sgl-project/sglang · error · ValueError

SGLang does not recognize target_modules='{config.target_mod

Error message

SGLang does not recognize target_modules='{config.target_modules}'. Please use a list of module name suffixes in the adapter's PEFT config, or explicitly specify --lora-target-modules during server startup.

What it means

Raised in init_lora_shapes when an adapter's PEFT config specifies target_modules in a form SGLang cannot interpret (not a recognizable list of suffixes and not matching a known mapping) and no explicit --lora-target-modules was given. SGLang only auto-infers target modules from a list of module-name suffixes in the PEFT config.

Source

Thrown at python/sglang/srt/lora/lora_manager.py:697

                        # per-adapter inference for this adapter.
                        continue
                    else:
                        # Resolve by scanning the base model for all
                        # LoRA-compatible linear modules.
                        adapter_target_modules = auto_detect_lora_target_modules(
                            self.base_model
                        )
                        logger.info(
                            "LoRA adapter '%s' uses target_modules='%s'. "
                            "Resolved to %s by inspecting the base model.",
                            self.lora_refs[lora_id].lora_name,
                            config.target_modules,
                            sorted(adapter_target_modules),
                        )
                        self.target_modules.update(adapter_target_modules)
                        continue
                else:
                    raise ValueError(
                        f"SGLang does not recognize target_modules="
                        f"'{config.target_modules}'. Please use a list of module "
                        "name suffixes in the adapter's PEFT config, or explicitly "
                        "specify --lora-target-modules during server startup."
                    )

            if not isinstance(config.target_modules, list):
                raise ValueError(
                    f"SGLang currently only supports inferring LoRA target modules when a list of "
                    "suffixes is provided in `target_modules` field of PEFT config. Please explicitly "
                    "specify `--lora-target-modules` during server startup. You can specify `all` to "
                    "enable all support modules types. "
                )

            adapter_target_modules = get_normalized_target_modules(
                config.target_modules
            )

View on GitHub (pinned to 0132848349)

Solutions

1. Start the server with explicit --lora-target-modules listing the suffixes (or 'all')
2. Edit the adapter's adapter_config.json to store an explicit list of module suffixes like ['q_proj','v_proj']

Example fix

# before
--enable-lora  # adapter config has target_modules=".*q_proj"
# after
--enable-lora --lora-target-modules q_proj v_proj o_proj

import json
cfg = json.load(open(f'{path}/adapter_config.json'))
tm = cfg.get('target_modules')
if not isinstance(tm, list):
    raise SystemExit('start server with --lora-target-modules, or fix adapter config')

def has_list_target_modules(cfg) -> bool:
    tm = cfg.get('target_modules')
    return isinstance(tm, list) and all(isinstance(x, str) for x in tm)

When it happens

Trigger: Loading an adapter whose adapter_config.json has target_modules as a regex string (e.g. '.*q_proj|.*v_proj') or a keyword like 'all-linear', while the server was started without --lora-target-modules.

Common situations: Adapters trained with PEFT defaults that store a regex in target_modules; moving an adapter from vLLM/transformers to SGLang without adjusting config.

Understand the failure class

Background: Config validation failed: what "invalid value for {key}" and settings-rejection errors mean across 19 open-source libraries — this error's family across 19 libraries.

Related errors

• SGLang currently only supports inferring LoRA target modules
• LoRA adapter '{lora_name}' contains target modules {sorted(u
• DSA indexer weights_proj LoRA is incompatible with piecewise
• Failed to load LoRA adapter {lora_ref.lora_name} because it
• LoRA adapter {lora_ref.lora_name} with rank {lora_config.r}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/8d886e7111599287. Report an issue: GitHub.