Class VLLMInProcessWeightReloadAdapter

rl_engine/executors/bridge.py:327–433

Install a manifest through vLLM's in-process `reload_weights` utility path. This adapter is for single-process vLLM deployments, for example vLLM V1 with `VLLM_ENABLE_V1_MULTIPROCESSING=0`. It is a real hot-weight install path, but it is not CUDA IPC zero-copy: tensors are passed d

class VLLMInProcessWeightReloadAdapter:
    """
    Install a manifest through vLLM's in-process `reload_weights` utility path.

    This adapter is for single-process vLLM deployments, for example vLLM V1
    with `VLLM_ENABLE_V1_MULTIPROCESSING=0`. It is a real hot-weight install
    path, but it is not CUDA IPC zero-copy: tensors are passed directly to the
    in-process worker and vLLM performs the model reload/copy into GPU weights.
    Multiprocess vLLM should use the IPC or NCCL public `update_weights` APIs
    once those transports are validated on the target hardware.
    """

    def __init__(
        self,
        engine: Any,
        *,
        target_dtype: Optional[torch.dtype] = None,
        target_device: Optional[torch.device | str] = None,
        is_checkpoint_format: bool = True,
        synchronize_cuda: bool = True,
    ):
        self.engine = engine
        self.target_dtype = target_dtype
        self.target_device = target_device
        self.is_checkpoint_format = bool(is_checkpoint_format)
        self.synchronize_cuda = bool(synchronize_cuda)
        self.active_weight_version: Optional[int] = None
        self.active_update_id: Optional[str] = None

    def install(
        self,
        manifest: WeightUpdateManifest,
        tensors: Mapping[str, torch.Tensor],
    ) -> None:
        WeightLayout.from_metadata(manifest.metadata).validate_supported()
        tensor_map = dict(tensors)
        if set(tensor_map) != set(manifest.tensors):
            missing = sorted(set(manifest.tensors) - set(tensor_map))
            extra = sorted(set(tensor_map) - set(manifest.tensors))
            raise WeightManifestValidationError(
                f"vLLM reload tensor set mismatch: missing={missing}, extra={extra}"
            )

        weights: list[tuple[str, torch.Tensor]] = []
        for name, descriptor in manifest.tensors.items():
            tensor = tensor_map[name]
            if tuple(int(dim) for dim in tensor.shape) != descriptor.shape:
                raise WeightManifestValidationError(
                    f"vLLM reload tensor shape mismatch for {name}: "
                    f"expected {descriptor.shape}, got {tuple(tensor.shape)}"
                )
            if str(tensor.dtype) != descriptor.dtype:
                raise WeightManifestValidationError(
                    f"vLLM reload tensor dtype mismatch for {name}: "
                    f"expected {descriptor.dtype}, got {tensor.dtype}"
                )

            weight = tensor.detach()