vLLM#52617: LMCache P/D receiver reports full hit but decoding diverges
forum
Fields
id
sources/src5-gotcha-vllm-52617-lmcache-pd-divergencePublisher
vLLM
Kind
forum
Published
2026-08-17
Accessed
2026-10-04
notes
LMCache P/D receiver reports complete KV-cache hit but decoding diverges from monolithic baseline at token 0. First generated token differs, complete output differs, logprob differences reach 2.58-3.61. NIXL and Mooncake connectors work correctly. Demonstrates KV cache corruption in disaggregated serving.
Notes
LMCache P/D receiver reports complete KV-cache hit but decoding diverges from monolithic baseline at token 0. First generated token differs, complete output differs, logprob differences reach 2.58-3.61. NIXL and Mooncake connectors work correctly. Demonstrates KV cache corruption in disaggregated serving.
Cross-references
This record references no other records.