Issue evidence · Serve · vllm-project/vllm

NIXL handshake crashes with StopIteration for Mamba-only models (no FullAttention KV group)

GitHub issue: https://github.com/vllm-project/vllm/issues/61617

Reproduced

For a model with no attention KV group, the NIXL decode-side handshake raises StopIteration and no KV transfer can be set up.

What Badgr ran

Served state-spaces/mamba-130m-hf as a prefill (kv_producer) and a decode (kv_consumer) instance with NixlConnector on one RTX 4090 through Badgr (vllm/vllm-openai cu134-nightly, vLLM 0.32.0rc1.dev100, NIXL 1.4.1), with VLLM_SSM_CONV_STATE_LAYOUT=DS, and sent a prefill then a decode request carrying the returned kv_transfer_params.

What came back

  • The prefill request succeeded and returned remote_block_ids.
  • The decode request returned HTTP 500; the engine logged "NIXL transfer failure: handshake_failed" with StopIteration from _build_fa_remote (base_worker.py line 2016).

Without VLLM_SSM_CONV_STATE_LAYOUT=DS the engines refuse to start for Mamba NIXL transfers, so that setting was needed to reach the handshake. The reporter used an A100 on an earlier main commit.

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

← All issue evidence