sgl-project/sglang · error · NotImplementedError

weight cache transport backend {VMM_FD_BACKEND!r} is not imp

Error message

weight cache transport backend {VMM_FD_BACKEND!r} is not implemented in this build

What it means

VmmFdTransportBackend (the 'vmm_fd' weight cache transport) is a stub in this build: every method raises NotImplementedError. Selected when the backend name is vmm_fd, but the feature was not compiled/implemented here.

Source

Thrown at python/sglang/srt/weight_cache/transport.py:154

        return MultiprocessingSerializer.deserialize(entry["handle"])


class VmmFdTransportBackend(WeightCacheTransportBackend):
    """Placeholder for the CUDA VMM + fd-passing transport.

    The backend is not wired up yet: can_export_state reports False so the
    daemon keeps selecting torch_ipc, and every other entry point fails loudly
    instead of silently returning None.
    """

    name = VMM_FD_BACKEND

    def __init__(self):
        self._raise_not_implemented()

    @staticmethod
    def _raise_not_implemented() -> NoReturn:
        raise NotImplementedError(
            f"weight cache transport backend {VMM_FD_BACKEND!r} is not "
            f"implemented in this build"
        )

    @classmethod
    def can_export_state(
        cls, state_tensors: Mapping[str, Tuple[torch.Tensor, bool]]
    ) -> bool:
        return False

    def prepare_export(
        self, state_tensors: Mapping[str, Tuple[torch.Tensor, bool]]
    ) -> Dict[str, Dict[str, Any]]:
        self._raise_not_implemented()

    def send_fetch_state_response(
        self,
        conn: socket.socket,

View on GitHub (pinned to 0132848349)

Solutions

1. Use the default torch IPC backend (unset the backend name)
2. Upgrade sglang to a build that implements vmm_fd

Example fix

# before
backend = get_client_transport_backend("vmm_fd")
# after
backend = get_client_transport_backend(None)  # torch IPC

from sglang.srt.weight_cache.transport import TORCH_IPC_BACKEND, VMM_FD_BACKEND
assert name in (None, TORCH_IPC_BACKEND) or implemented(name)

When it happens

Trigger: Setting the weight cache transport backend to 'vmm_fd' via config/env; any call to its __init__, prepare_export, send/recv_fetch_state_response, or import_tensor.

Common situations: Using a docs/blog config that references the experimental VMM FD backend on a release that lacks it.

Related errors

• [weight_cache:{where}] quantization method {quant_method!r}
• {env_field.name}={template!r} must contain '{{global_rank}}'
• Unknown weight cache transport backend {name!r}
• This browser cannot encode H.264 MP4
• H.264 encoder did not return MP4 decoder config

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/e37b011d7d818118. Report an issue: GitHub.