Search results

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

Status: Open.

#133 In Ascend/pytorch;

· JiaryCoder opened on May 13, 2026

CANN 9.0 Support

Status: Open.

#130 In Ascend/pytorch;

· chengchen666 opened on May 8, 2026

torch.zero_ in compile

Status: Open.

#128 In Ascend/pytorch;

· 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)

Status: Open.

#125 In Ascend/pytorch;

· 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

Status: Open.

#123 In Ascend/pytorch;

· ycaibb opened on Apr 22, 2026

【feature】当配置为cuda和nccl时，ptx 不支持 创建通信域

Status: Open.

#121 In Ascend/pytorch;

· littleyellowbicycle opened on Apr 20, 2026

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

Status: Open.

#110 In Ascend/pytorch;

· 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？

Status: Open.

#109 In Ascend/pytorch;

· 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.

Status: Open.

#108 In Ascend/pytorch;

· 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%+

Status: Open.

#107 In Ascend/pytorch;

· cuizhiqiang0 opened on Feb 3, 2026

编译安装的时候报错：FileNotFoundError-v2r6/op_plugin_functions.yaml

Status: Open.

#106 In Ascend/pytorch;

· NiYueLiuFeng opened on Jan 29, 2026

需要一个torch-npu 2.9.1版本

Status: Open.

#105 In Ascend/pytorch;

· zy1257 opened on Jan 29, 2026