A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix] Disable cascade attention with FlashInfer (#26130)
Signed-off-by: mgoin <mgoin64@gmail.com> Signed-off-by: Michael Goin <mgoin64@gmail.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Michael Goin committed
Oct 2, 2025 at 19:30 UTC
f1fc2107a314ddfdba31348dc134bb861ccd77a1
1 file changed
+3
-2
vllm/v1/attention/backends/flashinfer.py
+3
-2
@@ -29,7 +29,6 @@ from vllm.utils.flashinfer import (can_use_trtllm_attention,
29
flashinfer_disable_q_quantization,
30
supports_trtllm_attention,
31
use_trtllm_attention)
32
-from vllm.v1.attention.backends.flash_attn import use_cascade_attention
32
# yapf conflicts with isort for this block
33
# yapf: disable
34
from vllm.v1.attention.backends.utils import (AttentionCGSupport,
@@ -677,7 +676,9 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
676
# TODO: The cascade wrapper currently does not support setting
677
# kv cache dtype to something different from query dtype.
678
return False
680
- return use_cascade_attention(*args, **kwargs)
679
+ # TODO: Cascade attention doesn't work, disable it for now
680
+ # return use_cascade_attention(*args, **kwargs)
681
+ return False
682
683
684
class FlashInferImpl(AttentionImpl):