AI API receipt · vllm-project/vllm

grouped_topk selects wrong experts with more than eight groups

vllm-project/vllm#60752 · upstream bug reproduced by running the issue's own reproduction

Badgr cannot tell this response is wrong

With 9+ expert groups the kernel never scores groups 8 and above, so MoE routing picks wrong experts.

wrong expert routing lowers answer quality without breaking the format; the reply looks valid. The receipt therefore records success. This is a limit of what the contract can prove, shown here so it is not mistaken for protection.

What the upstream returned (replayed)

The output the bug produces, taken from the issue’s reproduction and fed to Badgr’s validators. Not a live provider call.

text: "The result is approximately forty two, give or take."
tool_calls: none

Badgr receipt

Written by Badgr’s receipt ledger and read back from GET /v1/receipts/replay-vllm-60752-grouped-topk-wrong-experts.

{
  "request_id": "replay-vllm-60752-grouped-topk-wrong-experts",
  "status": "success",
  "endpoint": "/v1/chat/completions",
  "stream": false,
  "failure_stage": null,
  "error_code": null,
  "error_message": null,
  "retryable": null,
  "provider_mode": "replay",
  "routing_decision": "replayed_upstream_output",
  "created_at": "1791524108.89506"
}

Recorded 2026-10-09 in Badgr’s local development environment. Replay cases and verdicts are pinned by backend/test_ai_api_replay_cases.py.

← All AI API receipts