0
代码
代码
Issues
Pull Requests
讨论
Wiki
分析
0
ascend-robotascend-robot【bugfix】Add npu device type check for int8 tensor fractal_nz format
912c6e9a创建于 22 小时前历史提交
文件最后提交记录最后更新时间
core
【bugfix】Add npu device type check for int8 tensor fractal_nz format Co-authored-by: mnhdxnh<947098055@qq.com> # message auto-generated for no-merge-commit merge: !744 merge master into master 【bugfix】Add npu device type check for int8 tensor fractal_nz format Created-by: mnhdxnh Commit-by: mnhdxnh Merged-by: ascend-robot Description: fix: Add npu device type check for int8 tensor fractal_nz format See merge request: Ascend/msprobe!74422 小时前
infer
【安全】delete rest owner check in offline dump and compare Co-authored-by: ylw1234<lwying007@126.com> # message auto-generated for no-merge-commit merge: !741 merge master into master 【安全】delete rest owner check in offline dump and compare Created-by: ylw1234 Commit-by: ylw1234 Merged-by: ascend-robot Description: delete rest owner check in offline dump and compare See merge request: Ascend/msprobe!74122 小时前
lib
[feat]Device MD5构建工程 Co-authored-by: Tjh-UKN<2559659915@qq.com> # message auto-generated for no-merge-commit merge: !711 merge md5 into master [feat]Device MD5构建工程 Created-by: Tjh-UKN Commit-by: Tjh-UKN Merged-by: ascend-robot Description: 背景：MD5计算需要host同步，现在提供device上计算MD5的算子，加速MD5计算 自验证：https://www.yuque.com/taejohnson/tzgbbf/cbai414pbcpgr9mc See merge request: Ascend/msprobe!7111 天前
mindspore
【安全】delete chmod in python dir Co-authored-by: ylw1234<lwying007@126.com> # message auto-generated for no-merge-commit merge: !707 merge master into master 【安全】delete chmod in python dir Created-by: ylw1234 Commit-by: ylw1234 Merged-by: ascend-robot Description: delete chmod in python dir See merge request: Ascend/msprobe!7074 天前
msaccucmp
【安全】delete rest owner check in offline dump and compare Co-authored-by: ylw1234<lwying007@126.com> # message auto-generated for no-merge-commit merge: !741 merge master into master 【安全】delete rest owner check in offline dump and compare Created-by: ylw1234 Commit-by: ylw1234 Merged-by: ascend-robot Description: delete rest owner check in offline dump and compare See merge request: Ascend/msprobe!74122 小时前
overflow_check
【优化】overflow_check性能优化 Co-authored-by: wugengjun<wugengjun1@huawei.com> # message auto-generated for no-merge-commit merge: !456 merge master0317 into master 【优化】overflow_check性能优化 Created-by: wugengjun Commit-by: wugengjun Merged-by: ascend-robot Description: 针对 overflow_check 模块中的 _connect_comm_nodes() 方法，通过引入哈希索引和缓存机制，将时间复杂度从 O(R² × N²) 降低到 O(R × N) 在256卡场景下测试，由原本的6小时多，优化为1小时45分左右 See merge request: Ascend/msprobe!4562 个月前
pytorch
【bugfix】monitor删除format配置项 Co-authored-by: l30036321<lvkaimeng@gmail.com> # message auto-generated for no-merge-commit merge: !745 merge master into master 【bugfix】monitor删除format配置项 Created-by: lv-kaimeng Commit-by: l30036321 Merged-by: ascend-robot Description: 【bugfix】monitor删除format配置项 See merge request: Ascend/msprobe!7451 天前
response_anomaly
推理服务亚健康检查模块：添加功能材料、脚本转换代码和一些代码优化 Co-authored-by: g30066442<gaoyelu@h-partners.com> # message auto-generated for no-merge-commit merge: !685 merge master into master 推理服务亚健康检查模块：添加功能材料、脚本转换代码和一些代码优化 Created-by: gaoyelu1996 Commit-by: g30066442 Merged-by: ascend-robot Description: 添加功能材料、模型参数和词表映射脚本转换代码和一些代码优化 See merge request: Ascend/msprobe!6857 天前
visualization
【安全】delete chmod in python dir Co-authored-by: ylw1234<lwying007@126.com> # message auto-generated for no-merge-commit merge: !707 merge master into master 【安全】delete chmod in python dir Created-by: ylw1234 Commit-by: ylw1234 Merged-by: ascend-robot Description: delete chmod in python dir See merge request: Ascend/msprobe!7074 天前
__init__.py
【bugfix】解决环境中有tensorflow环境，导入工具模块报错的问题 Co-authored-by: h00613304<hekunkun@huawei.com> # message auto-generated for no-merge-commit merge: !651 merge master into master 【bugfix】解决环境中有tensorflow环境，导入工具模块报错的问题 Created-by: kun_8 Commit-by: h00613304 Merged-by: ascend-robot Description: 执行以下代码时发生段错误： ```python from msprobe.pytorch import PrecisionDebugger # ← 触发崩溃 ``` **错误输出：** ``` Segmentation fault (core dumped) ``` --- ## 崩溃堆栈分析 ### 调用链 ``` Python 导入链: test_core_dump_error.py └── msprobe.pytorch └── msprobe/pytorch/monitor/data_writers.py:24 └── torch.utils.tensorboard.SummaryWriter └── tensorboard.compat.tf (懒加载) └── tensorflow └── tensorflow.python.platform.self_check.preload_check() └── _pywrap_cpu_feature_guard.InfoAboutUnusedCPUFeatures() ← 崩溃点 ``` ### 崩溃点代码 ```python # tensorflow/python/platform/self_check.py:59-64 else: # Load a library that performs CPU feature guard checking. from tensorflow.python.platform import _pywrap_cpu_feature_guard _pywrap_cpu_feature_guard.InfoAboutUnusedCPUFeatures() # ← Segfault here ``` ## 根因分析 TensorFlow 的 _pywrap_cpu_feature_guard 是一个 C++ 扩展模块，用于检测 CPU 支持的指令集特性（如 AVX, SSE, NEON 等）。 在 **ARM aarch64** 架构上，TensorFlow 2.21.0 的预编译 wheel 可能存在以下问题： 1. **CPU 特性检测代码路径未正确适配 ARM 架构** 2. **与 NPU 驱动/运行时环境存在内存冲突** 3. **缺少必要的 CPU 指令集支持检测** ### 触发路径 msprobe 使用 PyTorch 的 SummaryWriter 进行日志记录，而 PyTorch 的 tensorboard 集成会尝试导入 TensorBoard。TensorBoard 的 compat 模块会懒加载 TensorFlow 以提供兼容性支持，最终导致 TensorFlow 的初始化代码被执行。 ## 解决方案 ### 创建自动补丁文件 ```python """msprobe 导入补丁 - 在导入 msprobe 前加载此模块""" import sys try: import tensorboard.compat.tensorflow_stub as tensorflow_stub sys.modules['tensorflow'] = tensorflow_stub sys.modules['tensorflow.python'] = tensorflow_stub except ImportError: pass ``` See merge request: Ascend/msprobe!65119 天前
config.json
add extra info config to enable stack&construct Co-authored-by: TAJh<taojiaheng1@huawei.com> 2 个月前
msprobe.py
【安全】add Input Path Security Statement (softlink) Co-authored-by: ylw1234<lwying007@126.com> # message auto-generated for no-merge-commit merge: !740 merge master into master 【安全】add Input Path Security Statement (softlink) Created-by: ylw1234 Commit-by: ylw1234 Merged-by: ascend-robot Description: add Input Path Security Statement (softlink) See merge request: Ascend/msprobe!7401 天前