Issue evidence · Serve · vllm-project/vllm

GLM-5.3-Flash with --mm-encoder-tp-mode data gives wrong image and video answers

GitHub issue: https://github.com/vllm-project/vllm/issues/61681

Not run

With multimodal encoder data parallelism across tensor-parallel ranks, GLM-5.3-Flash answers about images and video are mixed up.

What Badgr ran

Nothing was run: the report is for a specific B200 TP=2 setup with a large model, which did not fit the $5 cap.

What came back

  • The issue includes a 4,450-character script that was not executed.

A funded B200 pair would be needed.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence