add fp8 support on bw

vllm/attention/utils/fa_utils.py

@@ -69,7 +69,7 @@ def get_flash_attn_version(requires_alibi: bool = False) -> Optional[int]:
 
 
 def flash_attn_supports_fp8() -> bool:
-    if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938":
+    if current_platform.is_rocm():
         return True
     return get_flash_attn_version() == 3 and \
         current_platform.get_device_capability().major == 9