[Misc] Log the reason for falling back to FlexAttention (#20699)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
2025-07-14 19:16:51 +08:00
parent a4851cfe68
commit e8cc53af5e
10 changed files with 105 additions and 33 deletions
--- a/vllm/v1/attention/backends/flashinfer.py
+++ b/vllm/v1/attention/backends/flashinfer.py
@@ -42,6 +42,10 @@ class FlashInferBackend(AttentionBackend):
    accept_output_buffer: bool = True
    cached_sm100a_supported: Optional[bool] = None

+    @classmethod
+    def get_supported_dtypes(cls) -> list[torch.dtype]:
+        return [torch.float16, torch.bfloat16]
+
    @classmethod
    def get_supported_head_sizes(cls) -> list[int]:
        # https://github.com/flashinfer-ai/flashinfer/blob/3d55c71a62052c590c130897d3a3db49b14fcc34/include/flashinfer/utils.cuh#L157