hpcaitech / colossalai-benchmark Goto Github PK

🐛 Describe the bug

I am installing the latest colossalai using the python setup.py install command, and the NUS HPC prompt some errors like:
subprocess.CalledProcessError: Command '['ninja', '-v']' returned non-zero exit status 1.
#error "You're running a too old version of GCC. We need GCC 5 or later."

If I use pip install colossalai, another error happens!

Well, pip install need Internet, and if I am using GPU in NUS, no internet can be accessed, so I have to use python setup.py install, but the GCC version is too old...

Lucky for me (maybe unlucky for other students using the server), I saved the old version's colossalai....

Thank you!!!

Environment

No response

Describe the feature

There's no README in Bert Benchmark now.

Hello, thanks for the wonderful project. Did you consider aligning the results with some commonly used ones?
https://github.com/mlcommons/training
https://github.com/Oneflow-Inc/DLPerf

🐛 Describe the bug

I try to train vision transformer following instructions in README.md. However, it throws an error in imagenet1k/train.py that there is no module named model_zoo. Corresponding code is from model_zoo.vit import vit_small_patch16_224. I tried to find this module in all repos in hpcaitech organization and required python wheels, but nothing was found.

My training script is DATA=../dataset/tfrecord torchrun --nproc_per_node=8 train.py --config=configs/vit_vanilla.py , which is executed in imagenet1k folder.

Environment

CUDA 11.3
Torch 1.12.1
Torchvision 0.13.1

🐛 Describe the bug

As a place to show the best practice for users, I believe it is necessary to help users to skip the annoying dataset preparation stage.

Environment

No response

In 'Usage' part, the first command needs launchers, eg. OpenMPI, but this is not mentioned. It's easy to mislead newbies to waste time and effort if they are running on their local machine.

Some parameters in the second command seem not necessary, eg. I can run the example by the following command.
DATA=/data/cifar-10 torchrun --nproc_per_node=2 --master_port=29501 train.py --config=configs/vit_vanilla.py

In addition, it's hard for newbies to know what content they should provide for those parameters. eg. how to know the RANK, IP_ADDRESS and PORT. It would be better if you can provide some explanation and example.

Describe the feature

Zero init context of DeepSpeed is not provided now. Let's add this feature so taht we can benchmark larger models, but need to take care of numel count.

The submodule in ColossalAI should be update its commit ID when there is any update in this repository. This may be done via github action and we should definitely automate this process to save some trouble.