使用 ForkingPickler 在 NPU 序列化 tensor 报错

一、问题现象（附报错日志上下文）：
使用 ForkingPickler 在 NPU 序列化 tensor 报错

```
import io
import pybase64
import torch
from multiprocessing.reduction import ForkingPickler

class MultiprocessingSerializer:
    @staticmethod
    def serialize(tensor: torch.Tensor, output_str: bool = False):
        buf = io.BytesIO()

# Prefer GPU/NPU if available, else use CPU
        if torch.cuda.is_available() and tensor.device.type == "cuda":
            ForkingPickler(buf).dump(tensor)
        elif torch.npu.is_available() and tensor.device.type == "npu":
            ForkingPickler(buf).dump(tensor)
        else:
            # Fallback to CPU tensor
            ForkingPickler(buf).dump(tensor.detach().cpu())

buf.seek(0)
        output = buf.read()

if output_str:
            output = pybase64.b64encode(output).decode("utf-8")

return output

@staticmethod
    def deserialize(data, device: str = None):
        if isinstance(data, str):
            data = pybase64.b64decode(data, validate=True)

buf = io.BytesIO(data)
        tensor = ForkingPickler(buf).load()

if device is not None:
            tensor = tensor.to(device)

return tensor

if __name__ == "__main__":
    if torch.cuda.is_available():
        device = "cuda"
    elif torch.npu.is_available():
        device = "npu"
    else:
        device = "cpu"

x = torch.randn(2, 3, device=device)
    print("Original:", x, x.device)

serialized = MultiprocessingSerializer.serialize(x)

recovered = MultiprocessingSerializer.deserialize(serialized, device=device)
    print("Recovered:", recovered, recovered.device)
```
尝试使用 ForkingPickler 序列化 tensor 在 NPU 上传输，出现如下问题：

```
Traceback (most recent call last):
  File "/pathto/pickler_test.py", line 54, in <module>
    serialized = MultiprocessingSerializer.serialize(x)
  File "/pathto/pickler_test.py", line 16, in serialize
    ForkingPickler(buf).dump(tensor)
  File "/pathto/miniconda3/envs/roll/lib/python3.10/site-packages/torch/multiprocessing/reductions.py", line 618, in reduce_storage
    fd, size = storage._share_fd_cpu_()
  File "/pathto/miniconda3/envs/roll/lib/python3.10/site-packages/torch/storage.py", line 437, in wrapper
    return fn(self, *args, **kwargs)
  File "/pathto/miniconda3/envs/roll/lib/python3.10/site-packages/torch/storage.py", line 512, in _share_fd_cpu_
    return super()._share_fd_cpu_(*args, **kwargs)
RuntimeError: _share_fd_: only available on CPU
```
与https://gitee.com/ascend/pytorch/issues/IBLJAA?from=project-issue&search_text=_share_fd_ 问题基本一致
请问除了放在 CPU 上传输，是否有其他解决方案

Ascend/pytorch
暂停

内容风险标识

评论 (1)

Ascend/pytorch暂停 .gitee-modal { width: 500px !important; }

内容风险标识