No safe way to set the CPU thread count from Python: _unsafe_reset_threadpool after load_method crashes or hangs XNNPACK

未关闭
  • 哪里不对？
#24,765 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

@shoumikhin 已经在做这个了。

开始于 2026年11月7日。

  • #24770 来自 @shoumikhin —— 未关闭

评估

难度
4/5
预计耗时
3-5 天
新手友好度
25/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
停滞
技术栈
linux, python, pytorch
领域
backend, machine-learning

调研方向

Start with the Python binding for _unsafe_reset_threadpool and the thread pool lifetime rules in threadpool.h; trace how loaded methods retain the pool pointer. The issue includes a reproducer using XNNPACK and Runtime.load_method, so use it to check the reset-before-load and reset-after-load cases. Done means Python has a supported way to choose the thread count or rejects a reset after loading before it can crash or hang; an open linked pull request indicates work is already underway.

由索引模型根据 Issue 内容生成。

描述

bug
🐛 Describe the bug

From Python, _unsafe_reset_threadpool(n) is the only way to set the thread count that XNNPACK uses. Called after a method is loaded, it makes the following execute calls segfault or hang forever (all threads waiting on a futex). Called before load_method, it works.

import torch
from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
from executorch.exir import to_edge_transform_and_lower
from executorch.extension.pybindings.portable_lib import (
    _threadpool_get_thread_count,
    _unsafe_reset_threadpool,
)
from executorch.runtime import Runtime

model = torch.nn.Sequential(torch.nn.Linear(256, 256), torch.nn.ReLU(), torch.nn.Linear(256, 256)).eval()
x = torch.randn(64, 256)
ep = torch.export.export(model, (x,))
pte = to_edge_transform_and_lower(ep, partitioner=[XnnpackPartitioner()]).to_executorch()
method = Runtime.get().load_program(pte.buffer).load_method("forward")
_unsafe_reset_threadpool(2 * _threadpool_get_thread_count())
for _ in range(20):
    method.execute([x])  # segfaults or hangs

On Linux x86_64 with 8 threads by default, a reset to 16 after load_method crashed or hung in 10 of 11 runs, and a reset to 4 in 3 of 5. The same resets before load_method passed 10 of 10. Linux aarch64 shows the same pattern, including the hang and an abort with The futex facility returned an unexpected error code. A reset to the current count does nothing, so it always looks fine.

Cause: the XNNPACK runtime gets the thread pool pointer when the method is loaded (xnn_create_runtime_v4(..., get_pthreadpool(), ...)), and the reset frees that pool and creates a new one.

threadpool.h warns that other code may still hold the old pointer and says to reset before loading. The Python function has no such note and does not stop the call. This is different from #14321 (a macOS crash from two copies of pthreadpool): here there is one copy, and the trigger is the order of the calls.

Ask:

  1. A supported way to set the thread count from Python, for example a runtime or backend option.
  2. Until then, raise an error instead of crashing when the reset comes after a method is loaded.
Versions
  • executorch 1.6.0 (PyPI) and the 2026-11-05 nightly
  • torch 2.15.0, and 2.16 nightly with the ExecuTorch nightly
  • Linux x86_64 and Linux aarch64, CPU
主要语言
Python
星标
5.2k
派生
1.2k
平均合并
1 天 23 小时
30 天内合并 PR
700

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

pytorch/executorch 的其他 Issue

查看 pytorch/executorch 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。