---
title: '54634'
pageLayout: default
slug: thoughts/inference/vllm-prs/54634
permalink: https://aarnphm.xyz/thoughts/inference/vllm-prs/54634.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
# [#54634](https://github.com/aarnphm/aarnphm.github.io/issues/54634) \[K3 Bug\] Fix Kimi-K3 RecoverSSM startup failure `'MambaAttentionBackendEnum.GDN_ATTN declares 4 states, but provides 2 state copy funcs'`

[vllm-project/vllm#54634](https://github.com/vllm-project/vllm/pull/54634) · yewentao256 · MERGED · created 2026-08-31T21:14:59Z · merged 2026-08-31T23:22:49Z

## Purpose

`vllm serve moonshotai/Kimi-K3   --trust-remote-code   --tensor-parallel-size 8 --
load-format fastsafetensors   --gpu-memory-utilization 0.85   --enable-p
refix-caching   --use-replayssm   --reasoning-parser kimi_k3   --enable-
auto-tool-choice   --tool-call-parser kimi_k3   --host 0.0.0.0   --port
30000   --speculative-config '{"model":"RadixArk/Kimi-K3-DSpark","method
":"dspark","num_speculative_tokens":7,"draft_sample_method":"probabilist
ic","rejection_sample_method":"standard"}'`



```bash
(EngineCore pid=3344247)   File "/home/yewentao256/vllm-source/vllm/v1/executor/multiproc_executor.py", line 437, in get_response
(EngineCore pid=3344247)     raise RuntimeError(
(EngineCore pid=3344247) RuntimeError: Worker failed with error 'MambaAttentionBackendEnum.GDN_ATTN declares 4 states, but provides 2 state copy funcs', please check the stack trace above for the root cause
```

RecoverSSM adds two temporary workspace tensors after its two persistent states. These workspace tensors are rewritten during every speculative verification step and do not require state migration.

This PR fixes the issue.

---

## comments

**yewentao256** (2026-08-31T21:15:13Z):

/ci run

**github-actions** (2026-08-31T21:15:31Z):

✅ Triggered [Buildkite CI #86426](https://buildkite.com/vllm/ci/builds/86426) for commit `ae31c043bff7`.

