vLLM#52627: Kimi-K3 silent corruption with Mooncake+NIXL MultiConnector

Fields

id

sources/src5-gotcha-vllm-52627-kimi-k3-mooncake-corruption

URL

https://github.com/vllm-project/vllm/issues/52627

Publisher

vLLM

Kind

forum

Published

2026-08-17

Accessed

2026-10-04

notes

Kimi-K3 TP=8 1P1D deployment with MooncakeStoreConnector+NixlConnector shows intermittent silent output corruption. NIXL-only is stable. Corruption appears with Mooncake enabled, even without speculative decoding. Suspected: hybrid cache interaction between Kimi-K3, MooncakeStoreConnector, MultiConnector, and KDA recurrent state. Demonstrates KV cache corruption with external cache.

Notes

Kimi-K3 TP=8 1P1D deployment with MooncakeStoreConnector+NixlConnector shows intermittent silent output corruption. NIXL-only is stable. Corruption appears with Mooncake enabled, even without speculative decoding. Suspected: hybrid cache interaction between Kimi-K3, MooncakeStoreConnector, MultiConnector, and KDA recurrent state. Demonstrates KV cache corruption with external cache.

Cross-references

This record references no other records.