[Bugfix] fix Qwen3VLMoe load when pp > 1

[JJJYmmm]

Purpose

fix QwenLM/Qwen3-VL#1523

Test Plan

test with command

python3 -m vllm.entrypoints.openai.api_server   
  --model  /path/to/Qwen3-VL-235B-A22B-Instruct   
  --served-model-name Qwen3-VL-235B-A22B-Instruct   
  --tensor-parallel-size 4
  --pipeline-parallel-size 2   
  --mm-encoder-tp-mode data   
  --limit-mm-per-prompt.video 0   
  --mm-processor-cache-type shm   
  --enable-expert-parallel   
  --host 0.0.0.0   
  --port 22002   
  --dtype bfloat16   
  --gpu-memory-utilization 0.95   
  --distributed-executor-backend mp 
  --max-model-len 40960

Test Result

before

(EngineCore_DP0 pid=9735) ERROR 09-28 20:05:29 [core.py:712] param = params_dict[name]
(EngineCore_DP0 pid=9735) ERROR 09-28 20:05:29 [core.py:712] ~~~~~~~~~~~^^^^^^
(EngineCore_DP0 pid=9735) ERROR 09-28 20:05:29 [core.py:712] KeyError: 'layers.0.mlp.experts.w2_weight'

now fixed!

[gemini-code-assist]

Code Review

This pull request addresses a KeyError that occurs when loading Qwen3VLMoe models with pipeline parallelism (pp > 2). The issue stems from the is_pp_missing_parameter check being incorrectly placed within the logic for non-fused experts only, causing it to be skipped for fused experts. The fix correctly relocates this check to a common path before the branching logic for fused and non-fused experts. This ensures that weights for experts on other pipeline stages are correctly skipped, resolving the loading error. The change is logical, well-targeted, and effectively fixes the reported bug.

[Isotr0py]

Thanks for fixing!