oneflow

oneflow copied to clipboard

Fix inplace mul 0size check bug

2

#close https://github.com/Oneflow-Inc/oneflow/issues/9131

BBuf

one embedding eager

9

主要改动： 1、OneEmbedding要在eager下使用，需要对optimizer做如下操作 ```python opt = flow.one_embedding.Optimizer( opt, embeddings=[dlrm_module.embedding.one_embedding] ) ``` 在optimizer的step方法中，会执行参数更新，如果漏掉了以上操作，会报错 2、补充一些缺失的functor 3、修改update op/kernel，允许learning_rate 为optional，另传参数learning_rate_val，lazy下learning_rate为tensor，eager下learning_rate通过attr传递 4、在单卡时走unique_key_value->lookup->embedding_gather，重写了embedding_gather以支持动态内存分配 eager下不支持各个操作之间的overlap，在dlrm上训练可以达到目标精度

guo-ran

support ci tag: pr sym link

14

howin98

针对full cache，对更新过的条目标记flag为True SaveSnapshot的时候，只保存这些更新过的条目

MARD1NO

Implement exponential_ and multinomial

8

需求来源： https://github.com/Oneflow-Inc/OneTeam/issues/1184#issuecomment-1232440993 ## Todo lists - [x] 实现 exponential_ 算子 - [x] functor 逻辑 - [x] cpu kernel - [x] cuda kernel - [x] 测试 - [x] 实现 multinomial 算子...

Ldpe2G

module.to aligned with pytorch

10

module.to 和 pytorch 对齐

daquexian

来源：yolov5加载yolov5s.onnx做预测的时候。 pytorch脚本： ``` import torch as flow targets = flow.randn((0, 6), device="cuda") print(targets.shape) height, width = 640, 640 targets[:, 2:] *= flow.tensor((width, height, width, height), device="cuda") ``` 输出： ``` torch.Size([0,...

BBuf

## Summary eager global 下 detach 接口获得结果的 autograd 属性不符合预期，影响后续后向图构图。 ## Code to reproduce bug ```python cuda_placement = flow.placement("cuda", [0, 1]) cpu_placement = flow.placement("cpu", [0, 1]) B = flow.sbp.broadcast a =...

wyg1997

Can oneflow do split-convolution operation ?

5

## Summary I am wondering if oneflow support this kind of operations. For example, I have an input tensor of [1, 3, 200, 200] ( [batch_size, channel, width, height] )...

DarrenYing

add spmm_coo op & kernel (as a part of GCN)

1

## SpMM COO NOT ready for merging! Add the basic spmm_coo to perform SpMM operations needed by GCN. Only the CUDA kernel is offered. Code pieces are borrowed from prior...

yuang-chen