Search results

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

pytorch-npu hosting website description update

Status: Open.

#104 In Ascend/pytorch;

· iambowen opened on Jan 22, 2026

DataLoader加载数据集时遇到：AttributeError: '_MultiProcessingDataLoaderIter' object has no attribute '_in_order'

Status: Open.

#103 In Ascend/pytorch;

· Persistence2021 opened on Jan 20, 2026

对HSTU模型启用编译，编译 layer_norm 和 addmm_silu 算子报错 error: ub overflow

Status: Open.

#102 In Ascend/pytorch;

· zkdliushuo opened on Jan 19, 2026

SFTTrainer Out of Memory: Full-parameter fine-tuning on 8*910B (64G) for long context model Qwen3-4B-Instruct-2507.

Status: Open.

#101 In Ascend/pytorch;

· chenweiyj opened on Jan 19, 2026

模型是否支持FP8格式模型推理？

Status: Open.

#100 In Ascend/pytorch;

· MirrorYu opened on Jan 16, 2026

torch_npu2.7.1 和CANN 8.3.RC1不兼容

Status: Open.

#99 In Ascend/pytorch;

· ljljkatakuli opened on Jan 6, 2026

补图过多导致压测卡死

Status: Open.

#98 In Ascend/pytorch;

· chenxu214 opened on Dec 30, 2025

对HSTU模型启用编译报错 No module named 'torch._inductor.triton_heuristics'

Status: Open.

#97 In Ascend/pytorch;

· zkdliushuo opened on Dec 25, 2025

910b多进程出现RuntimeError: devptr INTERNAL ASSERT FAILED

Status: Open.

#96 In Ascend/pytorch;

· Limerence7 opened on Dec 22, 2025