The TorchShard project welcomes your expertise and enthusiasm! If you are interested in torchshard, you are welcome to help to make it better and more professional.
For major changes, please write full tests and make sure all the changes pass them.
Test cases for a new core feature should:
- have its non-parallel implementation and parallel implementation
- have a test with DDP to wrap up the target layer or module
- pass forward equality checks, including layer outputs, loss values, and any other kinds of intermediate outputs
- pass backward equality checks, including outputs and gradients
- pass all the checks with the threshold value
1e-6
.
The tests folder holds the test scripts.
# under the project root
python3 setup.py install && rm -fr ./build ./dist ./torchshard.egg-info
➝ Back to the main page