sgl-project/sglang · critical · RuntimeError

Failed to load LoRA adapter {lora_ref.lora_name}: {result.er

Error message

Failed to load LoRA adapter {lora_ref.lora_name}: {result.error_message}

What it means

Raised during LoRAManager.init_lora_adapters (called from init_state at startup) when an adapter listed in startup lora_paths fails to load and the internal _load_lora_adapter result reports success=False. It wraps the underlying error_message from the load attempt (weight conversion, file IO, config parse, etc.).

Source

Thrown at python/sglang/srt/lora/lora_manager.py:612

    def init_lora_adapters(self, lora_paths: Optional[List[LoRARef]] = None):
        # Configs of all active LoRA adapters, indexed by LoRA ID.
        self.configs: Dict[str, LoRAConfig] = {}

        # LoRA adapter weights cached in CPU memory, indexed by LoRA ID.
        self.loras: Dict[str, LoRAAdapter] = {}

        # Mapping from LoRA ID to LoRARef object.
        self.lora_refs: Dict[str, LoRARef] = {}

        # Count of pinned LoRA adapters.
        self.num_pinned_loras: int = 0

        if lora_paths:
            for lora_ref in lora_paths:
                result = self._load_lora_adapter(lora_ref)
                if not result.success:
                    raise RuntimeError(
                        f"Failed to load LoRA adapter {lora_ref.lora_name}: {result.error_message}"
                    )

    def _detect_shared_outer_loras(self) -> bool:
        """Auto-detect shared outer LoRA format from loaded adapter weights.

        MoE adapters with shared outer experts store 3D tensors where
        dim[0]=1 indicates weights shared across all experts, while
        dim[0]=num_experts indicates per-expert weights.
        Returns True if gate_up lora_A has expert_dim=1 (shared).

        All loaded adapters that expose a 3D gate_up lora_A must agree;
        mixed formats raise RuntimeError.
        """
        shared_outer: Optional[bool] = None
        for adapter_id, adapter in self.loras.items():
            for layer in adapter.layers:
                for name, weight in layer.weights.items():

View on GitHub (pinned to 0132848349)

Solutions

1. Check the wrapped result.error_message in the traceback to identify the root cause
2. Verify the adapter path exists and contains a valid PEFT adapter (adapter_config.json + safetensors)
3. Fix or remove the offending entry from --lora-paths and restart
4. Confirm the adapter is plain LoRA (not DoRA) and its rank/targets match server flags

import os
for name, path in lora_paths:
    assert os.path.isdir(path) and os.path.exists(os.path.join(path, 'adapter_config.json')), f"bad lora path: {path}"

try:
    manager.init_lora_adapters(lora_paths)
except RuntimeError as e:
    log.error("startup LoRA load failed: %s", e)  # e carries the inner error_message
    raise SystemExit(1)

When it happens

Trigger: Passing --lora-paths (or the equivalent lora_paths list) at server startup where at least one path is invalid, missing adapter_config.json/weights, or fails CUDA-side weight preparation, making result.success false.

Common situations: Typo in the adapter path in server args, an adapter saved in an unsupported format (DoRA, non-PEFT layout), or an OOM/corruption during weight loading at startup.

Related errors

• SGLANG_RUST_SERVER is not supported with the offline Engine
• Multi-node weight cache daemons (nnodes > 1) require --dist-
• Weight cache daemon for pp_rank={pp_rank} tp_rank={tp_rank}
• Weight cache daemon (pid={p.pid}) exited prematurely with co
• engine_info_bootstrap_port {bootstrap_port} is already in us

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/b35ec79cedd828ce. Report an issue: GitHub.