xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-24 16:35:55 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] fix Qwen3VLMoe load when pp > 1 (#25838)

Browse Source 
Signed-off-by: liuye.hj <liuye.hj@alibaba-inc.com>
Co-authored-by: liuye.hj <liuye.hj@alibaba-inc.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
JJJYmmm 2025-09-29 01:56:12 +08:00 committed by yewentao256
parent e40c12696a
commit 0b343e3218
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/models/qwen3_vl_moe.py
View File

@ -212,6 +212,8 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
# attempted to load as other weights later
is_expert_weight = True
name_mapped = name.replace(weight_name, param_name)
if is_pp_missing_parameter(name_mapped, self):
continue
if is_fused_expert:
loaded_weight = loaded_weight.transpose(-1,
-2) # no bias
@ -230,8 +232,6 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
name_mapped, params_dict, loaded_weight,
shard_id, num_experts)
else:
if is_pp_missing_parameter(name_mapped, self):
continue
# Skip loading extra parameters for GPTQ/modelopt models
if name_mapped.endswith(
ignore_suffixes