Issue evidence · Serve · vllm-project/vllm

NIXL failed KV transfers are silently treated as loaded for HMA models (invalid blocks never reported)

GitHub issue: https://github.com/vllm-project/vllm/issues/61616

Confirmed from source (not run)

For hybrid-KV-cache models, a failed NIXL receive is not reported as invalid blocks, so the scheduler decodes from unwritten KV.

What Badgr ran

Read vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py inside the cu134-nightly image (vLLM 0.32.0rc1.dev100). Not run: no failed transfer was injected.

What came back

  • Line 2912 holds "TODO (NickLucche) handle failed transfer for HMA." and line 2913 guards the invalid-block queueing with "if not self._is_hma_required", the code the issue describes.

Code reading only (not run): the silent wrong output was not demonstrated.

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

← All issue evidence