oneflow icon indicating copy to clipboard operation
oneflow copied to clipboard

OneFlow is a deep learning framework designed to be user-friendly, scalable and efficient.

Results 591 oneflow issues
Sort by recently updated
recently updated
newest added

enhancement
system

## Summary 由于 autograd.grad 的机制问题,会借用 tensor.retain_grad 属性,把临时的 grad 存在 tensor.grad 上,导致的结果错误。 ## Code to reproduce bug ```python import oneflow as flow # import torch as flow a = flow.ones(2, 3).requires_grad_()...

bug
community

Libai T5 ``` # T5-large model config model.cfg.num_attention_heads = 12 model.cfg.hidden_size = 384 model.cfg.hidden_layers = 6 ``` 生成 op graph 用于 debug ``` (MODULE:model.t5_model.encoder.layers.4.post_attention_layernorm:LayerNorm((384,), eps=1e-05, elementwise_affine=True)): ( (INPUT:_model.t5_model.encoder.layers.4.post_attention_layernorm_input.0.0_2:tensor(..., placement=oneflow.placement(type="cuda", ranks=[0,...

enhancement
graph

## 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...

bug
community

```python import oneflow as flow flow.cuda.manual_seed_all(1) ``` 程序调用了flow.cuda.manual_seed_all之后会立刻占用每一张显卡的部分显存,我这里每张卡都被占用了500-600MiB,可能是异常现象。pytorch调用这个接口并没有观察到会占用显卡的显存。

bug
community

## Summary https://github.com/Oneflow-Inc/oneflow/pull/9363 这个 PR 中,我用一个 named_tuple 把返回 TensorTuple 的算子结果包了一层,使其能够跟 torch 对齐: ``` In [1]: import torch In [2]: import oneflow as flow In [3]: x = torch.randn(2, 3, 4)...

bug
community

lazy tensor indexing 支持。 主要处理了: - indexing functor 内对 eager/lazy mode 的管理,value tensor 可以做合适的预处理 - 解决 lazy boxing 中对 0size tensor 没有处理,导致的程序崩溃

enhancement
feature
automerge
graph

背景:https://github.com/Oneflow-Inc/OneCloud/issues/147#issuecomment-1195307571 概述: 在 functional_api.yaml 中,支持了一种新的写法: ``` "TensorTuple (Tensor input, Int32 dim, Bool keepdim=False) => Max", ``` 来表明返回值是一个 named tuple,第一个元素的 name 是 values,第二个元素的 name 是 indices 实现: 在生成 pybind 接口的时候,判断 tensortuple...

enhancement
system
api

```python import oneflow as flow m = flow.nn.BatchNorm2d(10).to("cuda") m.half() print(m.running_mean.dtype) x = flow.randn(1, 10, 20, 20).to("cuda").half() print(m(x)) ``` torch可以正常运行,oneflow挂在 op 的type推导: ```shell oneflow.float16 Traceback (most recent call last): File "../../debug.py",...

bug
community