sgl-project/sglang · error · RuntimeError

Error raised in subprocess: {returned.stderr.decode()}

Error message

Error raised in subprocess:
{returned.stderr.decode()}

What it means

registry.inspect_model_cls loads candidate model classes in a subprocess to avoid polluting the parent process; this RuntimeError wraps any non-zero exit of that subprocess and surfaces its stderr. The real failure is whatever the child process printed (import error, CUDA init error, etc.).

Source

Thrown at python/sglang/multimodal_gen/runtime/models/registry.py:233

    # issues like https://stackoverflow.com/questions/23212435/permission-denied-to-write-to-my-temporary-file
    with tempfile.TemporaryDirectory() as tempdir:
        output_filepath = os.path.join(tempdir, "registry_output.tmp")

        # `cloudpickle` allows pickling lambda functions directly
        input_bytes = cloudpickle.dumps((fn, output_filepath))

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
            raise RuntimeError(
                f"Error raised in subprocess:\n" f"{returned.stderr.decode()}"
            ) from e

        with open(output_filepath, "rb") as f:
            return cast(_T, pickle.load(f))


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """

    module_name: str
    class_name: str

    # Performed in another process to avoid initializing CUDA
    def inspect_model_cls(self) -> _ModelInfo:

View on GitHub (pinned to 0132848349)

Solutions

1. Read the stderr text embedded in the message — it contains the child's traceback; fix that root cause first
2. Reproduce manually: python -c "import <model_module>" in the same environment
3. Install missing optional dependencies (pip install the package named in the child traceback)
4. If the model fails only under subprocess isolation (e.g. env vars, CUDA_VISIBLE_DEVICES), make import-time code lazy/conditional

import subprocess
r = subprocess.run([sys.executable, '-c', f'import {module}'], capture_output=True)
if r.returncode != 0:
    raise ImportError(f'cannot import {module}: {r.stderr.decode()}')

try:
    arch_info = registry.inspect_model_cls(config)
except RuntimeError as e:
    stderr = str(e)
    log.error('subprocess failure: %s', stderr)
    raise  # fix the root cause in the child traceback, don't retry blindly

When it happens

Trigger: Calling inspect_model_cls for an architecture whose module fails to import in a clean subprocess — missing optional dependency, syntax/attribute error in the model file, driver/CUDA unavailable, or heavy import-time side effects crashing the child.

Common situations: Adding a new model whose file imports a package not in the environment; model code that imports torch.cuda at module scope on CPU-only machines; registry inspection running in a sandboxed/minimal CI container where flash-attn or diffusers extras are absent.

Related errors

• Failed to load serve backend {name!r} from {self._entry_poin
• Can not import FA3 in sgl_kernel. Please check your installa
• {_METALLIB_NAME} not found next to the native Metal extensio
• scalar_type_id {scalar_type_id} doesn't exists.
• FlashAttention-4 CUTE is not available. Install flash-attn-4

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/8112397a0ca79beb. Report an issue: GitHub.