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 64, 64, 4, 2, 2, 2, true, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 32, 128, 4, 1, 4, 1, false, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 16, 128, 4, 1, 4, 1, false, false, elem_type>>(params, stream, configure);
-//     // M=128, N=64 is quite slow, I think because we need to read/write dQaccum twice as many times
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 128, 64, 8, 2, 2, 2, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 128, 64, 8, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 16, 256, 8, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 128, 128, 8, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 64, 64, 4, false, elem_type>>(params, stream, configure);
-
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 256, 64, 8, 8, 4, 8, false, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 256, 64, 8, 8, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     // run_flash_bwd<Flash_bwd_kernel_traits<64, 128, 64, 4, 4, 2, 4, false, false, elem_type>>(params, stream, configure);
-// }
-
 template<>
 void run_mha_bwd_<cutlass::half_t, 64>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim64<cutlass::half_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim96_bf16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
-//     using elem_type = cutlass::bfloat16_t;
-//     if (params.h == params.h_k) {
-//         run_flash_bwd<Flash_bwd_kernel_traits<96, 64, 128, 8, 2, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     } else {
-//         run_flash_bwd_seqq_parallel<Flash_bwd_kernel_traits<96, 128, 64, 8, 4, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     }
-// }
-
 template<>
 void run_mha_bwd_<cutlass::bfloat16_t, 96>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim96<cutlass::bfloat16_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_bwd_hdim96_fp16_sm80.cu

 // Copyright (c) 2023, 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 &params, cudaStream_t stream, const bool configure) {
-//     using elem_type = cutlass::half_t;
-//     if (params.h == params.h_k) {
-//         // run_flash_bwd<Flash_bwd_kernel_traits<96, 64, 128, 8, 2, 4, 4, true, false, elem_type>>(params, stream, configure);
-//         // This is very slightly faster
-//         run_flash_bwd<Flash_bwd_kernel_traits<96, 64, 128, 8, 2, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     } else {
-//         run_flash_bwd_seqq_parallel<Flash_bwd_kernel_traits<96, 128, 64, 8, 4, 4, 4, false, false, elem_type>>(params, stream, configure);
-//     }
-// }
-
 template<>
 void run_mha_bwd_<cutlass::half_t, 96>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
     run_mha_bwd_hdim96<cutlass::half_t>(params, stream, configure);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_fwd_hdim128_bf16_sm80.cu

 // Copyright (c) 2023, 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_fwd_launch_template.h"
 
-// template<>
-// void run_mha_fwd_<cutlass::bfloat16_t, 128>(Flash_fwd_params &params, cudaStream_t stream) {
-//     using elem_type = cutlass::bfloat16_t;
-//     if (params.p_dropout == 1.f) {
-//         run_flash_fwd<Flash_fwd_kernel_traits<128, 128, 64, 4, false, false, elem_type>, false>(params, stream);
-//     } else {
-//         run_flash_fwd<Flash_fwd_kernel_traits<128, 128, 32, 4, false, false, elem_type>, true>(params, stream);
-//     }
-// }
 template<>
 void run_mha_fwd_<cutlass::bfloat16_t, 128>(Flash_fwd_params &params, cudaStream_t stream) {
     run_mha_fwd_hdim128<cutlass::bfloat16_t>(params, stream);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_fwd_hdim128_fp16_sm80.cu

 // Copyright (c) 2023, 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_fwd_launch_template.h"
 
-// template<>
-// void run_mha_fwd_<cutlass::half_t, 128>(Flash_fwd_params &params, cudaStream_t stream) {
-//     using elem_type = cutlass::half_t;
-//     if (params.p_dropout == 1.f) {
-//         // Using 8 warps (128 x 128 and 256 x 64) is 28% slower for seqlen=2k
-//         run_flash_fwd<Flash_fwd_kernel_traits<128, 128, 64, 4, false, false, elem_type>, false>(params, stream);
-//         // run_flash_fwd<Flash_fwd_kernel_traits<128, 128, 64, 4, true, false, elem_type>, false>(params, stream);
-//         // run_flash_fwd<Flash_fwd_kernel_traits<128, 128, 64, 4, false, true, elem_type>, false>(params, stream);
-//         // run_flash_fwd<Flash_fwd_kernel_traits<128, 128, 64, 4, true, true, elem_type>, false>(params, stream);
-//         run_flash_fwd<Flash_fwd_kernel_traits<128, 128, 32, 4, false, false, elem_type>, false>(params, stream);
-//         run_flash_fwd<Flash_fwd_kernel_traits<128, 64, 64, 4, false, false, elem_type>, false>(params, stream);
-//         run_flash_fwd<Flash_fwd_kernel_traits<128, 64, 128, 4, false, false, elem_type>, false>(params, stream);
-//         // 1st ones are good for H100, A100
-//         // 2nd one is good for A6000 bc we get slightly better occupancy
-//     } else {
-//         run_flash_fwd<Flash_fwd_kernel_traits<128, 128, 32, 4, false, false, elem_type>, true>(params, stream);
-//         run_flash_fwd<Flash_fwd_kernel_traits<128, 128, 32, 4, true, false, elem_type>, true>(params, stream);
-//         run_flash_fwd<Flash_fwd_kernel_traits<128, 128, 32, 4, true, true, elem_type>, true>(params, stream);
-//         // 1st one is good for H100, A100, A6000
-//     }
-// }
-
 template<>
 void run_mha_fwd_<cutlass::half_t, 128>(Flash_fwd_params &params, cudaStream_t stream) {
     run_mha_fwd_hdim128<cutlass::half_t>(params, stream);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_fwd_hdim160_bf16_sm80.cu

 // Copyright (c) 2023, 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_fwd_launch_template.h"
 
-// template<>
-// void run_mha_fwd_<cutlass::bfloat16_t, 160>(Flash_fwd_params &params, cudaStream_t stream) {
-//     using elem_type = cutlass::bfloat16_t;
-//     BOOL_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
-//         run_flash_fwd<Flash_fwd_kernel_traits<160, 128, 32, 4, false, false, elem_type>, Is_dropout>(params, stream);
-//     });
-// }
 template<>
 void run_mha_fwd_<cutlass::bfloat16_t, 160>(Flash_fwd_params &params, cudaStream_t stream) {
     run_mha_fwd_hdim160<cutlass::bfloat16_t>(params, stream);
-}
\ No newline at end of file
+}

csrc/flash_attn/src/flash_fwd_hdim160_fp16_sm80.cu

 // Copyright (c) 2023, 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_fwd_launch_template.h"
 
-// template<>
-// void run_mha_fwd_<cutlass::half_t, 160>(Flash_fwd_params &params, cudaStream_t stream) {
-//     using elem_type = cutlass::half_t;
-//     BOOL_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
-//         run_flash_fwd<Flash_fwd_kernel_traits<160, 128, 32, 4, false, false, elem_type>, Is_dropout>(params, stream);
-//         run_flash_fwd<Flash_fwd_kernel_traits<160, 128, 32, 4, false, true, elem_type>, Is_dropout>(params, stream);
-//         run_flash_fwd<Flash_fwd_kernel_traits<160, 128, 64, 4, false, false, elem_type>, Is_dropout>(params, stream);
-//         run_flash_fwd<Flash_fwd_kernel_traits<160, 64, 64, 4, false, false, elem_type>, Is_dropout>(params, stream);
-//         // run_flash_fwd<Flash_fwd_kernel_traits<160, 128, 64, 4, false, elem_type>>(params, stream);
-//         // run_flash_fwd<Flash_fwd_kernel_traits<160, 64, 128, 4, false, elem_type>>(params, stream);
-//         // run_flash_fwd<Flash_fwd_kernel_traits<160, 64, 64, 4, false, elem_type>>(params, stream);
-//         // run_flash_fwd<Flash_fwd_kernel_traits<160, 128, 64, 8, false, elem_type>>(params, stream);
-//         // run_flash_fwd<Flash_fwd_kernel_traits<160, 128, 128, 8, false, elem_type>>(params, stream);
-//         // For A6000, no-causal, 1st is fastest. causal, 4th is fastest.
-//         // For A100, H100, 1st is fastest.
-//     });
-// }
 template<>
 void run_mha_fwd_<cutlass::half_t, 160>(Flash_fwd_params &params, cudaStream_t stream) {
     run_mha_fwd_hdim160<cutlass::half_t>(params, stream);
-}
\ No newline at end of file
+}