sgl-project/sglang · error · ValueError

Failed to implicitly load LoRA adapter {lora_path}: {load_re

Error message

Failed to implicitly load LoRA adapter {lora_path}: {load_result.error_message}

What it means

The server attempted an implicit load of a requested LoRA adapter and the load result reported failure (other than 'already loaded'). The tokenizer manager surfaces the underlying worker error message, e.g. missing files, incompatible rank/target modules, or OOM during weight load.

Source

Thrown at python/sglang/srt/managers/tokenizer_manager.py:3378

                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(
                    f"Failed to implicitly load LoRA adapter {lora_path}: {load_result.error_message}"
                )

        # Look up the LoRA ID from the registry and start tracking ongoing LoRA requests.
        obj.lora_id = await self.lora_registry.acquire(obj.lora_path)
        # Propagate lora_id to any sub-objects already cached by __getitem__.
        for i, sub_obj in obj.__dict__.get("_sub_obj_cache", {}).items():
            sub_obj.lora_id = (
                obj.lora_id[i] if isinstance(obj.lora_id, list) else obj.lora_id
            )

    def _init_req_state(
        self,
        obj: Union[GenerateReqInput, EmbeddingReqInput],
        request: Optional[fastapi.Request] = None,
    ):
        created_time = obj.received_time

View on GitHub (pinned to 0132848349)

Solutions

1. Check the appended load_result.error_message for the root cause (missing file, rank mismatch, OOM)
2. Verify the adapter path exists and matches the base model, then reload via /load_lora_adapter
3. Increase --max-lora-rank / free GPU memory / reduce concurrently loaded adapters
4. Fix the path in --lora-paths and restart the server

import os
for name, path in lora_paths.items():
    assert os.path.isdir(path) and os.path.exists(os.path.join(path, 'adapter_config.json')), path

try:
    out = await engine.generate(prompt, lora_path=p)
except ValueError as e:
    if "Failed to implicitly load LoRA adapter" in str(e):
        log.error("adapter load failed: %s", e); alert_ops(e)
    raise

When it happens

Trigger: Requesting lora_path that triggers implicit load while the adapter directory is missing, has an incompatible architecture (rank > max_lora_rank, unsupported target modules), or loading exhausts GPU memory.

Common situations: Wrong adapter path in --lora-paths or request; adapter trained on a different base model; --max-lora-rank too small; GPU memory exhausted by too many loaded adapters.

Related errors

• LoRA batch_info must provide max_len or seg_lens.
• LoRA batch_info must provide max_len or seg_lens.
• Initialization failed. Please see the error messages above.
• lora_nickname cannot be empty
• Failed to set LoRA adapter: {str(e)}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/a97719f8a25aa0c8. Report an issue: GitHub.