first commit

csrc/flash_attn/src/flash_fwd_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_fwd_launch_template.h"
+
+template<>
+void run_mha_fwd_<cutlass::half_t, 224, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim224<cutlass::half_t, false>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_hdim256_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_<cutlass::bfloat16_t, 256, true>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim256<cutlass::bfloat16_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_fwd_launch_template.h"
+
+template<>
+void run_mha_fwd_<cutlass::bfloat16_t, 256, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim256<cutlass::bfloat16_t, false>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_hdim256_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_<cutlass::half_t, 256, true>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim256<cutlass::half_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_fwd_launch_template.h"
+
+template<>
+void run_mha_fwd_<cutlass::half_t, 256, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim256<cutlass::half_t, false>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_<cutlass::bfloat16_t, 32, true>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim32<cutlass::bfloat16_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_<cutlass::bfloat16_t, 32, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim32<cutlass::bfloat16_t, false>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_<cutlass::half_t, 32, true>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim32<cutlass::half_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_<cutlass::half_t, 32, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim32<cutlass::half_t, false>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_hdim512_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_<cutlass::bfloat16_t, 512, true>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim512<cutlass::bfloat16_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_hdim512_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, 512, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim512<cutlass::bfloat16_t, false>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_hdim512_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_<cutlass::half_t, 512, true>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim512<cutlass::half_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_hdim512_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, 512, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim512<cutlass::half_t, false>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_<cutlass::bfloat16_t, 64, true>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim64<cutlass::bfloat16_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_<cutlass::bfloat16_t, 64, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim64<cutlass::bfloat16_t, false>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_<cutlass::half_t, 64, true>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim64<cutlass::half_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_<cutlass::half_t, 64, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim64<cutlass::half_t, false>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_hdim96_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_<cutlass::bfloat16_t, 96, true>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim96<cutlass::bfloat16_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_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_fwd_launch_template.h"
+
+template<>
+void run_mha_fwd_<cutlass::bfloat16_t, 96, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim96<cutlass::bfloat16_t, false>(params, stream);
+}

csrc/flash_attn/src/flash_fwd_hdim96_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_<cutlass::half_t, 96, true>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim96<cutlass::half_t, true>(params, stream);
+}