vLLM issue #30931: Prefix Cache Corruption with LoRA adapters sharing a name but differing in id

The adapter-swap instance of a stale prefix cache. Prefix cache blocks computed under one LoRA adapter are reused under another, because the block hash does not discriminate between adapters that share a name but differ in lora_init_id. Maintainer jeejeelee confirms the mechanism and scopes it: 'I understand this issue should only occur in offline mode, it shouldn't be a problem when online server'; the reporter confirms it was offline/research use. Resolution direction discussed: do not expose lora_init_id, PR #31489 to hide it. Env: A100-SXM4-80GB x4, vLLM around 0.11-era, PyTorch 2.9.0+cu128. Severity caveat matters - this one is scoped to offline/batch entrypoints, not the online server.