sgl-project/sglang · error · ValueError

Dit target weight name

Error message

Dit target weight name {target_name} already exists in lora_adapters[{lora_nickname}]

What it means

load_lora_adapter stores LoRA B weights keyed by target parameter name inside lora_adapters[nickname]; if two source entries in the checkpoint map to the same target_name (after SwiGLU fc1 name swapping), the second insert is rejected to avoid silently overwriting weights.

Solutions

1. Inspect the checkpoint keys (safetensors index) and remove the duplicate entry
2. Re-export the LoRA adapter from the original training framework
3. If this is a pipeline bug for your model arch, report it with the checkpoint key list

from safetensors import safe_open
with safe_open(path, framework='pt') as f:
    keys = list(f.keys())
assert len(keys) == len(set(keys)) and len(keys) <= expected_param_count

try:
    pipeline.set_lora(...)
except ValueError as e:
    if 'already exists in lora_adapters' in str(e): log_checkpoint_issue(path)

When it happens

Trigger: A LoRA checkpoint containing duplicate/aliased weight names, e.g. both the original and _swap_peft_swiglu_fc1_lora_b-renamed variants of the same fc1 lora_b tensor; loading such an adapter via set_lora.

Common situations: Exporting a PEFT adapter for a SwiGLU model that duplicates gate/up names; converting checkpoints between naming conventions that produce collisions; corrupted or re-saved adapters.

Related errors

• A GGUF encoder checkpoint cannot be combined with a second…
• A quantized checkpoint requires an in-tree native encoder…
• Action normalization stats not found at
• adapter_config.json lora_alpha conflicts with safetensors…
• Adapter not found in the pipeline. Please provide lora_path…

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/3ec1c942be02d72e. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/lora/pipeline.py:914

            if merge_index is not None:
                to_merge_params[target_name][merge_index] = weight
                # A/B of one fused layer must be laid out together (GQA B cannot stack).
                if target_name.endswith((".lora_A", ".lora_B")):
                    continue
                if len(to_merge_params[target_name]) == num_params_to_merge:
                    sorted_tensors = [
                        to_merge_params[target_name][i]
                        for i in range(num_params_to_merge)
                    ]
                    # Use stack instead of cat because it needs to be compatible with TP.
                    weight = torch.stack(sorted_tensors, dim=0)
                    del to_merge_params[target_name]
                else:
                    continue

            weight = _swap_peft_swiglu_fc1_lora_b(name, target_name, weight)
            if target_name in self.lora_adapters[lora_nickname]:
                raise ValueError(
                    f"Dit target weight name {target_name} already exists in lora_adapters[{lora_nickname}]"
                )
            self.lora_adapters[lora_nickname][target_name] = weight.to(self.device)

        _store_fused_lora_groups(
            self.lora_adapters[lora_nickname],
            to_merge_params,
            adapter_lora_alpha,
            self.device,
        )
        transformer = self.modules["transformer"]
        if isinstance(transformer, BaseDiT):
            self.lora_adapters[lora_nickname] = transformer.prepare_lora_adapter(
                self.lora_adapters[lora_nickname]
            )

        self.loaded_adapter_paths[lora_nickname] = lora_path
        self.loaded_adapter_alphas[lora_nickname] = adapter_lora_alpha

View on GitHub (pinned to 0132848349)