From 48eb8eba581f0e45272f4e763bf5ec342f77091a Mon Sep 17 00:00:00 2001 From: Chenheli Hua Date: Wed, 29 Oct 2025 16:17:48 -0700 Subject: [PATCH] [Temp fix] Disable torch.compile for Qwen2.5 VL's VisionBlock temporarily. (#27760) Signed-off-by: Chenheli Hua Signed-off-by: Roger Wang Co-authored-by: Roger Wang --- vllm/model_executor/models/qwen2_5_vl.py | 20 +++++++++++--------- 1 file changed, 11 insertions(+), 9 deletions(-) diff --git a/vllm/model_executor/models/qwen2_5_vl.py b/vllm/model_executor/models/qwen2_5_vl.py index 41cb7084057d..dfaeb663bbe2 100644 --- a/vllm/model_executor/models/qwen2_5_vl.py +++ b/vllm/model_executor/models/qwen2_5_vl.py @@ -460,15 +460,17 @@ class Qwen2_5_VisionAttention(nn.Module): return output -@support_torch_compile( - dynamic_arg_dims={ - "x": 0, - "cu_seqlens": 0, - "rotary_pos_emb": 0, - "seqlens": 0, - }, - mark_unbacked_dims={"seqlens": 0}, -) +# (FIXME): Enable this after dynamic slicing is fixed +# See https://github.com/vllm-project/vllm/pull/27760 +# @support_torch_compile( +# dynamic_arg_dims={ +# "x": 0, +# "cu_seqlens": 0, +# "rotary_pos_emb": 0, +# "seqlens": 0, +# }, +# mark_unbacked_dims={"seqlens": 0}, +# ) class Qwen2_5_VisionBlock(nn.Module): def __init__( self,