Split into more .cu files to speed up compilation

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

@@ -5,6 +5,6 @@
 #include "flash_fwd_launch_template.h"
 
 template<>
-void run_mha_fwd_<cutlass::half_t, 256>(Flash_fwd_params &params, cudaStream_t stream) {
-    run_mha_fwd_hdim256<cutlass::half_t>(params, stream);
+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

@@ -5,6 +5,6 @@
 #include "flash_fwd_launch_template.h"
 
 template<>
-void run_mha_fwd_<cutlass::bfloat16_t, 32>(Flash_fwd_params &params, cudaStream_t stream) {
-    run_mha_fwd_hdim32<cutlass::bfloat16_t>(params, stream);
+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

@@ -5,6 +5,6 @@
 #include "flash_fwd_launch_template.h"
 
 template<>
-void run_mha_fwd_<cutlass::half_t, 32>(Flash_fwd_params &params, cudaStream_t stream) {
-    run_mha_fwd_hdim32<cutlass::half_t>(params, stream);
+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_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

@@ -5,6 +5,6 @@
 #include "flash_fwd_launch_template.h"
 
 template<>
-void run_mha_fwd_<cutlass::bfloat16_t, 64>(Flash_fwd_params &params, cudaStream_t stream) {
-    run_mha_fwd_hdim64<cutlass::bfloat16_t>(params, stream);
+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

@@ -5,6 +5,6 @@
 #include "flash_fwd_launch_template.h"
 
 template<>
-void run_mha_fwd_<cutlass::half_t, 64>(Flash_fwd_params &params, cudaStream_t stream) {
-    run_mha_fwd_hdim64<cutlass::half_t>(params, stream);
+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

@@ -5,6 +5,6 @@
 #include "flash_fwd_launch_template.h"
 
 template<>
-void run_mha_fwd_<cutlass::bfloat16_t, 96>(Flash_fwd_params &params, cudaStream_t stream) {
-    run_mha_fwd_hdim96<cutlass::bfloat16_t>(params, stream);
+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);
+}

csrc/flash_attn/src/flash_fwd_hdim96_fp16_sm80.cu

@@ -5,6 +5,6 @@
 #include "flash_fwd_launch_template.h"
 
 template<>
-void run_mha_fwd_<cutlass::half_t, 96>(Flash_fwd_params &params, cudaStream_t stream) {
-    run_mha_fwd_hdim96<cutlass::half_t>(params, stream);
+void run_mha_fwd_<cutlass::half_t, 96, false>(Flash_fwd_params &params, cudaStream_t stream) {
+    run_mha_fwd_hdim96<cutlass::half_t, false>(params, stream);
 }

csrc/flash_attn/src/flash_fwd_split_hdim128_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, 128, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim128_bf16_sm80.cu

@@ -4,4 +4,4 @@
 
 #include "flash_fwd_launch_template.h"
 
-template void run_mha_fwd_splitkv_dispatch<cutlass::bfloat16_t, 128>(Flash_fwd_params &params, cudaStream_t stream);
+template void run_mha_fwd_splitkv_dispatch<cutlass::bfloat16_t, 128, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim128_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, 128, true>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim128_fp16_sm80.cu

@@ -4,4 +4,4 @@
 
 #include "flash_fwd_launch_template.h"
 
-template void run_mha_fwd_splitkv_dispatch<cutlass::half_t, 128>(Flash_fwd_params &params, cudaStream_t stream);
+template void run_mha_fwd_splitkv_dispatch<cutlass::half_t, 128, false>(Flash_fwd_params &params, cudaStream_t stream);

csrc/flash_attn/src/flash_fwd_split_hdim160_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, 160, true>(Flash_fwd_params &params, cudaStream_t stream);