[release/2.4] [ROCM] Properly disable Flash Attention/Efficient Attention with environment variables (#1570)

xinyazhang · jithunnair-amd · commit 98d727f5a229 · 2025-03-17T19:16:56.000Z
Now `USE_FLASH_ATTENTION=0 USE_MEM_EFF_ATTENTION=0 python setup.py` can compile correctly. This is cherry-picked version of pytorch#133866
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -874,6 +874,16 @@ cmake_dependent_option(
   Will be disabled if not supported by the platform" ON
   "USE_CUDA OR USE_ROCM" OFF)
 
+#
+# Cannot be put into Dependencies.cmake due circular dependency:
+# USE_FLASH_ATTENTION -> USE_ROCM -> Dependencies.cmake -> aotriton.cmake
+#
+if(USE_ROCM)
+  if(USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION)
+    include(cmake/External/aotriton.cmake)
+  endif()
+endif()
+
 if(DEBUG_CUDA)
   string(APPEND CMAKE_CUDA_FLAGS_DEBUG " -lineinfo")
   string(APPEND CMAKE_CUDA_FLAGS_RELWITHDEBINFO " -lineinfo")
diff --git a/aten/src/ATen/native/transformers/cuda/sdp_utils.cpp b/aten/src/ATen/native/transformers/cuda/sdp_utils.cpp
@@ -20,7 +20,10 @@
 #include <c10/util/string_view.h>
 
 #if USE_ROCM
+#if defined(USE_FLASH_ATTENTION) || defined(USE_MEM_EFF_ATTENTION)
 #include <aotriton/flash.h>
+#define USE_AOTRITON 1
+#endif
 #endif
 
 /**
@@ -185,6 +188,7 @@ bool check_flash_attention_hardware_support(sdp_params const& params, bool debug
   using sm80 = SMVersion<8, 0>;
   using sm90 = SMVersion<9, 0>;
 #if USE_ROCM
+#if USE_AOTRITON
   auto stream = at::cuda::getCurrentCUDAStream().stream();
   if (hipSuccess != aotriton::v2::flash::check_gpu(stream)) {
       auto dprops = at::cuda::getCurrentDeviceProperties();
@@ -194,6 +198,9 @@ bool check_flash_attention_hardware_support(sdp_params const& params, bool debug
       }
       return false;
   }
+#else
+  return false;
+#endif
 #else
   auto dprops = at::cuda::getCurrentDeviceProperties();
   if (!check_sm_version<sm80, sm90>(dprops)) {
@@ -216,6 +223,7 @@ bool check_mem_efficient_hardware_support(sdp_params const& params, bool debug)
   using sm50 = SMVersion<5, 0>;
   using sm90 = SMVersion<9, 0>;
 #if USE_ROCM
+#if USE_AOTRITON
   auto stream = at::cuda::getCurrentCUDAStream().stream();
   if (hipSuccess != aotriton::v2::flash::check_gpu(stream)) {
       auto dprops = at::cuda::getCurrentDeviceProperties();
@@ -225,6 +233,9 @@ bool check_mem_efficient_hardware_support(sdp_params const& params, bool debug)
       }
       return false;
   }
+#else
+  return false;
+#endif
 #else
   auto dprops = at::cuda::getCurrentDeviceProperties();
   if (!check_sm_version<sm50, sm90>(dprops)) {
@@ -238,8 +249,9 @@ bool check_mem_efficient_hardware_support(sdp_params const& params, bool debug)
     }
     return false;
   }
-#endif
   return true;
+#endif
+  return false;
 }
 
 bool check_requires_grad_and_head_dim_gt192_constraints_on_sm86_89(
diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
@@ -1097,7 +1097,6 @@ if(USE_ROCM)
       message(STATUS "Disabling Kernel Assert for ROCm")
     endif()
 
-    include(${CMAKE_CURRENT_LIST_DIR}/External/aotriton.cmake)
     if(USE_CUDA)
       caffe2_update_option(USE_MEM_EFF_ATTENTION OFF)
     endif()
