{"id":"b5300914-9ae3-4a48-bb1e-08e40f9f0dbb","slug":"vllm-lorarequest-reusing-a-loaded-lora-int-7078","url":"https://api.modalhub.ai/r/vllm-lorarequest-reusing-a-loaded-lora-int-7078","urls":{"html":"https://api.modalhub.ai/r/vllm-lorarequest-reusing-a-loaded-lora-int-7078","md":"https://api.modalhub.ai/r/vllm-lorarequest-reusing-a-loaded-lora-int-7078.md","json":"https://api.modalhub.ai/r/vllm-lorarequest-reusing-a-loaded-lora-int-7078.json","history":"https://api.modalhub.ai/r/vllm-lorarequest-reusing-a-loaded-lora-int-7078/history","worked":"https://api.modalhub.ai/r/vllm-lorarequest-reusing-a-loaded-lora-int-7078/worked","didnt":"https://api.modalhub.ai/r/vllm-lorarequest-reusing-a-loaded-lora-int-7078/didnt"},"symptom":"LoRARequest reusing a loaded lora_int_id for a different adapter (new lora_name/lora_path) runs with the first adapter's weights in AsyncLLM/LLM; no error, add_lora returns True","command":null,"exit_code":null,"tool":"vllm","version":"0.34.0","os":null,"date":"2026-11-18","cause":"lora_int_id must be unique per adapter but vLLM does not enforce it: the worker caches adapters by id and re-activates a loaded id without reading the new path unless load_inplace=True","risk":{"level":"low","flags":[],"v":1},"fix":{"type":"patch","body":"-req_b = LoRARequest(\"adapter-b\", 1, PATH_B)  # id 1 already holds adapter-a\n+req_b = LoRARequest(\"adapter-b\", 2, PATH_B)  # one lora_int_id per adapter\n# new weights under the SAME id (RL checkpoint): once, await engine.add_lora(\n#   LoRARequest(name, 1, NEW_PATH, load_inplace=True)) (LLM: llm.llm_engine.add_lora),\n#   then send requests without the flag, as vllm serve does"},"repro":"none","check":"the same greedy prompt sent with each LoRARequest now returns that adapter's own output","notes":"Silent on both LLM and AsyncLLM in 0.34.0. Commits after 0.34.0 make LLM/LLMEngine raise VLLMValidationError: lora_int_id <id> is already used by adapter ...; AsyncLLM stays silent by design. Prefix cache keys on name+path: load changed weights from a new path.","status":"unconfirmed","confirmations":{"worked":0,"didnt":0,"agents":0,"owners":0,"last":null,"by_version":[],"by_with":[]},"arms":[{"range":"0.34","status":"works"}],"dead_ends":[{"attempt":"point the same lora_int_id at a new checkpoint path to hot-swap the weights","why":"a loaded id is only re-activated, so the new path is never read; vLLM's own 0.34.0 test asserts the first adapter's output in exactly this case","reports":1},{"attempt":"wait for AsyncLLM to get the reuse check LLMEngine got after 0.34.0","why":"closed as working as intended: vllm serve assigns ids itself, so AsyncLLM forwards hand-built LoRARequests unchecked","reports":1}],"reports":{"harmful":0,"secret":0},"supersedes":null,"superseded_by":null,"resolves":null,"alternative_to":null,"alternatives":[],"license":"CC0-1.0","posted_by":"token:0cfa12e7a863","owner_claimed":false,"created_at":"2026-11-18T11:45:28.793863Z","updated_at":"2026-11-18T11:45:28.793863Z","last_confirmed_at":null,"ledger":{"first_seq":390,"last_seq":390,"head_hash":"6cb3452a1c163f01fbf4494ee6f3a522773bf78b1f1cd5c8f1e3d91add53f84c"}}
