MindSpore replicate_params backward prefetch ST teardown 阶段 segfault

问题描述

HyperParallel 门禁流水线中，MindSpore fully_shard replicate_params backward prefetch 回归用例在主体执行通过后，msrun worker 在进程退出/GC 阶段触发 native segmentation fault，导致流水线失败。

失败流水线：
https://build.mindspore.cn/blue/organizations/jenkins/Hyper-parallel_Atomgit_Gate/detail/Hyper-parallel_Atomgit_Gate/4372/pipeline

失败用例：
tests/mindspore/st/fully_shard/test_fully_shard_replicate_prefetch_regression.py::test_ms_fully_shard_replicate_prefetch_regression_suite

实际失败的子用例：
test_ms_fully_shard_replicate_params_backward_prefetch_regression

关键日志

_test_fully_shard_replicate_prefetch_regression.py::test_ms_fully_shard_replicate_params_backward_prefetch_regression
======================= 1 passed, 30 warnings in 38.07s ========================
Fatal Python error: Segmentation fault
Current thread ...:
  Garbage-collecting
  <no Python frame>

Worker process 24341 exit with exception. Error code: -11.
RuntimeError: Distributed job exited with exception.
AssertionError: List cases failed: ['test_ms_fully_shard_replicate_params_backward_prefetch_regression']

同一 suite 中 test_ms_hsdp_replicate_dtensor_state_visible_after_backward 已通过，并打印：

rank: 0, max_logits_sum_after_backward: 112.0

test_ms_fully_shard_replicate_params_backward_prefetch_regression 也已打印 loss：

rank: 0, regression step loss: 2.7763614654541016

说明失败不是 Python 断言不满足，而是用例结束后的 MindSpore 分布式运行时清理阶段 native crash。

临时规避

PR #851 临时跳过该 ST，避免当前门禁被 teardown 阶段 segfault 阻塞。待 msrun worker 退出阶段 core dump 修复后，需要恢复该用例。

期望结果

• test_ms_fully_shard_replicate_params_backward_prefetch_regression 执行通过后，msrun worker 正常退出。
• 外层 parallel_run 不应因 worker -11 退出而失败。
• 恢复 test_ms_fully_shard_replicate_prefetch_regression_suite 的门禁覆盖。

schema_version: 1
source: gitcode
gitcode_repo: mindspore/hyper-parallel
gitcode_issue: 219
source_url: https://gitcode.com/mindspore/hyper-parallel/issues/219

Start with tests/mindspore/st/fully_shard/test_fully_shard_replicate_prefetch_regression.py and reproduce test_ms_fully_shard_replicate_params_backward_prefetch_regression through the HyperParallel_Atomgit_Gate pipeline. Check the msrun worker teardown after the passing loss output and compare it with the HSDP case that exits normally. Done means the worker exits without code -11, parallel_run succeeds, and the full regression suite is restored.