Issues

• os.environ.get取值异常

  Ascend/pytorch#170 · Filter by author bean002 opened on Sep 24, 2026
  2

• torch_npu.profiler 在 warmup → RECORD 跃迁时丢失窗口头部的设备侧数据，实际采集次数少于 active

  Ascend/pytorch#167 · Filter by author PHLoveLife opened on Sep 11, 2026
  1

• ProcessGroupHCCL creation fails with fftsplus SDMA device error (507049) when another process uses CaMem sleep mode on the same NPU

  Ascend/pytorch#166 · Filter by author daniel-ding530 opened on Aug 31, 2026
  1

• [Bug]: NPU 环境下 Dynamo 处理零参数 id() 时抛出 IndexError 而非 TypeError，与 PyTorch 原生行为不一致，影响设备测试解耦

  Ascend/pytorch#163 · Filter by author taochong123456 opened on Aug 11, 2026
  1

• [Requirement] 支持torch.uint64在npu上的相关计算

  Ascend/pytorch#158 · Filter by author daifucong233 opened on Aug 3, 2026

• [feat] is_cuda设计意义不明

  Ascend/pytorch#157 · Filter by author daifucong233 opened on Aug 3, 2026

• [Bug]: torch.compile在编译时会出现硬件访问超时问题

  Ascend/pytorch#154 · Filter by author ShaohonChen opened on Jul 28, 2026
  1

• [Bug]: 310P 布尔掩码 index_put_ 导致设备上下文崩溃（aicpu 错误 507018）

  Ascend/pytorch#153 · Filter by author ash934141435 opened on Jul 27, 2026

• [Requirement|需求建议]: 昇腾有独立分支支持pytorch-lightning训练框架

  Ascend/pytorch#151 · Filter by author zhoufei-tech opened on Jul 13, 2026

• [bug] interpolate doesn't work with small scale_factor values

  Ascend/pytorch#150 · Filter by author vladgor opened on Jul 3, 2026

• torch_npu 2.1.0 + CANN 8.0.0 在昇腾910B上微调Qwen3-ASR 模型

  Ascend/pytorch#148 · Filter by author lrs02 opened on Jun 25, 2026
  1

• 【bug】torch_npu不支持flex_attention，使用flex_attention进行训练会报错

  Ascend/pytorch#147 · Filter by author sunny-infra opened on Jun 25, 2026
  2

• torch_npu 2.10 + CANN 9.0 on soc_version 104: torch.nn.functional.linear fails with MatMul/TransData

  Ascend/pytorch#144 · Filter by author wakaka6 opened on Jun 8, 2026
  1

• 算子launch线程数量有误，全部由acl_thread进行算子下发

  Ascend/pytorch#140 · Filter by author AcK1ng opened on May 24, 2026
  1

• RFC: 为 torch_npu MemPool 增加 sub_block_alignment 参数

  Ascend/pytorch#139 · Filter by author whybeyoung opened on May 22, 2026

• torch_npu profiler 输出的 trace JSON 与 Perfetto 不兼容：ts 为字符串类型 + 绝对时间戳导致精度丢失

  Ascend/pytorch#133 · Filter by author JiaryCoder opened on May 13, 2026
  2

• CANN 9.0 Support

  Ascend/pytorch#130 · Filter by author chengchen666 opened on May 8, 2026

• torch.zero_ in compile

  Ascend/pytorch#128 · Filter by author arabel1a opened on May 4, 2026

• [Bug] ATC allow_mix_precision / allow_fp32_to_fp16 在 SAM3 vision encoder FPN neck 上造成灾难性精度崩溃 (与 torch_npu fp16 行为不一致)

  Ascend/pytorch#125 · Filter by author ZhaoyuZhong0913 opened on Apr 24, 2026

• torch_npu/_inductor 中 MLIRProcessor 使用 shell=True 执行由环境变量拼接的命令，存在条件性命令注入风险

  Ascend/pytorch#123 · Filter by author ycaibb opened on Apr 22, 2026
  2

• 【feature】当配置为cuda和nccl时，ptx 不支持 创建通信域

  Ascend/pytorch#121 · Filter by author littleyellowbicycle opened on Apr 20, 2026

• [Usage] AclnnGroupedMatmulWeightNz does not support x dtype int8 and weight dtype float16

  Ascend/pytorch#110 · Filter by author ZhongsJie opened on Mar 23, 2026

• Question: 在torch.npu的使用环境下，通过pytorch官方文档实现的自定义的Process Group Backend会有效果吗？

  Ascend/pytorch#109 · Filter by author 525309178 opened on Mar 19, 2026

• 执行deepseekv3.2-exp模型量化报错：torch_npu.memory: [ERROR] [269128] Get a block from the existing pool failed. Try to free cached blocks and reallocate. This error log can be ignored.

  Ascend/pytorch#108 · Filter by author mominhua opened on Mar 17, 2026

• Acend 910b和 nvidia H20 相同超参数，同一条训练数据， 对qwen-3vl-8b sft训练forward阶段loss差距在1%以内，backward后最后一层(35层)main_grad的误差小于1%， 传播到第一层 main_grad误差超50%-100%+

  Ascend/pytorch#107 · Filter by author cuizhiqiang0 opened on Feb 3, 2026
  1