sgl-project/sglang · error · ValueError

Received request with {len(unique_lora_paths)} unique loras

Error message

Received request with {len(unique_lora_paths)} unique loras requested but max loaded loras is {get_lora().max_loaded_loras}

What it means

A single request referenced more unique LoRA adapters than the server's max_loaded_loras limit. The tokenizer manager counts distinct paths in the request's lora_path list and rejects it if the count exceeds the configured maximum number of concurrently loaded adapters.

Source

Thrown at python/sglang/srt/managers/tokenizer_manager.py:3346

            raise ValueError(
                f"LoRA adapter '{first_adapter}' was requested, but LoRA is not enabled. "
                "Please launch the server with --enable-lora flag and preload adapters "
                "using --lora-paths or /load_lora_adapter endpoint."
            )

        await self._resolve_lora_path(obj)

    async def _resolve_lora_path(self, obj: Union[GenerateReqInput, EmbeddingReqInput]):
        if isinstance(obj.lora_path, str):
            unique_lora_paths = set([obj.lora_path])
        else:
            unique_lora_paths = set(obj.lora_path)

        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()}."
                )

View on GitHub (pinned to 0132848349)

Solutions

Relaunch the server with a higher --max-loras / --max-loaded-loras (and enough GPU memory via --lora-target-modules etc.)
Split the request so each generate call uses at most max_loaded_loras unique adapters
Deduplicate lora_path lists client-side before sending

Example fix

# before
python -m sglang.launch_server --model-path M --enable-lora --max-loras 1
# request: lora_path=["a","b"]

# after
python -m sglang.launch_server --model-path M --enable-lora --max-loras 4
# or split into two requests, one per adapter

Defensive patterns

Strategy: validation

Validate before calling

max_loaded = get_max_loaded_loras()  # from server info
uniq = {p for p in lora_paths if p}
if len(uniq) > max_loaded:
    lora_paths = split_or_reduce(uniq, max_loaded)

Type guard

def fits_lora_limit(paths: list, limit: int) -> bool:
    return len({p for p in paths if p}) <= limit

Try / catch

try:
    out = engine.generate(prompt, lora_path=paths)
except ValueError as e:
    if "max loaded loras" in str(e):
        # chunk the batch by unique adapter
        ...
    raise

Prevention

Set --max-loras to the maximum per-request adapter diversity you need
Deduplicate adapter paths before batching requests
Monitor adapter count per request in client middleware

When it happens

Trigger: Passing a list of lora_path values containing more unique adapters than --max-loaded-loras (derived from max_loras / lora capacity) in one generate request.

Common situations: Multi-LoRA batching with a long adapter list while server launched with default max_loras (often 1); raising client-side batch diversity without raising --max-loras/--max-loaded-loras; duplicate paths collapsing to fewer uniques than expected after refactoring.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/c7c6735e7a477a3d. Report an issue: GitHub.