[BUG] conda-cpp-tests on gb300: routing PDP test hits illegal memory access, set_shmem_of_kernel fails after it

未关闭
#2,212 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

@0z5a 已经在做这个了。

开始于 2026年11月3日。

  • #2260 来自 @0z5a —— 未关闭

评估

难度
4/5
预计耗时
1-2 天
新手友好度
48/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃

调研方向

Start with the failing cost_boundary_initialization.pdp_set_route_data_resets_cost_boundaries test and the set_shmem_of_kernel tests in the routing test binary; the report says the latter failures may be fallout from a sticky CUDA error. Check the nightly run and reproduce on gb300 if available, then trace the illegal memory access in the PDP test. Done when the PDP test passes and the shared memory tests no longer see error 700.

由索引模型根据 Issue 内容生成。

描述

bug

Describe the bug
In nightly run 40167716943 (2026-10-25, b5ad4d07), conda-cpp-tests / 13.3.0, 3.14, arm64, ubuntu26.04, gb300, latest-driver, latest-deps fails (job 121092087617). 3 routing tests fail, in every retry:

  • cost_boundary_initialization.pdp_set_route_data_resets_cost_boundaries: C++ exception from cuda/__driver/driver_api.h:662, an illegal memory access was encountered
  • set_shmem_of_kernel.zero_request: cudaGetLastError() returns 700 (cudaErrorIllegalAddress), expected cudaSuccess (set_shmem_of_kernel.cu:32)
  • set_shmem_of_kernel.normal_request: set_shmem_of_kernel(kernel_normal, 4096) returns false (set_shmem_of_kernel.cu:38)

The two set_shmem_of_kernel failures look like fallout of the illegal access in the preceding PDP test: the sticky error 700 is still set when they run. Within the routing test binary, the other 57 tests pass.

  • All 12 other conda-cpp-tests jobs in the run passed.
  • The same commit passed on gb300 the next night (40277562656), so this is intermittent.
  • This is not the failure set in #2176 (pdlp_class, Population, ReducedMatrix). None of those tests failed in this run.

Expected behavior
cost_boundary_initialization.pdp_set_route_data_resets_cost_boundaries passes on gb300 as on the other GPUs.

Environment details

  • Environment location: CI, gb300 arm64, CUDA 13.3.0, Python 3.14, ubuntu26.04, latest driver
  • Method of cuOpt install: conda build from source (rattler-build)

Additional context
Related: #2176 (other gb300-only C++ test failures).

主要语言
Cuda
星标
1.1k
派生
249
平均合并
1 天 15 小时
30 天内合并 PR
124

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。

NVIDIA/cuopt 的其他 Issue

查看 NVIDIA/cuopt 的全部 Issue

相似的 Issue

更多 Backend & API Design Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。