[Bugfix] Disable cascade attention with FlashInfer (#26130)

Signed-off-by: mgoin <[email protected]>
Signed-off-by: Michael Goin <[email protected]>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Signed-off-by: simon-mo <[email protected]>

Author: 2 people

vllm/v1/attention/backends/flashinfer.py

@@ -29,7 +29,6 @@
                                    flashinfer_disable_q_quantization,
                                    supports_trtllm_attention,
                                    use_trtllm_attention)
-from vllm.v1.attention.backends.flash_attn import use_cascade_attention
 # yapf conflicts with isort for this block
 # yapf: disable
 from vllm.v1.attention.backends.utils import (AttentionCGSupport,
@@ -677,7 +676,9 @@ def use_cascade_attention(self, *args, **kwargs) -> bool:
             # TODO: The cascade wrapper currently does not support setting
             # kv cache dtype to something different from query dtype.
             return False
-        return use_cascade_attention(*args, **kwargs)
+        # TODO: Cascade attention doesn't work, disable it for now
+        # return use_cascade_attention(*args, **kwargs)
+        return False
 
 
 class FlashInferImpl(AttentionImpl):