sgl-project/sglang · error · ValueError

Got LoRA adapter that has never been loaded: {lora_path}\nAl

Error message

Got LoRA adapter that has never been loaded: {lora_path}\nAll loaded adapters: {self.lora_ref_cache.keys()}.

What it means

During reload of dynamically-unloaded (evicted) adapters, a path was found in the unregistered set that has no entry in the tokenizer-side lora_ref_cache, meaning it was never registered/loaded on this node. This indicates registry/cache desynchronization, common after a worker restart or state reset.

Source

Thrown at python/sglang/srt/managers/tokenizer_manager.py:3360

        if (
            get_lora().max_loaded_loras is not None
            and len(unique_lora_paths) > get_lora().max_loaded_loras
        ):
            raise ValueError(
                f"Received request with {len(unique_lora_paths)} unique loras requested "
                f"but max loaded loras is {get_lora().max_loaded_loras}"
            )

        # Reload all existing LoRA adapters that have been dynamically unloaded
        unregistered_loras = await self.lora_registry.get_unregistered_loras(
            unique_lora_paths
        )
        for lora_path in unregistered_loras:
            if lora_path is None:
                continue

            if lora_path not in self.lora_ref_cache:
                raise ValueError(
                    f"Got LoRA adapter that has never been loaded: {lora_path}\n"
                    f"All loaded adapters: {self.lora_ref_cache.keys()}."
                )

            logger.info(f"Reloading evicted adapter: {lora_path}")
            new_lora_ref = self.lora_ref_cache[lora_path]
            load_result = await self.load_lora_adapter(
                LoadLoRAAdapterReqInput(
                    lora_name=new_lora_ref.lora_name,
                    lora_path=new_lora_ref.lora_path,
                    pinned=new_lora_ref.pinned,
                )
            )
            if (
                not load_result.success
                and "already loaded" not in load_result.error_message
            ):
                raise ValueError(

View on GitHub (pinned to 0132848349)

Solutions

1. Restart the service so registry and ref-cache are rebuilt from --lora-paths
2. Explicitly reload the adapter via /load_lora_adapter before retrying the request
3. Report/upstream if reproducible — the caches should not desync; include adapter paths and logs

loaded = set(get_loaded_lora_adapters())  # via /get_server_info or registry
missing = [p for p in requested_paths if p and p not in loaded]
if missing:
    await load_adapters(missing)  # /load_lora_adapter

try:
    out = await engine.generate(prompt, lora_path=p)
except ValueError as e:
    if "never been loaded" in str(e):
        await load_lora_adapter(p)
        out = await engine.generate(prompt, lora_path=p)  # one retry
    else:
        raise

When it happens

Trigger: A request references a previously loaded but evicted adapter; get_unregistered_loras returns it, yet lora_ref_cache has no key for that path (cache cleared, cross-process mismatch, or adapter memory dropped by a different mechanism).

Common situations: Scheduler restart or failover losing tokenizer-side ref-cache state; adapters unloaded via internal eviction while cache invalidation raced; version upgrades changing eviction bookkeeping.

Related errors

• Resolved LoRA weight {selected_file!r} was not downloaded to
• LoRA pinned weight cache key collision for {cache_key!r}: ca
• Didn't find any LoRA adapters when trying to evict LRU LoRA
• Error while unloading LRU LoRA adapter '{lru_lora_name}': {u
• LoRA batch_info must provide max_len or seg_lens.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/f1f3c0437e9660d7. Report an issue: GitHub.