first commit

csrc/flash_attn/src/flash_fwd_split_hdim256_q_bf16_kv_e5m2_causal_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_splitkv_dispatch_kv_fp8<cutlass::bfloat16_t, cutlass::float_e5m2_t,256, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim256_q_bf16_kv_e5m2_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_splitkv_dispatch_kv_fp8<cutlass::bfloat16_t, cutlass::float_e5m2_t,256, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim256_q_fp16_kv_e5m2_causal_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_splitkv_dispatch_kv_fp8<cutlass::half_t, cutlass::float_e5m2_t,256, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim256_q_fp16_kv_e5m2_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_splitkv_dispatch_kv_fp8<cutlass::half_t, cutlass::float_e5m2_t,256, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim32_bf16_causal_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_splitkv_dispatch<cutlass::bfloat16_t, 32, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_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_fwd_launch_template.h"
+
+template void run_mha_fwd_splitkv_dispatch<cutlass::bfloat16_t, 32, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim32_fp16_causal_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_splitkv_dispatch<cutlass::half_t, 32, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_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_fwd_launch_template.h"
+
+template void run_mha_fwd_splitkv_dispatch<cutlass::half_t, 32, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim64_bf16_causal_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_splitkv_dispatch<cutlass::bfloat16_t, 64, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_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_fwd_launch_template.h"
+
+template void run_mha_fwd_splitkv_dispatch<cutlass::bfloat16_t, 64, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim64_fp16_causal_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_splitkv_dispatch<cutlass::half_t, 64, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_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_fwd_launch_template.h"
+
+template void run_mha_fwd_splitkv_dispatch<cutlass::half_t, 64, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim64_fp8_causal_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_splitkv_dispatch_fp8<cutlass::float_e4m3_t,cutlass::bfloat16_t, 64, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim64_fp8_e5m2_causal_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_splitkv_dispatch_fp8<cutlass::float_e5m2_t,cutlass::bfloat16_t, 64, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim64_fp8_e5m2_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_splitkv_dispatch_fp8<cutlass::float_e5m2_t,cutlass::bfloat16_t, 64, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim64_fp8_outfp16_causal_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_splitkv_dispatch_fp8<cutlass::float_e4m3_t,cutlass::half_t, 64, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim64_fp8_outfp16_e5m2_causal_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_splitkv_dispatch_fp8<cutlass::float_e5m2_t,cutlass::half_t, 64, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim64_fp8_outfp16_e5m2_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_splitkv_dispatch_fp8<cutlass::float_e5m2_t,cutlass::half_t, 64, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim64_fp8_outfp16_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_splitkv_dispatch_fp8<cutlass::float_e4m3_t,cutlass::half_t, 64, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim64_fp8_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_splitkv_dispatch_fp8<cutlass::float_e4m3_t,cutlass::bfloat16_t, 64, false>(Flash_fwd_params &params, cudaStream_t stream);