When i request a longer audio , return {'error': {'message': 'The decoder prompt contains a(n) audio item with length 3476, which exceeds the pre-allocated encoder cache size 2048. Please reduce the input size or increase the encoder cache size by setting --limit-mm-per-prompt at startup.', 'type': 'BadRequestError', 'param': None, 'code': 400}}，

TL;DR

Increase the --limit-mm-per-prompt parameter value to allow for longer audio inputs.

Guidance

• The error message indicates that the audio input length exceeds the pre-allocated encoder cache size, which is currently set to 2048.
• To fix this, you need to increase the --limit-mm-per-prompt parameter value, specifically the audio.length property, to a value greater than 3476.
• You can do this by modifying the command to include a higher value for --limit-mm-per-prompt, for example: --limit-mm-per-prompt '{"audio":{"count":1,"length":65536}}'.
• Verify that the new value allows for longer audio inputs without exceeding the maximum allowed length.

Example

python3 -m vllm.entrypoints.openai.api_server \
  --model /Qwen3-ASR-1.7B \
  --host 0.0.0.0 \
  --port 30000 \
  --served-model-name qwen3-asr \
  --max-model-len "65536" \
  --limit-mm-per-prompt '{"audio":{"count":1,"length":65536}}'

Notes

The optimal value for --limit-mm-per-prompt will depend on the specific requirements of your application and the available resources.

Recommendation

Apply workaround: Increase the --limit-mm-per-prompt parameter value to allow for longer audio inputs, as this will resolve the immediate issue and allow for further testing and evaluation.