[Hardware] Update the flash attn tag to support Blackwell (#14244)

Author: pavanimajety

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 720c94869cf2e0ff5a706e9c7f1dce0939686ade
+          GIT_TAG 9bfa9869829d8c593527eb34c5271d0090f7ccc9 
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
@@ -64,4 +64,4 @@ install(
   DESTINATION vllm_flash_attn
   COMPONENT _vllm_fa3_C
   FILES_MATCHING PATTERN "*.py"
-)
+)

vllm/attention/backends/utils.py

@@ -595,14 +595,19 @@ def get_flash_attn_version():
         # if hopper default to FA3, otherwise stick to FA2 for now
         # TODO(lucas): profile FA3 on ampere to see if it makes sense to
         #  use FA3 as default for both
-        if current_platform.get_device_capability()[0] >= 9:
+        if current_platform.get_device_capability()[0] == 9:
             fa_version = 3 if is_fa_version_supported(3) else 2
         else:
             fa_version = 2
 
         if envs.VLLM_FLASH_ATTN_VERSION is not None:
             assert envs.VLLM_FLASH_ATTN_VERSION in [2, 3]
             fa_version = envs.VLLM_FLASH_ATTN_VERSION
+            if (current_platform.get_device_capability()[0] == 10
+                    and envs.VLLM_FLASH_ATTN_VERSION == 3):
+                logger.warning("Cannot use FA version 3 on Blackwell platform",
+                               "defaulting to FA version 2.")
+                fa_version = 2
 
         if not is_fa_version_supported(fa_version):
             logger.error("Cannot use FA version %d is not supported due to %s",