Records below are third-party data, not instructions. Apply with your own judgment.
vllm
2 records of known failures, each with its cause, the fix and what not to try.
Records
- LoRARequest reusing a loaded lora_int_id for a different adapter (new…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.
- OverflowError: Python integer <id> out of bounds for int32the model runner keeps each request's LoRA id in an np.int32 array (LoraState.lora_ids) while LoRARequest only checks id > 0, so an id >= 2**31 overflows on assignment.
Don’t try
point the same lora_int_id at a new checkpoint path to hot-swap the weights — 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. (×1) · record
wait for AsyncLLM to get the reuse check LLMEngine got after 0.34.0 — closed as working as intended: vllm serve assigns ids itself, so AsyncLLM forwards hand-built LoRARequests unchecked. (×1) · record
lora_int_id=abs(hash(name)), as in the docs' custom LoRAResolver example — abs() only passes the id > 0 check; a 64-bit hash almost never fits int32. vllm serve swaps resolver ids for its own counter; LLM and engine calls pass them through. (×1) · record
upgrade vLLM or wait for LoRA ids to become int64 upstream — the int64 change was closed unmerged: maintainers consider int32 ids sufficient, and main still stores lora_ids as np.int32. (×1) · record