Skip to content

vllm.distributed.device_communicators.xpu_communicator

XpuCommunicator

Bases: DeviceCommunicatorBase

Source code in vllm/distributed/device_communicators/xpu_communicator.py 
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
class XpuCommunicator(DeviceCommunicatorBase):

    def __init__(self,
                 cpu_group: ProcessGroup,
                 device: Optional[torch.device] = None,
                 device_group: Optional[ProcessGroup] = None,
                 unique_name: str = ""):
        super().__init__(cpu_group, device, device_group, unique_name)

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

    def gather(self,
               input_: torch.Tensor,
               dst: int = 0,
               dim: int = -1) -> Optional[torch.Tensor]:
        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()
        # For xpu path, gather doesn't work properly together with ray
        # cluster so we use all_gather instead for now.
        input_size = input_.size()
        # Allocate output tensor.
        output_tensor = torch.empty((self.world_size, ) + input_size,
                                    dtype=input_.dtype,
                                    device=input_.device)
        # All-gather.
        dist.all_gather_into_tensor(output_tensor,
                                    input_,
                                    group=self.device_group)
        if self.rank_in_group == dst:
            # Reshape
            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:])
        else:
            output_tensor = None
        return output_tensor

    def broadcast(self, input_: torch.Tensor, src: int = 0) -> None:
        dist.broadcast(input_, src=src, group=self.device_group)

__init__ 

__init__(
    cpu_group: ProcessGroup,
    device: Optional[device] = None,
    device_group: Optional[ProcessGroup] = None,
    unique_name: str = "",
)
Source code in vllm/distributed/device_communicators/xpu_communicator.py 
15
16
17
18
19
20
def __init__(self,
             cpu_group: ProcessGroup,
             device: Optional[torch.device] = None,
             device_group: Optional[ProcessGroup] = None,
             unique_name: str = ""):
    super().__init__(cpu_group, device, device_group, unique_name)

all_reduce 

all_reduce(input_) -> Tensor
Source code in vllm/distributed/device_communicators/xpu_communicator.py 
22
23
24
def all_reduce(self, input_) -> torch.Tensor:
    dist.all_reduce(input_, group=self.device_group)
    return input_

broadcast 

broadcast(input_: Tensor, src: int = 0) -> None
Source code in vllm/distributed/device_communicators/xpu_communicator.py 
57
58
def broadcast(self, input_: torch.Tensor, src: int = 0) -> None:
    dist.broadcast(input_, src=src, group=self.device_group)

gather 

gather(
    input_: Tensor, dst: int = 0, dim: int = -1
) -> Optional[Tensor]
Source code in vllm/distributed/device_communicators/xpu_communicator.py 
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
def gather(self,
           input_: torch.Tensor,
           dst: int = 0,
           dim: int = -1) -> Optional[torch.Tensor]:
    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()
    # For xpu path, gather doesn't work properly together with ray
    # cluster so we use all_gather instead for now.
    input_size = input_.size()
    # Allocate output tensor.
    output_tensor = torch.empty((self.world_size, ) + input_size,
                                dtype=input_.dtype,
                                device=input_.device)
    # All-gather.
    dist.all_gather_into_tensor(output_tensor,
                                input_,
                                group=self.device_group)
    if self.rank_in_group == dst:
        # Reshape
        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:])
    else:
        output_tensor = None
    return output_tensor