[Bugfix] Fix Qwen25VL packed_modules_mapping (vllm-project#23604)

jeejeelee · mgoin · amd-xiaoyu12 · commit 4892eb68b79d · 2025-08-28T15:53:52.000Z
Signed-off-by: Jee Jee Li &lt;pandaleefree@gmail.com&gt;
Co-authored-by: Michael Goin &lt;mgoin64@gmail.com&gt;
Signed-off-by: Xiao Yu &lt;xiao.yu@amd.com&gt;
diff --git a/vllm/model_executor/models/qwen2_5_vl.py b/vllm/model_executor/models/qwen2_5_vl.py
@@ -853,6 +853,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
                                          SupportsQuant):
 
     packed_modules_mapping = {
+        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
     }
 

Original file line number	Diff line number	Diff line change
`@@ -853,6 +853,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,`
`853`	`853`	`SupportsQuant):`
`854`	`854`
`855`	`855`	`packed_modules_mapping = {`
	`856`	`+ "qkv_proj": ["q_proj", "k_proj", "v_proj"],`
`856`	`857`	`"gate_up_proj": ["gate_proj", "up_proj"],`
`857`	`858`	`}`
`858`	`859`