A Questions about asynchronous saving

[JOjoker-world]

""" serializes the tensors (copies GPU tensors to CPU), and persists the snapshot to disk.

Fsync the file after writing to gaurantee persistence

Call this in a new process to perform in bgk
""" 
def _serialize_and_persist(
    self, 
    filepath,
    snapshot,
    active_snapshot,
    lock,
    linkpath=None,
    iter_chk = None,
    epoch_chk = None, 
    overwrite = True):
    print("[{}] START ASYNC".format(time.time()))

    with lock:
        if active_snapshot.value == 0:
            self.logger.error("Cannot persist. Empty snapshot")
            return
    #Create new stream
    s = torch.cuda.Stream()
    torch.cuda.stream(s)

    #print("Saving : {}".format(filepath))
    torch.save(snapshot, filepath)
    #print("Saved : {}".format(filepath))
    # Clear the snapshot.
    with lock:
        active_snapshot.value = 0

    # Ensure its persisted
    f = open(filepath, 'a+')
    os.fsync(f.fileno())
    f.close()

    update_stats(
            filepath,
            iter_chk=iter_chk,
            overwrite=overwrite,
            epoch_chk = epoch_chk,
            linkpath=linkpath)
    print("[{}] END ASYNC".format(time.time()))

**这一段代码中

Create new stream

s = torch.cuda.Stream() torch.cuda.stream(s) torch.save(snapshot, filepath) 我实际测试发现并没有实现计算和save的重叠，CheckFreq似乎也没有调用整个函数，请教一下这个函数究竟是否可行呢**

[Miracle0x0]

我当时也观察到这个问题，但没有深挖下去了……我个人认为CheckFreq给出的这部分代码不是论文中实验部分的代码，可能需要问一下他们团队的相关人员？