FengWen

nn.Conv2d Segmentation fault (core dumped)

5

## 描述 当调用oneflow.nn.Conv2d函数出现程序异常终断 Segmentation fault (core dumped) ## 最小复现代码 ```python import oneflow.nn as nn def oneflow_cal(in_channels=512, out_channels=256, kernel_size=128, stride=1, groups=1, bias=False): padding = kernel_size // 2 conv = nn.Conv2d( in_channels,...

设置CUDA_VISIBLE_DEVICES 环境变量没生效

4

## Summary import oneflow 后再通过设置 CUDA_VISIBLE_DEVICES 环境变量 对于OneFlow后端 并没有生效。 ## Code to reproduce bug OneFlow 设置 `CUDA_VISIBLE_DEVICES` 代码： ``` python import oneflow as flow import os os.environ["CUDA_VISIBLE_DEVICES"] = "1" device...

## Summary ddp模式打开flow.cuda.synchronize()， oneflow会卡死 ## Code to reproduce bug ### flow-time_sync.py ```python """ Usage: $ python flow-time_sync.py $ python -m oneflow.distributed.launch --nproc_per_node 2 flow-time_sync.py """ import random import oneflow as...

flow.save 通过追加的形式保存模型

2

## Summary flow.save(ckpt,path_save) 是通过追加的形式保存模型，导致保存的权重文件大小爆炸。 ## Code to reproduce bug ```python """ Usage: $ python test_flow_save.py --epochs 1 --save ./epochs_1 $ python test_flow_save.py --epochs 2 --save ./epochs_2 du -smh ./epochs* """...

## Summary oneflow 和torch 对于 amp使用没有对齐 - 出现问题代码地点：https://github.com/Oneflow-Inc/one-yolov5/blob/fix_align_code/models/common.py#L650-L669 ## Code to reproduce bug ### torch版代码 ```python from torch.cuda import amp autocast = False with amp.autocast(autocast): print(autocast) print("is ok") ``` 对应输出...

dev_scaled_dot_product_attention_math

2

Tensor attn_mask=None, Float dropout_p=0.0, Bool is_causal=False, Tensor dropout_mask=None ## Todo ### 测试 - [x] attention(query, key, value) - [x] attn_mask - [ ] dropout_p , dropout_mask - [ ] is_causal...

[Feature Request]: oneflow.nn.functional 缺失 scaled_dot_product_attention 属性

1

### Background and motivation https://github.com/huggingface/diffusers/blob/705c592ea98ba4e288d837b9cba2767623c78603/src/diffusers/models/attention_processor.py#L806-L807 ```shell >>> import oneflow.nn.functional as F >>> hasattr(F,"scaled_dot_product_attention") False >>> import oneflow >>> oneflow.__version__ '0.9.1+cu117.git.e4118c70a' ``` oneflow__version__ = 0.9.1+cu117.git.e4118c70a diffusers >= 0.19.3 ### API Proposal...

feature/add-oneflow.distributions.normal.Normal

1

需求: https://github.com/Oneflow-Inc/oneflow/issues/10122

### Background and motivation  添加一个.tensor.resize()方法可以直接在原tensor上调整大小,避免创建新的tensor与拷贝值的操作。这可以提高tensor大小调整的效率,也使得tensor值的变更更加简洁高效。 ### API Proposal ``` >>> x = torch.tensor([[1, 2], [3, 4], [5, 6]]) >>> x.resize_(2, 2) tensor([[ 1, 2], [ 3, 4]]) ``` `.resize_()`方法的用法与效果:...

## Install OneDiff and its dependencies. ```shell python3 -m pip install --pre oneflow -f https://oneflow-pro.oss-cn-beijing.aliyuncs.com/branch/community/cu121 && git clone https://github.com/siliconflow/onediff.git && cd onediff && python3 -m pip install -e . &&...