Chang Liu

[Example][Refactor] Refactor RGCN example

13

## Description Referring to: https://github.com/dmlc/dgl/pull/4186, this PR is for refactoring RGCN example. Within this PR, only entity classification task with full graph training (`entity.py`) and mini-batch training (`entity_sample.py`) are refactored....

Running error for all mini-batch runs within examples/pytorch/gcmc

2

## 🐛 Bug Mini-batch sampling pipline update, one of the major features released in v0.8, has not been propagated to examples/pytorch/gcmc/train_sampling.py. ## To Reproduce Step(s) to reproduce the behavior: (single...

[Example][Refactor] Refactor RGAT example

21

## Description To resolve https://github.com/dmlc/dgl/issues/4411, this PR refactors rgat example according to the golden example https://github.com/dmlc/dgl/pull/4186. ## Checklist Please feel free to remove inapplicable items for your PR. - [x]...

The RGAT example needs refactor ([rgat.py](https://github.com/dmlc/dgl/blob/master/examples/pytorch/rgat/rgat.py)).

## 🚀 Feature When it comes to using RelGraphConv module, it seems we always converted to a homogeneous graph for all current examples; while for hetero-graph RGCN, we used another...

In unit test [test_save_embeddings], it initialized torch process group, but never destroys it. As a result, if another test follows that invokes the call to `torch.distributed.init_process_group()`, it will result in...

[Bug] [Example] Transformer example is dated and does not work with latest networkx

3

## 🐛 Bug The following example is outdated and does not work with later networkx pkg (`networkx==3.3`): https://github.com/dmlc/dgl/tree/master/examples/pytorch/transformer ## To Reproduce Steps to reproduce the behavior: 1. python3 translation_train.py --gpus...

[Performance] Remove unnecessary synchronization using thrust::cuda::par_nosync policy

1

We are always using asynchronous thrust launch on a cuda stream, which involves extra `cudaStreamSync` within thrust calls, e.g., https://github.com/rapidsai/wholegraph/blob/9f290c4ff814f0e527324c8f7095c04cfb461a79/cpp/src/wholememory_ops/functions/exchange_ids_nccl_func.cu#L63 https://github.com/rapidsai/wholegraph/blob/9f290c4ff814f0e527324c8f7095c04cfb461a79/cpp/src/wholegraph_ops/unweighted_sample_without_replacement_func.cuh#L340 It would be better to change to `thrust::cuda::par_nosync`, to...

Destroy the entire temp_memory handle before each allocation can be error-prone

1

Hi, Within `temp_memory_handle` class, can we avoid completely free the object when invoking each type of `*_malloc()` function? I thought, by design if needed, users/developers should be able to allocate...

This PR demonstrates how to integrate the [NVIDIA WholeGraph](https://github.com/rapidsai/wholegraph) into PyG’s graph and feature store base classes, providing a modular and PyG-like way to extend PyG's dataloader for better GPU...