diff --git a/vllm/distributed/device_communicators/xpu_communicator.py b/vllm/distributed/device_communicators/xpu_communicator.py index 58c64e7920..e094f6fd50 100644 --- a/vllm/distributed/device_communicators/xpu_communicator.py +++ b/vllm/distributed/device_communicators/xpu_communicator.py @@ -99,12 +99,24 @@ class XpuCommunicator(DeviceCommunicatorBase): output = torch.empty( output_shape, dtype=input_tensor.dtype, device=input_tensor.device ) - if sizes is not None and sizes.count(sizes[0]) != len(sizes): - # if inputs shape in different ranks is not the same using reduce_scatter + if sizes is None or sizes.count(sizes[0]) == len(sizes): + dist.reduce_scatter_tensor(output, input_tensor, group=self.device_group) + elif torch.xpu.device_count() < self.world_size: + # uneven reduce_scatter is not supported when the + # process cannot see every device in the group ( + # e.g. external-LBDP). + max_size = max(sizes) + padded = input_tensor.new_zeros( + (world_size * max_size,) + input_tensor.shape[1:] + ) + for rank, split in enumerate(input_tensor.split(sizes, dim=0)): + padded[rank * max_size : rank * max_size + sizes[rank]] = split + padded_output = input_tensor.new_empty((max_size,) + input_tensor.shape[1:]) + dist.reduce_scatter_tensor(padded_output, padded, group=self.device_group) + output = padded_output[:chunk_size] + else: input_splits = list(input_tensor.split(sizes, dim=0)) dist.reduce_scatter(output, input_splits, group=self.device_group) - else: - dist.reduce_scatter_tensor(output, input_tensor, group=self.device_group) # Reshape before returning return output.movedim(0, dim).contiguous()