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 11, 08:45 UTC
Update note comment for flashinfer attention warmup (#30711)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Dec 17, 2025 at 00:29 UTC
d4d2751732c3ccae162a5a0160c7d4fe05d2779a
1 file changed
+3
-4
vllm/model_executor/warmup/kernel_warmup.py
+3
-4
@@ -49,13 +49,12 @@ def kernel_warmup(worker: "Worker"):
49
except NotImplementedError:
50
return False
51
52
- # NOTE: we add check for empty attn_groups to avoid errors when
53
- # deploying models such as E instances and encoder-only models.
54
- # As for those models, worker.model_runner.attn_groups is empty.
55
- # This change is made during EPD feature development.
52
if (
53
not worker.model_runner.is_pooling_model
54
and worker.model_runner.attn_groups
55
+ # NOTE: This should be `any` instead of `all` but other hybrid attention
56
+ # backends don't support this dummy run. Once we remove
57
+ # `build_for_cudagraph_capture`, we can change it to `any`.
58
and all(
59
_is_flashinfer_backend(group.backend)
60
for groups in worker.model_runner.attn_groups