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, 01:45 UTC
[Temp fix] Disable torch.compile for Qwen2.5 VL's VisionBlock temporarily. (#27760)
Signed-off-by: Chenheli Hua <huachenheli@outlook.com> Signed-off-by: Roger Wang <hey@rogerw.io> Co-authored-by: Roger Wang <hey@rogerw.io>
Chenheli Hua committed
Oct 29, 2025 at 16:17 UTC
48eb8eba581f0e45272f4e763bf5ec342f77091a
1 file changed
+11
-9
vllm/model_executor/models/qwen2_5_vl.py
+11
-9
@@ -460,15 +460,17 @@ class Qwen2_5_VisionAttention(nn.Module):
460
return output
461
462
463
-@support_torch_compile(
464
- dynamic_arg_dims={
465
- "x": 0,
466
- "cu_seqlens": 0,
467
- "rotary_pos_emb": 0,
468
- "seqlens": 0,
469
- },
470
- mark_unbacked_dims={"seqlens": 0},
471
-)
463
+# (FIXME): Enable this after dynamic slicing is fixed
464
+# See https://github.com/vllm-project/vllm/pull/27760
465
+# @support_torch_compile(
466
+# dynamic_arg_dims={
467
+# "x": 0,
468
+# "cu_seqlens": 0,
469
+# "rotary_pos_emb": 0,
470
+# "seqlens": 0,
471
+# },
472
+# mark_unbacked_dims={"seqlens": 0},
473
+# )
474
class Qwen2_5_VisionBlock(nn.Module):
475
def __init__(
476
self,