Issue evidence · Serve · huggingface/text-generation-inference

Zero config not working for VLMs

GitHub issue: https://github.com/huggingface/text-generation-inference/issues/3181

Not reproduced

TGI 3.2.3 started with only --model-id fails to serve Qwen/Qwen2.5-VL-3B-Instruct.

What Badgr ran

Ran ghcr.io/huggingface/text-generation-inference:3.2.3 with --model-id Qwen/Qwen2.5-VL-3B-Instruct on an RTX 4090 through Badgr, then sent a 256x256 image plus text to /v1/chat/completions.

What came back

  • The server became healthy (/health 200).
  • The image request returned HTTP 200 with a one-sentence description of the image.

The reporter used an A6000 and their failure text was not read here, so this shows it works on a 4090, not that their setup is fine. A vLLM serve for the same model was attempted but never started on the hosts available.

Checked 2026-10-14 in Badgr’s local development environment.

← All issue evidence