某种设备在运行pytorch算子时,如果不支持会自动fallback到cpu,输出的tensor.device却不是cpu,我希望能获取到这个状态。本文通过捕获标准输出,根据终端是否输出fallback字符串,判断是否触发了fallback

一.代码

import threading
import sys
import os

class CheckFallback:
    def __init__(self,enable=True):        
        self.is_fallback=False
        self.enable=enable
        if self.enable:
            self.stdout_fileno_origin = sys.stdout.fileno()
            self.stdout_fileno_dup = os.dup(self.stdout_fileno_origin)
            self.stdout_pipe = os.pipe()
            os.dup2(self.stdout_pipe[1], self.stdout_fileno_origin)
            os.close(self.stdout_pipe[1])
            self.stdout_messages = ''
            self.running=True
            self.task = threading.Thread(target=self.read_pipe)
            self.task.start()

    def read_pipe(self):
        while self.running:
            msg = os.read(self.stdout_pipe[0], 8192)
            if msg:
                self.stdout_messages+=msg.decode('utf-8')
    
    def __enter__(self):
        return self

    def __exit__(self, exc_type, exc_val, exc_tb):
        if self.enable:
            self.running=False
            os.close(self.stdout_fileno_origin)
            self.task.join()
            os.close(self.stdout_pipe[0])
            os.dup2(self.stdout_fileno_dup, self.stdout_fileno_origin)
            os.close(self.stdout_fileno_dup)
            #检查终端是否有fallback信息输出
            if self.stdout_messages.find("fallback")>-0:
                self.is_fallback=True

import torch
A=torch.ones((512,65024),dtype=torch.float16).to("your_device")
with CheckFallback() as f:
    C=torch.ops.aten.gelu.default(A)    
print(f.is_fallback)
print(C.shape,C.device)

with CheckFallback() as f:
    A=torch.ones((1,32),dtype=torch.float16).to("your_device")
    C=torch.ops.aten.pow(A,A)
print(f.is_fallback)
print(C.shape,C.device)