sgl-project/sglang · error · ValueError

LoRA is not enabled. Please set `--enable-lora` to enable Lo

Error message

LoRA is not enabled. Please set `--enable-lora` to enable LoRA.

What it means

The /load_lora_adapter HTTP/control API refuses to load a LoRA adapter because the server was not started with --enable-lora, so no LoRA support is initialized.

Source

Thrown at python/sglang/srt/managers/tokenizer_control_mixin.py:613

        # Initiate the actual unloading operation at the backend processes only after all
        # ongoing requests using this LoRA adapter are finished.
        await self.lora_registry.wait_for_unload(lora_id)
        result = _merge_lora_update_results(
            await self.update_lora_adapter_communicator(obj)
        )

        return result

    async def load_lora_adapter(
        self: TokenizerManager,
        obj: LoadLoRAAdapterReqInput,
        _: Optional[fastapi.Request] = None,
    ) -> LoadLoRAAdapterReqOutput:
        self.auto_create_handle_loop()

        try:
            if not get_lora().enable_lora:
                raise ValueError(
                    "LoRA is not enabled. Please set `--enable-lora` to enable LoRA."
                )

            assert (
                get_parallel().dp_size == 1 or get_parallel().enable_dp_attention
            ), "dp_size must be 1 or dp attention must be enabled for dynamic lora loading"
            logger.info(
                "Start load Lora adapter. Lora name=%s, path=%s",
                obj.lora_name,
                obj.lora_path,
            )

            async with self.lora_update_lock:
                # Generate new uniquely identifiable LoRARef object.
                new_adapter = LoRARef(
                    lora_name=obj.lora_name,
                    lora_path=obj.lora_path,
                    pinned=obj.pinned,

View on GitHub (pinned to 0132848349)

Solutions

1. Relaunch the server with --enable-lora (and set --max-lora-rank/--lora-modules as needed)
2. Verify get_lora().enable_lora in your runtime config before issuing the request

Example fix

# before
python -m sglang.launch_server --model-path ...
# after
python -m sglang.launch_server --model-path ... --enable-lora --max-lora-rank 64

resp = requests.get(server + '/get_server_info')
if not resp.json().get('enable_lora'):
    skip_or_fail('server lacks --enable-lora')

try:
    await client.load_lora_adapter(req)
except ValueError as e:
    if 'enable-lora' in str(e):
        raise ConfigError('relaunch server with --enable-lora') from e
    raise

When it happens

Trigger: Calling load_lora_adapter on a server launched without --enable-lora (or with enable_lora false in runtime config).

Common situations: Forgetting the launch flag when adding dynamic LoRA loading to a serving stack; config presets that drop the flag.

Related errors

• PEFT adapter_config.json must contain a JSON object
• PEFT lora_alpha must be a positive integer
• LoRA adapter {lora_ref.lora_name} with rank {lora_config.r}
• Failed to load LoRA adapter {lora_ref.lora_name} as a pinned
• Streaming sessions are disabled. Please relaunch with --enab

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/98f852d56cb9b0d9. Report an issue: GitHub.