Split bwd into more .cu files to speed up compilation

csrc/flash_attn/src/flash_bwd_hdim32_fp16_causal_sm80.cu

+// 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 &params, cudaStream_t stream) {
+    run_mha_bwd_hdim32<cutlass::half_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_bwd_hdim32_fp16_sm80.cu

-// 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 &params, cudaStream_t stream) {
-    run_mha_bwd_hdim32<cutlass::half_t>(params, stream);
+void run_mha_bwd_<cutlass::half_t, 32, false>(Flash_bwd_params &params, cudaStream_t stream) {
+    run_mha_bwd_hdim32<cutlass::half_t, false>(params, stream);
 }

csrc/flash_attn/src/flash_bwd_hdim64_bf16_causal_sm80.cu

+// 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 &params, cudaStream_t stream) {
+    run_mha_bwd_hdim64<cutlass::bfloat16_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_bwd_hdim64_bf16_sm80.cu

-// 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 &params, cudaStream_t stream) {
-    run_mha_bwd_hdim64<cutlass::bfloat16_t>(params, stream);
+void run_mha_bwd_<cutlass::bfloat16_t, 64, false>(Flash_bwd_params &params, cudaStream_t stream) {
+    run_mha_bwd_hdim64<cutlass::bfloat16_t, false>(params, stream);
 }

csrc/flash_attn/src/flash_bwd_hdim64_fp16_causal_sm80.cu

+// 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 &params, cudaStream_t stream) {
+    run_mha_bwd_hdim64<cutlass::half_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_bwd_hdim64_fp16_sm80.cu

-// 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 &params, cudaStream_t stream) {
-    run_mha_bwd_hdim64<cutlass::half_t>(params, stream);
+void run_mha_bwd_<cutlass::half_t, 64, false>(Flash_bwd_params &params, cudaStream_t stream) {
+    run_mha_bwd_hdim64<cutlass::half_t, false>(params, stream);
 }

csrc/flash_attn/src/flash_bwd_hdim96_bf16_causal_sm80.cu

+// 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 &params, cudaStream_t stream) {
+    run_mha_bwd_hdim96<cutlass::bfloat16_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_bwd_hdim96_bf16_sm80.cu

-// 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 &params, cudaStream_t stream) {
-    run_mha_bwd_hdim96<cutlass::bfloat16_t>(params, stream);
+void run_mha_bwd_<cutlass::bfloat16_t, 96, false>(Flash_bwd_params &params, cudaStream_t stream) {
+    run_mha_bwd_hdim96<cutlass::bfloat16_t, false>(params, stream);
 }

csrc/flash_attn/src/flash_bwd_hdim96_fp16_causal_sm80.cu

+// 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 &params, cudaStream_t stream) {
+    run_mha_bwd_hdim96<cutlass::half_t, true>(params, stream);
+}

csrc/flash_attn/src/flash_bwd_hdim96_fp16_sm80.cu

-// 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 &params, cudaStream_t stream) {
-    run_mha_bwd_hdim96<cutlass::half_t>(params, stream);
+void run_mha_bwd_<cutlass::half_t, 96, false>(Flash_bwd_params &params, cudaStream_t stream) {
+    run_mha_bwd_hdim96<cutlass::half_t, false>(params, stream);
 }

csrc/flash_attn/src/flash_bwd_launch_template.h

@@ -65,7 +65,7 @@ __global__ void flash_bwd_convert_dkv_kernel(const Flash_bwd_params params) {
     flash::convert_dKV<Kernel_traits>(params);
 }
 
-template<typename Kernel_traits, bool Is_dropout>
+template<typename Kernel_traits, bool Is_dropout, bool Is_causal>
 void run_flash_bwd_seqk_parallel(Flash_bwd_params &params, cudaStream_t stream) {
     const int num_m_block = (params.seqlen_q + Kernel_traits::kBlockM - 1) / Kernel_traits::kBlockM;
     dim3 grid_m(num_m_block, params.b, params.h);
@@ -90,7 +90,6 @@ void run_flash_bwd_seqk_parallel(Flash_bwd_params &params, cudaStream_t stream)
     const bool is_even_K = params.d == Kernel_traits::kHeadDim;
     constexpr int smem_size_dq_dk_dv = Kernel_traits::kSmemSize1colblock;
     // printf("smem_size_dq_dk_dv = %d\n", smem_size_dq_dk_dv);
-    BOOL_SWITCH(params.is_causal, Is_causal, [&] {
     BOOL_SWITCH(is_even_MN, IsEvenMNConst, [&] {
         EVENK_SWITCH(is_even_K, IsEvenKConst, [&] {
             LOCAL_SWITCH((params.window_size_left >= 0 || params.window_size_right >= 0) && !params.is_causal, Is_local, [&] {
@@ -112,7 +111,6 @@ void run_flash_bwd_seqk_parallel(Flash_bwd_params &params, cudaStream_t stream)
             });
         });
     });
-    });
 
     auto kernel_dq = &flash_bwd_convert_dq_kernel<Kernel_traits>;
     if (Kernel_traits::kSmemdQSize >= 48 * 1024)  {
@@ -123,14 +121,14 @@ void run_flash_bwd_seqk_parallel(Flash_bwd_params &params, cudaStream_t stream)
     C10_CUDA_KERNEL_LAUNCH_CHECK();
 }
 
-template<typename Kernel_traits, bool Is_dropout>
+template<typename Kernel_traits, bool Is_dropout, bool Is_causal>
 void run_flash_bwd(Flash_bwd_params &params, cudaStream_t stream) {
 #ifndef FLASHATTENTION_DISABLE_BACKWARD
-    run_flash_bwd_seqk_parallel<Kernel_traits, Is_dropout>(params, stream);
+    run_flash_bwd_seqk_parallel<Kernel_traits, Is_dropout, Is_causal>(params, stream);
 #endif
 }
 
-template<typename T>
+template<typename T, bool Is_causal>
 void run_mha_bwd_hdim32(Flash_bwd_params &params, cudaStream_t stream) {
     constexpr static int Headdim = 32;
     int device;
@@ -144,17 +142,17 @@ void run_mha_bwd_hdim32(Flash_bwd_params &params, cudaStream_t stream) {
     DROPOUT_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
         if (max_smem_per_block >= 2 * ((3 * 128 + 2 * 128) * Headdim + 2 * 128 * 128)) { // 104 KB
             if constexpr(!Is_dropout) {  // We can afford more registers to keep V in registers
-                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, true, false, T>, Is_dropout>(params, stream);
+                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, true, false, T>, Is_dropout, Is_causal>(params, stream);
             } else {
-                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, false, false, T>, Is_dropout>(params, stream);
+                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, false, false, T>, Is_dropout, Is_causal>(params, stream);
             }
         } else {  // 96 KB
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, true, false, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, true, false, T>, Is_dropout, Is_causal>(params, stream);
         }
     });
 }
 
-template<typename T>
+template<typename T, bool Is_causal>
 void run_mha_bwd_hdim64(Flash_bwd_params &params, cudaStream_t stream) {
     constexpr static int Headdim = 64;
     int device;
@@ -174,13 +172,13 @@ void run_mha_bwd_hdim64(Flash_bwd_params &params, cudaStream_t stream) {
         // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 64, 8, 4, 2, 4, false, false, T>, Is_dropout>(params, stream);
         // This is slightly faster. We want to split M more so we need fewer registers to store LSE.
         if (max_smem_per_block >= 144 * 1024) {
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, false, false, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, false, false, T>, Is_dropout, Is_causal>(params, stream);
             // This has a lot of register spilling
             // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, true, false, T>, Is_dropout>(params, stream);
         } else {
             // if (params.h == params.h_k) {
                 // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 4, false, false, T>, Is_dropout>(params, stream);
-                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 4, true, false, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 4, true, false, T>, Is_dropout, Is_causal>(params, stream);
                 // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 64, 8, 4, 2, 4, false, false, T>, Is_dropout>(params, stream);
                 // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 64, 8, 4, 2, 4, true, false, T>, Is_dropout>(params, stream);
             // } else {
@@ -199,7 +197,7 @@ void run_mha_bwd_hdim64(Flash_bwd_params &params, cudaStream_t stream) {
     // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 64, 4, 4, 2, 4, false, false, T>>(params, stream);
 }
 
-template<typename T>
+template<typename T, bool Is_causal>
 void run_mha_bwd_hdim96(Flash_bwd_params &params, cudaStream_t stream) {
     constexpr static int Headdim = 96;
     int device;
@@ -214,18 +212,18 @@ void run_mha_bwd_hdim96(Flash_bwd_params &params, cudaStream_t stream) {
     DROPOUT_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
         if (max_smem_per_block >= 116 * 1024) {
             if constexpr(!Is_dropout) {  // 92KB
-                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 4, true, false, T>, Is_dropout>(params, stream);
+                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 4, true, false, T>, Is_dropout, Is_causal>(params, stream);
             } else {  // 116 KB
                 // This is faster for dropout since we don't have many registers to spare
-                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 4, false, false, T>, Is_dropout>(params, stream);
+                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 4, false, false, T>, Is_dropout, Is_causal>(params, stream);
             }
         } else {
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 4, true, false, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 4, true, false, T>, Is_dropout, Is_causal>(params, stream);
         }
     });
 }
 
-template<typename T>
+template<typename T, bool Is_causal>
 void run_mha_bwd_hdim128(Flash_bwd_params &params, cudaStream_t stream) {
     constexpr static int Headdim = 128;
     int device;
@@ -243,7 +241,7 @@ void run_mha_bwd_hdim128(Flash_bwd_params &params, cudaStream_t stream) {
         // Out of these three, the 2nd one is slightly faster (2% faster than the first). Idk why.
         // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 2, 2, false, false, T>>(params, stream);
         if (max_smem_per_block >= 144 * 1024) {
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 2, false, false, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 2, false, false, T>, Is_dropout, Is_causal>(params, stream);
             // run_flash_bwd_seqk_parallel<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, false, false, T>, Is_dropout>(params, stream);
             // run_flash_bwd_seqk_parallel<Flash_bwd_kernel_traits<Headdim, 128, 128, 8, 4, 4, 4, false, true, T>, Is_dropout>(params, stream);
             // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 2, true, false, T>, Is_dropout>(params, stream);
@@ -251,7 +249,7 @@ void run_mha_bwd_hdim128(Flash_bwd_params &params, cudaStream_t stream) {
             // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 128, 64, 8, 4, 2, 2, true, false, T>, Is_dropout>(params, stream);
         } else {
             // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, false, false, T>, Is_dropout>(params, stream);
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, true, false, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, true, false, T>, Is_dropout, Is_causal>(params, stream);
         }
         // run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 128, 8, 2, 4, 4, false, false, T>>(params, stream);
 
@@ -259,7 +257,7 @@ void run_mha_bwd_hdim128(Flash_bwd_params &params, cudaStream_t stream) {
     });
 }
 
-template<typename T>
+template<typename T, bool Is_causal>
 void run_mha_bwd_hdim160(Flash_bwd_params &params, cudaStream_t stream) {
     constexpr static int Headdim = 160;
     int device;
@@ -272,14 +270,14 @@ void run_mha_bwd_hdim160(Flash_bwd_params &params, cudaStream_t stream) {
     }
     DROPOUT_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
         if (max_smem_per_block >= 116 * 1024) {
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 4, 4, false, false, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 4, 4, false, false, T>, Is_dropout, Is_causal>(params, stream);
         } else {
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 4, 4, false, true, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 4, 4, false, true, T>, Is_dropout, Is_causal>(params, stream);
         }
     });
 }
 
-template<typename T>
+template<typename T, bool Is_causal>
 void run_mha_bwd_hdim192(Flash_bwd_params &params, cudaStream_t stream) {
     constexpr static int Headdim = 192;
     int device;
@@ -292,14 +290,14 @@ void run_mha_bwd_hdim192(Flash_bwd_params &params, cudaStream_t stream) {
     }
     DROPOUT_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
         if (max_smem_per_block >= 136 * 1024) {
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, false, false, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, false, false, T>, Is_dropout, Is_causal>(params, stream);
         } else {
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, true, true, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, true, true, T>, Is_dropout, Is_causal>(params, stream);
         }
     });
 }
 
-template<typename T>
+template<typename T, bool Is_causal>
 void run_mha_bwd_hdim256(Flash_bwd_params &params, cudaStream_t stream) {
     constexpr static int Headdim = 256;
     int device;
@@ -312,12 +310,12 @@ void run_mha_bwd_hdim256(Flash_bwd_params &params, cudaStream_t stream) {
     }
     DROPOUT_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
         if (max_smem_per_block >= 176 * 1024) {  // H100
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, false, false, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, false, false, T>, Is_dropout, Is_causal>(params, stream);
         } else if (max_smem_per_block >= 144 * 1024) {  // A100, we don't do double buffering to save smem
-            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, false, true, T>, Is_dropout>(params, stream);
+            run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 2, 2, false, true, T>, Is_dropout, Is_causal>(params, stream);
         } else { // sm86 and sm89, max smem is 99 KB. Only works without dropout. V in regs and no double buffering.
             if constexpr (!Is_dropout) {
-                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 32, 8, 4, 1, 2, true, true, T>, false>(params, stream);
+                run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 32, 8, 4, 1, 2, true, true, T>, false, Is_causal>(params, stream);
             }
         }
     });

csrc/flash_attn/src/flash_fwd_hdim128_bf16_causal_sm80.cu

-// 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"
 

csrc/flash_attn/src/flash_fwd_hdim128_bf16_sm80.cu

-// 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"
 

csrc/flash_attn/src/flash_fwd_hdim128_fp16_causal_sm80.cu

-// 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"
 

csrc/flash_attn/src/flash_fwd_hdim128_fp16_sm80.cu

-// 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"
 

csrc/flash_attn/src/flash_fwd_hdim160_bf16_causal_sm80.cu

-// 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"
 

csrc/flash_attn/src/flash_fwd_hdim160_bf16_sm80.cu

-// 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"
 

csrc/flash_attn/src/flash_fwd_hdim160_fp16_causal_sm80.cu

-// 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"
 

csrc/flash_attn/src/flash_fwd_hdim160_fp16_sm80.cu

-// 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"
 

csrc/flash_attn/src/flash_fwd_hdim192_bf16_causal_sm80.cu

-// 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"