vllm.distributed ¶

class DeviceCommunicatorBase:
    """
    Base class for device-specific communicator.
    It can use the `cpu_group` to initialize the communicator.
    If the device has PyTorch integration (PyTorch can recognize its
    communication backend), the `device_group` will also be given.
    """

    def __init__(
        self,
        cpu_group: ProcessGroup,
        device: torch.device | None = None,
        device_group: ProcessGroup | None = None,
        unique_name: str = "",
    ):
        self.device = device or torch.device("cpu")
        self.cpu_group = cpu_group
        self.device_group = device_group
        self.unique_name = unique_name
        self.rank = dist.get_rank(cpu_group)
        self.world_size = dist.get_world_size(cpu_group)
        self.ranks = dist.get_process_group_ranks(cpu_group)
        self.global_rank = dist.get_rank()
        self.global_world_size = dist.get_world_size()
        self.rank_in_group = dist.get_group_rank(self.cpu_group, self.global_rank)

        use_ep = False
        all2all_backend = None
        from vllm.config import get_current_vllm_config

        config = get_current_vllm_config()
        if config is not None:
            # as long as we use data parallel (coupled data parallel
            # where all data parallel ranks execute forward together),
            # we initialize the all2all manager used in expert parallel.
            use_ep = config.parallel_config.data_parallel_size > 1
            all2all_backend = config.parallel_config.all2all_backend

        self.is_ep_communicator = "ep" in unique_name
        self.use_all2all = self.is_ep_communicator and use_ep
        self.all2all_backend = all2all_backend
        self.all2all_manager: All2AllManagerBase | None = None

    def all_reduce(self, input_: torch.Tensor) -> torch.Tensor:
        dist.all_reduce(input_, group=self.device_group)
        return input_

    def all_gather(self, input_: torch.Tensor, dim: int = -1) -> torch.Tensor:
        if dim < 0:
            # Convert negative dim to positive.
            dim += input_.dim()
        input_size = input_.size()
        # NOTE: we have to use concat-style all-gather here,
        # stack-style all-gather has compatibility issues with
        # torch.compile . see https://github.com/pytorch/pytorch/issues/138795
        output_size = (input_size[0] * self.world_size,) + input_size[1:]
        # Allocate output tensor.
        output_tensor = torch.empty(
            output_size, dtype=input_.dtype, device=input_.device
        )
        # All-gather.
        dist.all_gather_into_tensor(output_tensor, input_, group=self.device_group)
        # Reshape
        output_tensor = output_tensor.reshape((self.world_size,) + input_size)
        output_tensor = output_tensor.movedim(0, dim)
        output_tensor = output_tensor.reshape(
            input_size[:dim]
            + (self.world_size * input_size[dim],)
            + input_size[dim + 1 :]
        )
        return output_tensor

    def all_gatherv(
        self,
        input_: torch.Tensor | list[torch.Tensor],
        dim: int = 0,
        sizes: list[int] | None = None,
    ) -> torch.Tensor | list[torch.Tensor]:
        raise NotImplementedError

    def reduce_scatter(self, input_: torch.Tensor, dim: int = -1) -> torch.Tensor:
        world_size = self.world_size
        # Bypass the function if we are using only 1 GPU.
        if world_size == 1:
            return input_
        assert -input_.dim() <= dim < input_.dim(), (
            f"Invalid dim ({dim}) for input tensor with shape {input_.size()}"
        )

        if dim < 0:
            # Convert negative dim to positive.
            dim += input_.dim()

        # Note: This will produce an incorrect answer if we don't make
        # the input_tensor contiguous. Possible bug in reduce_scatter_tensor?
        input_tensor = input_.movedim(0, dim).contiguous()

        assert input_tensor.shape[0] % world_size == 0
        chunk_size = input_tensor.shape[0] // world_size
        output_shape = (chunk_size,) + input_tensor.shape[1:]

        output_tensor = torch.empty(
            output_shape, dtype=input_tensor.dtype, device=input_tensor.device
        )

        # Perform reduce-scatter operation
        torch.distributed.reduce_scatter_tensor(
            output_tensor, input_tensor, group=self.device_group
        )

        # Reshape before returning
        return output_tensor.movedim(0, dim).contiguous()

    def reduce_scatterv(
        self, input_: torch.Tensor, dim: int = -1, sizes: list[int] | None = None
    ) -> torch.Tensor:
        raise NotImplementedError

    def gather(
        self, input_: torch.Tensor, dst: int = 0, dim: int = -1
    ) -> torch.Tensor | None:
        """
        NOTE: We assume that the input tensor is on the same device across
        all the ranks.
        NOTE: `dst` is the local rank of the destination rank.
        """
        world_size = self.world_size
        assert -input_.dim() <= dim < input_.dim(), (
            f"Invalid dim ({dim}) for input tensor with shape {input_.size()}"
        )
        if dim < 0:
            # Convert negative dim to positive.
            dim += input_.dim()

        # Allocate output tensor.
        if self.rank_in_group == dst:
            gather_list = [torch.empty_like(input_) for _ in range(world_size)]
        else:
            gather_list = None
        # Gather.
        torch.distributed.gather(
            input_, gather_list, dst=self.ranks[dst], group=self.device_group
        )
        if self.rank_in_group == dst:
            output_tensor = torch.cat(gather_list, dim=dim)
        else:
            output_tensor = None
        return output_tensor

    def send(self, tensor: torch.Tensor, dst: int | None = None) -> None:
        """Sends a tensor to the destination rank in a blocking way"""
        """NOTE: `dst` is the local rank of the destination rank."""
        if dst is None:
            dst = (self.rank_in_group + 1) % self.world_size
        torch.distributed.send(tensor, self.ranks[dst], self.device_group)

    def recv(
        self, size: torch.Size, dtype: torch.dtype, src: int | None = None
    ) -> torch.Tensor:
        """Receives a tensor from the source rank."""
        """NOTE: `src` is the local rank of the source rank."""
        if src is None:
            src = (self.rank_in_group - 1) % self.world_size

        tensor = torch.empty(size, dtype=dtype, device=self.device)
        torch.distributed.recv(tensor, self.ranks[src], self.device_group)
        return tensor

    def destroy(self):
        pass

    def prepare_communication_buffer_for_model(self, model: torch.nn.Module) -> None:
        """
        Prepare the communication buffer for the model.
        """
        if not self.is_ep_communicator:
            return

        moe_modules = [
            module
            for module in model.modules()
            # TODO(bnell): Should use isinstance but can't.  Maybe search for
            # presence of quant_method.maybe_init_modular_kernel?
            if (
                module.__class__.__name__ == "FusedMoE"
                or module.__class__.__name__ == "SharedFusedMoE"
            )
        ]
        for module in moe_modules:
            module.maybe_init_modular_kernel()

    def dispatch(
        self,
        hidden_states: torch.Tensor,
        router_logits: torch.Tensor,
        is_sequence_parallel: bool = False,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Dispatch the hidden states and router logits to the appropriate device.
        This is a no-op in the base class.
        """
        return hidden_states, router_logits

    def combine(
        self, hidden_states: torch.Tensor, is_sequence_parallel: bool = False
    ) -> torch.Tensor:
        """
        Combine the hidden states and router logits from the appropriate device.
        This is a no-op in the base class.
        """
        return hidden_states

@dataclass
class GraphCaptureContext:
    stream: torch.cuda.Stream

class GroupCoordinator:
    """
    PyTorch ProcessGroup wrapper for a group of processes.
    PyTorch ProcessGroup is bound to one specific communication backend,
        e.g. NCCL, Gloo, MPI, etc.
    GroupCoordinator takes charge of all the communication operations among
        the processes in the group. It manages both CPU and device
        communication.
    """

    # available attributes:
    rank: int  # global rank
    ranks: list[int]  # global ranks in the group
    world_size: int  # size of the group
    # difference between `local_rank` and `rank_in_group`:
    # if we have a group of size 4 across two nodes:
    # Process | Node | Rank | Local Rank | Rank in Group
    #   0     |   0  |  0   |     0      |       0
    #   1     |   0  |  1   |     1      |       1
    #   2     |   1  |  2   |     0      |       2
    #   3     |   1  |  3   |     1      |       3
    local_rank: int  # local rank used to assign devices
    rank_in_group: int  # rank inside the group
    cpu_group: ProcessGroup  # group for CPU communication
    device_group: ProcessGroup  # group for device communication
    # device communicator (if use_device_communicator=True)
    device_communicator: DeviceCommunicatorBase | None
    mq_broadcaster: Any | None  # shared memory broadcaster

    def __init__(
        self,
        group_ranks: list[list[int]],
        local_rank: int,
        torch_distributed_backend: str | Backend,
        use_device_communicator: bool,  # whether to use device communicator
        use_message_queue_broadcaster: bool = False,
        group_name: str | None = None,
    ):
        group_name = group_name or "anonymous"
        self.unique_name = _get_unique_name(group_name)
        _register_group(self)

        self.rank = torch.distributed.get_rank()
        self.local_rank = local_rank

        self_device_group = None
        self_cpu_group = None

        for ranks in group_ranks:
            device_group = torch.distributed.new_group(
                ranks, backend=torch_distributed_backend
            )
            # a group with `gloo` backend, to allow direct coordination between
            # processes through the CPU.
            cpu_group = torch.distributed.new_group(ranks, backend="gloo")
            if self.rank in ranks:
                self.ranks = ranks
                self.world_size = len(ranks)
                self.rank_in_group = ranks.index(self.rank)
                self_device_group = device_group
                self_cpu_group = cpu_group

        assert self_cpu_group is not None
        assert self_device_group is not None

        self.cpu_group = self_cpu_group
        self.device_group = self_device_group

        from vllm.platforms import current_platform

        if current_platform.is_cuda_alike():
            self.device = torch.device(f"cuda:{local_rank}")
        elif current_platform.is_xpu():
            self.device = torch.device(f"xpu:{local_rank}")
        elif current_platform.is_out_of_tree():
            self.device = torch.device(f"{current_platform.device_name}:{local_rank}")
        else:
            self.device = torch.device("cpu")

        self.use_device_communicator = use_device_communicator
        self.device_communicator = None
        if use_device_communicator and self.world_size > 1:
            device_comm_cls = resolve_obj_by_qualname(
                current_platform.get_device_communicator_cls()
            )
            self.device_communicator = device_comm_cls(
                cpu_group=self.cpu_group,
                device=self.device,
                device_group=self.device_group,
                unique_name=self.unique_name,
            )

        from vllm.distributed.device_communicators.shm_broadcast import MessageQueue

        self.mq_broadcaster: MessageQueue | None = None
        if use_message_queue_broadcaster and self.world_size > 1:
            self.mq_broadcaster = MessageQueue.create_from_process_group(
                self.cpu_group, 1 << 22, 6
            )

        from vllm.platforms import current_platform

        self.use_custom_op_call = (
            current_platform.is_cuda_alike() or current_platform.is_tpu()
        )

        self.use_cpu_custom_send_recv = current_platform.is_cpu() and hasattr(
            torch.ops._C, "init_shm_manager"
        )

    def create_mq_broadcaster(
        self, writer_rank=0, external_writer_handle=None, blocking=True
    ):
        from vllm.distributed.device_communicators.shm_broadcast import MessageQueue

        return MessageQueue.create_from_process_group(
            self.cpu_group,
            1 << 22,
            6,
            writer_rank=writer_rank,
            external_writer_handle=external_writer_handle,
            blocking=blocking,
        )

    def create_single_reader_mq_broadcasters(
        self, reader_rank_in_group=0, blocking=False
    ):
        from vllm.distributed.device_communicators.shm_broadcast import MessageQueue

        return MessageQueue.create_from_process_group_single_reader(
            self.cpu_group,
            1 << 22,
            6,
            reader_rank=self.ranks[reader_rank_in_group],
            blocking=blocking,
        )

    @property
    def first_rank(self):
        """Return the global rank of the first process in the group"""
        return self.ranks[0]

    @property
    def last_rank(self):
        """Return the global rank of the last process in the group"""
        return self.ranks[-1]

    @property
    def is_first_rank(self):
        """Return whether the caller is the first process in the group"""
        return self.rank == self.first_rank

    @property
    def is_last_rank(self):
        """Return whether the caller is the last process in the group"""
        return self.rank == self.last_rank

    @property
    def next_rank(self):
        """Return the global rank of the process that follows the caller"""
        rank_in_group = self.rank_in_group
        world_size = self.world_size
        return self.ranks[(rank_in_group + 1) % world_size]

    @property
    def prev_rank(self):
        """Return the global rank of the process that precedes the caller"""
        rank_in_group = self.rank_in_group
        world_size = self.world_size
        return self.ranks[(rank_in_group - 1) % world_size]

    @contextmanager
    def graph_capture(self, graph_capture_context: GraphCaptureContext | None = None):
        if graph_capture_context is None:
            stream = torch.cuda.Stream()
            graph_capture_context = GraphCaptureContext(stream)
        else:
            stream = graph_capture_context.stream

        # only cuda uses this function,
        # so we don't abstract it into the base class
        maybe_ca_context = nullcontext()
        from vllm.distributed.device_communicators.cuda_communicator import (
            CudaCommunicator,
        )

        if self.device_communicator is not None:
            assert isinstance(self.device_communicator, CudaCommunicator)
            ca_comm = self.device_communicator.ca_comm
            if ca_comm is not None:
                maybe_ca_context = ca_comm.capture()  # type: ignore

        # ensure all initialization operations complete before attempting to
        # capture the graph on another stream
        curr_stream = torch.cuda.current_stream()
        if curr_stream != stream:
            stream.wait_stream(curr_stream)

        with torch.cuda.stream(stream), maybe_ca_context:
            yield graph_capture_context

    def all_reduce(self, input_: torch.Tensor) -> torch.Tensor:
        """
        User-facing all-reduce function before we actually call the
        all-reduce operation.

        We need this because Dynamo does not support passing an arbitrary
        object (`self` in this case) to a custom op. We need to pass the
         group name as a string, and then look up the group coordinator from
         the group name, dispatch the all-reduce operation to the group
         coordinator.

        In addition, PyTorch custom ops do not support mutation or returning
        a new tensor in the same op. So we always make the all-reduce operation
        out-of-place.
        """
        # Bypass the function if we are using only 1 GPU.
        if self.world_size == 1:
            return input_

        if self.use_custom_op_call:
            return torch.ops.vllm.all_reduce(input_, group_name=self.unique_name)
        else:
            return self._all_reduce_out_place(input_)

    def _all_reduce_out_place(self, input_: torch.Tensor) -> torch.Tensor:
        if self.device_communicator is None:
            raise ValueError("No device communicator found")
        return self.device_communicator.all_reduce(input_)

    def all_gather(self, input_: torch.Tensor, dim: int = -1) -> torch.Tensor:
        world_size = self.world_size
        # Bypass the function if we are using only 1 GPU.
        if world_size == 1:
            return input_
        assert -input_.dim() <= dim < input_.dim(), (
            f"Invalid dim ({dim}) for input tensor with shape {input_.size()}"
        )

        if self.use_custom_op_call:
            return torch.ops.vllm.all_gather(
                input_, dim, world_size, group_name=self.unique_name
            )
        else:
            return self._all_gather_out_place(input_, dim)

    def _all_gather_out_place(self, input_: torch.Tensor, dim: int) -> torch.Tensor:
        if self.device_communicator is None:
            raise ValueError("No device communicator found")
        return self.device_communicator.all_gather(input_, dim)

    def all_gatherv(
        self,
        input_: torch.Tensor | list[torch.Tensor],
        dim: int = 0,
        sizes: list[int] | None = None,
    ):
        if self.device_communicator is None:
            raise ValueError("No device communicator found")
        return self.device_communicator.all_gatherv(input_, dim, sizes)

    def reduce_scatter(self, input_: torch.Tensor, dim: int = -1) -> torch.Tensor:
        world_size = self.world_size
        # Bypass the function if we are using only 1 GPU.
        if world_size == 1:
            return input_
        assert -input_.dim() <= dim < input_.dim(), (
            f"Invalid dim ({dim}) for input tensor with shape {input_.size()}"
        )

        if self.use_custom_op_call:
            return torch.ops.vllm.reduce_scatter(
                input_, dim, world_size, group_name=self.unique_name
            )
        else:
            return self._reduce_scatter_out_place(input_, dim)

    def reduce_scatterv(
        self, input_: torch.Tensor, dim: int = -1, sizes: list[int] | None = None
    ) -> torch.Tensor:
        if self.device_communicator is None:
            raise ValueError("No device communicator found")
        return self.device_communicator.reduce_scatterv(input_, dim, sizes)

    def _reduce_scatter_out_place(self, input_: torch.Tensor, dim: int) -> torch.Tensor:
        if self.device_communicator is None:
            raise ValueError("No device communicator found")
        return self.device_communicator.reduce_scatter(input_, dim)

    def gather(
        self, input_: torch.Tensor, dst: int = 0, dim: int = -1
    ) -> torch.Tensor | None:
        """
        NOTE: We assume that the input tensor is on the same device across
        all the ranks.
        NOTE: `dst` is the local rank of the destination rank.
        """
        world_size = self.world_size
        # Bypass the function if we are using only 1 GPU.
        if world_size == 1:
            return input_
        if self.device_communicator is None:
            raise ValueError("No device communicator found")
        return self.device_communicator.gather(input_, dst, dim)

    def broadcast(self, input_: torch.Tensor, src: int = 0):
        """Broadcast the input tensor.
        NOTE: `src` is the local rank of the source rank.
        """
        assert src < self.world_size, f"Invalid src rank ({src})"

        # Bypass the function if we are using only 1 GPU.
        if self.world_size == 1:
            return input_
        # Broadcast.
        torch.distributed.broadcast(
            input_, src=self.ranks[src], group=self.device_group
        )
        return input_

    def broadcast_object(self, obj: Any | None = None, src: int = 0):
        """Broadcast the input object.
        NOTE: `src` is the local rank of the source rank.
        """
        assert src < self.world_size, f"Invalid src rank ({src})"

        # Bypass the function if we are using only 1 GPU.
        if self.world_size == 1:
            return obj
        if self.mq_broadcaster is not None:
            assert src == 0, "Message queue broadcaster only supports src=0"
            return self.mq_broadcaster.broadcast_object(obj)
        if self.rank_in_group == src:
            torch.distributed.broadcast_object_list(
                [obj], src=self.ranks[src], group=self.cpu_group
            )
            return obj
        else:
            recv = [None]
            torch.distributed.broadcast_object_list(
                recv, src=self.ranks[src], group=self.cpu_group
            )
            return recv[0]

    def broadcast_object_list(
        self, obj_list: list[Any], src: int = 0, group: ProcessGroup | None = None
    ):
        """Broadcast the input object list.
        NOTE: `src` is the local rank of the source rank.
        """
        assert src < self.world_size, f"Invalid src rank ({src})"

        # Bypass the function if we are using only 1 GPU.
        if self.world_size == 1:
            return obj_list
        # Broadcast.
        torch.distributed.broadcast_object_list(
            obj_list, src=self.ranks[src], group=self.device_group
        )
        return obj_list

    def send_object(self, obj: Any, dst: int) -> None:
        """Send the input object list to the destination rank."""
        """NOTE: `dst` is the local rank of the destination rank."""

        assert dst < self.world_size, f"Invalid dst rank ({dst})"

        assert dst != self.rank_in_group, (
            "Invalid destination rank. Destination rank is the same "
            "as the current rank."
        )

        # Serialize object to tensor and get the size as well
        object_tensor = torch.frombuffer(pickle.dumps(obj), dtype=torch.uint8)

        size_tensor = torch.tensor(
            [object_tensor.numel()], dtype=torch.long, device="cpu"
        )

        # Send object size

        torch.distributed.send(size_tensor, dst=self.ranks[dst], group=self.cpu_group)

        # Send object
        torch.distributed.send(object_tensor, dst=self.ranks[dst], group=self.cpu_group)

        return None

    def recv_object(self, src: int) -> Any:
        """Receive the input object list from the source rank."""
        """NOTE: `src` is the local rank of the source rank."""

        assert src < self.world_size, f"Invalid src rank ({src})"

        assert src != self.rank_in_group, (
            "Invalid source rank. Source rank is the same as the current rank."
        )

        size_tensor = torch.empty(1, dtype=torch.long, device="cpu")

        # Receive object size
        rank_size = torch.distributed.recv(
            size_tensor, src=self.ranks[src], group=self.cpu_group
        )

        # Tensor to receive serialized objects into.
        object_tensor = torch.empty(  # type: ignore[call-overload]
            size_tensor.item(),  # type: ignore[arg-type]
            dtype=torch.uint8,
            device="cpu",
        )

        rank_object = torch.distributed.recv(
            object_tensor, src=self.ranks[src], group=self.cpu_group
        )

        assert rank_object == rank_size, (
            "Received object sender rank does not match the size sender rank."
        )

        obj = pickle.loads(object_tensor.numpy().tobytes())

        return obj

    def broadcast_tensor_dict(
        self,
        tensor_dict: dict[str, torch.Tensor | Any] | None = None,
        src: int = 0,
        group: ProcessGroup | None = None,
        metadata_group: ProcessGroup | None = None,
    ) -> dict[str, torch.Tensor | Any] | None:
        """Broadcast the input tensor dictionary.
        NOTE: `src` is the local rank of the source rank.
        """
        # Bypass the function if we are using only 1 GPU.
        if not torch.distributed.is_initialized() or self.world_size == 1:
            return tensor_dict

        group = self.device_group
        metadata_group = self.cpu_group
        assert src < self.world_size, f"Invalid src rank ({src})"

        rank_in_group = self.rank_in_group
        if rank_in_group == src:
            metadata_list: list[tuple[Any, Any]] = []
            assert isinstance(tensor_dict, dict), (
                f"Expecting a dictionary, got {type(tensor_dict)}"
            )
            metadata_list, tensor_list = _split_tensor_dict(tensor_dict)
            # `metadata_list` lives in CPU memory.
            # `broadcast_object_list` has serialization & deserialization,
            # all happening on CPU. Therefore, we can use the CPU group.
            self.broadcast_object(metadata_list, src=src)
            async_handles = []
            for tensor in tensor_list:
                if tensor.numel() == 0:
                    # Skip broadcasting empty tensors.
                    continue
                if tensor.is_cpu:
                    # use metadata_group for CPU tensors
                    handle = torch.distributed.broadcast(
                        tensor, src=self.ranks[src], group=metadata_group, async_op=True
                    )
                else:
                    # use group for GPU tensors
                    handle = torch.distributed.broadcast(
                        tensor, src=self.ranks[src], group=group, async_op=True
                    )
                async_handles.append(handle)
            for async_handle in async_handles:
                async_handle.wait()

        else:
            metadata_list = self.broadcast_object(None, src=src)
            tensor_dict = {}
            async_handles = []
            for key, value in metadata_list:
                if isinstance(value, TensorMetadata):
                    tensor = torch.empty(
                        value.size, dtype=value.dtype, device=value.device
                    )
                    if tensor.numel() == 0:
                        # Skip broadcasting empty tensors.
                        tensor_dict[key] = tensor
                        continue
                    if tensor.is_cpu:
                        # use metadata_group for CPU tensors
                        handle = torch.distributed.broadcast(
                            tensor,
                            src=self.ranks[src],
                            group=metadata_group,
                            async_op=True,
                        )
                    else:
                        # use group for GPU tensors
                        handle = torch.distributed.broadcast(
                            tensor, src=self.ranks[src], group=group, async_op=True
                        )
                    async_handles.append(handle)
                    tensor_dict[key] = tensor
                else:
                    tensor_dict[key] = value
            for async_handle in async_handles:
                async_handle.wait()
        return tensor_dict

    def send_tensor_dict(
        self,
        tensor_dict: dict[str, torch.Tensor | Any],
        dst: int | None = None,
        all_gather_group: Optional["GroupCoordinator"] = None,
        all_gather_tensors: dict[str, bool] | None = None,
    ) -> dict[str, torch.Tensor | Any] | None:
        """Send the input tensor dictionary.
        NOTE: `dst` is the local rank of the source rank.

        all_gather_group: The group for the all-gather operation. If provided,
            an optimization is enabled where each rank in the group sends a
            slice of a tensor and the receiver reconstructs it using an
            all-gather, which can improve performance. This is typically the
            tensor-parallel group.
        all_gather_tensors: A dictionary to specify which tensors should use
            the all-gather optimization, which is only effective when
            `all_gather_group` is provided. By default, this optimization is
            on for any tensor whose size is divisible by the
            `all_gather_group`'s world size. However, it should be disabled
            for tensors that are not fully replicated across the group (e.g.,
            the residual tensor when sequence parallelism is enabled). This
            dictionary allows overriding the default behavior on a per-tensor
            basis.
        """
        # Bypass the function if we are using only 1 GPU.
        if not torch.distributed.is_initialized() or self.world_size == 1:
            return tensor_dict
        all_gather_size = 1 if all_gather_group is None else all_gather_group.world_size
        all_gather_rank = (
            0 if all_gather_group is None else all_gather_group.rank_in_group
        )

        group = self.device_group
        metadata_group = self.cpu_group

        if dst is None:
            dst = (self.rank_in_group + 1) % self.world_size
        assert dst < self.world_size, f"Invalid dst rank ({dst})"

        if self.use_cpu_custom_send_recv:
            if self.device_communicator is None:
                raise ValueError("No device communicator found")
            self.device_communicator.send_tensor_dict(  # type: ignore
                tensor_dict, dst
            )
            return None

        metadata_list: list[tuple[Any, Any]] = []
        assert isinstance(tensor_dict, dict), (
            f"Expecting a dictionary, got {type(tensor_dict)}"
        )
        metadata_list, tensor_list = _split_tensor_dict(tensor_dict)
        # `metadata_list` lives in CPU memory.
        # `send_object_list` has serialization & deserialization,
        # all happening on CPU. Therefore, we can use the CPU group.
        self.send_object(metadata_list, dst=dst)

        tensor_keys = [k for k, v in tensor_dict.items() if isinstance(v, torch.Tensor)]
        assert len(tensor_keys) == len(tensor_list)

        for key, tensor in zip(tensor_keys, tensor_list):
            if tensor.numel() == 0:
                # Skip sending empty tensors.
                continue

            # send-allgather: send only a slice, then do allgather.
            use_all_gather = (
                all_gather_group is not None and tensor.numel() % all_gather_size == 0
            )
            use_all_gather = (
                all_gather_tensors.get(key, use_all_gather)
                if all_gather_tensors
                else use_all_gather
            )
            if use_all_gather:
                tensor = tensor.reshape(all_gather_size, -1)[all_gather_rank]

            if tensor.is_cpu:
                # use metadata_group for CPU tensors
                torch.distributed.send(
                    tensor, dst=self.ranks[dst], group=metadata_group
                )
            else:
                # use group for GPU tensors
                torch.distributed.send(tensor, dst=self.ranks[dst], group=group)
        return None

    def recv_tensor_dict(
        self,
        src: int | None = None,
        all_gather_group: Optional["GroupCoordinator"] = None,
        all_gather_tensors: dict[str, bool] | None = None,
    ) -> dict[str, torch.Tensor | Any] | None:
        """Recv the input tensor dictionary.
        NOTE: `src` is the local rank of the source rank.

        all_gather_group: The group for the all-gather operation. If provided,
            an optimization is enabled where each rank in the group sends a
            slice of a tensor and the receiver reconstructs it using an
            all-gather, which can improve performance. This is typically the
            tensor-parallel group.
        all_gather_tensors: A dictionary to specify which tensors should use
            the all-gather optimization, which is only effective when
            `all_gather_group` is provided. By default, this optimization is
            on for any tensor whose size is divisible by the
            `all_gather_group`'s world size. However, it should be disabled
            for tensors that are not fully replicated across the group (e.g.,
            the residual tensor when sequence parallelism is enabled). This
            dictionary allows overriding the default behavior on a per-tensor
            basis.
        """
        # Bypass the function if we are using only 1 GPU.
        if not torch.distributed.is_initialized() or self.world_size == 1:
            return None
        all_gather_size = 1 if all_gather_group is None else all_gather_group.world_size
        all_gather_rank = (
            0 if all_gather_group is None else all_gather_group.rank_in_group
        )

        group = self.device_group
        metadata_group = self.cpu_group

        if src is None:
            src = (self.rank_in_group - 1) % self.world_size
        assert src < self.world_size, f"Invalid src rank ({src})"

        if self.use_cpu_custom_send_recv:
            if self.device_communicator is None:
                raise ValueError("No device communicator found")
            return self.device_communicator.recv_tensor_dict(  # type: ignore
                src
            )

        recv_metadata_list = self.recv_object(src=src)
        tensor_dict: dict[str, Any] = {}
        for key, value in recv_metadata_list:
            if isinstance(value, TensorMetadata):
                tensor = torch.empty(value.size, dtype=value.dtype, device=value.device)
                if tensor.numel() == 0:
                    # Skip broadcasting empty tensors.
                    tensor_dict[key] = tensor
                    continue

                # send-allgather: send only a slice, then do allgather.
                use_all_gather = (
                    all_gather_group is not None
                    and tensor.numel() % all_gather_size == 0
                )
                use_all_gather = (
                    all_gather_tensors.get(key, use_all_gather)
                    if all_gather_tensors
                    else use_all_gather
                )

                if use_all_gather:
                    orig_shape = tensor.shape
                    tensor = tensor.reshape(all_gather_size, -1)[all_gather_rank]

                if tensor.is_cpu:
                    # use metadata_group for CPU tensors
                    torch.distributed.recv(
                        tensor, src=self.ranks[src], group=metadata_group
                    )
                else:
                    # use group for GPU tensors
                    torch.distributed.recv(tensor, src=self.ranks[src], group=group)
                if use_all_gather:
                    # do the allgather
                    tensor = all_gather_group.all_gather(  # type: ignore
                        tensor, dim=0
                    )
                    tensor = tensor.reshape(orig_shape)

                tensor_dict[key] = tensor
            else:
                tensor_dict[key] = value
        return tensor_dict

    def barrier(self):
        """Barrier synchronization among the group.
        NOTE: don't use `device_group` here! `barrier` in NCCL is
        terrible because it is internally a broadcast operation with
        secretly created GPU tensors. It is easy to mess up the current
        device. Use the CPU group instead.
        """
        torch.distributed.barrier(group=self.cpu_group)

    def send(self, tensor: torch.Tensor, dst: int | None = None) -> None:
        """Sends a tensor to the destination rank in a blocking way"""
        """NOTE: `dst` is the local rank of the destination rank."""
        if self.device_communicator is None:
            raise ValueError("No device communicator found")
        self.device_communicator.send(tensor, dst)

    def recv(
        self, size: torch.Size, dtype: torch.dtype, src: int | None = None
    ) -> torch.Tensor:
        """Receives a tensor from the source rank."""
        """NOTE: `src` is the local rank of the source rank."""
        if self.device_communicator is None:
            raise ValueError("No device communicator found")
        return self.device_communicator.recv(size, dtype, src)

    def destroy(self):
        if hasattr(self, "device_group"):
            torch.distributed.destroy_process_group(self.device_group)
            del self.device_group
        if hasattr(self, "cpu_group"):
            torch.distributed.destroy_process_group(self.cpu_group)
            del self.cpu_group
        if self.device_communicator is not None:
            self.device_communicator.destroy()
        if self.mq_broadcaster is not None:
            self.mq_broadcaster = None

    def prepare_communication_buffer_for_model(self, model: torch.nn.Module):
        if self.device_communicator is not None:
            self.device_communicator.prepare_communication_buffer_for_model(model)

    def dispatch(
        self,
        hidden_states: torch.Tensor,
        router_logits: torch.Tensor,
        is_sequence_parallel: bool = False,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if self.device_communicator is not None:
            return self.device_communicator.dispatch(
                hidden_states, router_logits, is_sequence_parallel
            )
        else:
            return hidden_states, router_logits

    def combine(
        self, hidden_states, is_sequence_parallel: bool = False
    ) -> torch.Tensor:
        if self.device_communicator is not None:
            return self.device_communicator.combine(hidden_states, is_sequence_parallel)
        else:
            return hidden_states