PR #39877: fix(qwen3-vl): reject oversized video placeholders early

• Repository: vllm-project/vllm
• Author: T4mako
• State: closed | merged: False
• Link: https://github.com/vllm-project/vllm/pull/39877

Description (problem / solution / changelog)

Purpose

• #39876: reject excessively large video placeholders as early as possible.

I deployed qwen3.5-27b using vllm for a video summary task, my start shell：

nohup ./.venv/bin/vllm serve "$MODEL_DIR" \
  --host "$HOST" \
  --port "$PORT" \
  --served-model-name "Qwen3.5-27B" \
  --tensor-parallel-size 2 \
  --dtype float16 \
  --reasoning-parser "qwen3" \
  --gpu-memory-utilization "0.9" \
  --max-model-len 260000 \
  --enable-auto-tool-choice \
  --tool-call-parser "$TOOL_CALL_PARSER_VALUE" \
  --override-generation-config "$OVERRIDE_GENERATION_CONFIG_VALUE" \
  --api-key "$VLLM_API_KEY_VALUE \
  >> "$LOG_FILE" 2>&1 < /dev/null &

but encountered the following log:

Apr 06 03:13:54 ubuntu start_vllm_qwen.sh[873947]: (APIServer pid=873947) Token indices sequence length is longer than the specified maximum sequence length for this model (422640 > 262144). Running this sequence through the model will result in indexing errors.

After that, all other text reasoning tasks timed out and there are no request logs.

I believe it's necessary to reject excessively large video placeholders as early as possible.

Test Plan

Test Result

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

Changed files

• vllm/model_executor/models/qwen3_vl.py (modified, +9/-0)