pytorch-lightning VS fairscale

def on_validation_end(self, trainer: Trainer, pl_module: LightningModule) -> None: # Trainer calls `on_validation_end` for sanity check. Therefore, it is necessary to avoid # calling `trial.report` multiple times at epoch 0. For more details, see # https://github.com/PyTorchLightning/pytorch-lightning/issues/1391. if trainer.sanity_checking: return

Under the hood, the ModelWrapper object will create a ML model based on the config (so far, an XGBoost model and a PyTorch Lightning model). Each of those will have a wrapper that conducts training and evaluation (since from my understanding of Lightning, Trainers are required to be outside of the class). In lack of a better name, I call these wrappers Fitters. For uniformity, I thought about adding a common interface IFitter, which is inherited by all model wrappers as outlined below.

Join their Slack to ask the community questions and check out the GitHub here.

Pytorch lightning benchmarks against pytorch on every PR (benchmarks to make sure that it is mot slower.

There is already a ton of momentum around automating ML workflows. I would suggest you contribute to a preexisting project like, for instance, PyTorch Lightning or fast.ai.

Is the problem the sheer number of options, or the fact that they are all together in one place? Would it be better if they were organized into the different trainer entrypoints (fit, validate, ...)? If that is the case, there was an RFC proposing this which you might find interesting, feel free to drop by and comment on the issue: https://github.com/PyTorchLightning/pytorch-lightning/issues/10444

I wanted to make a quick performance comparison between the GPU (Tesla K80) and TPU (v2-8) available in Google Colab with PyTorch. To do so quickly, I used an MNIST example from pytorch-lightning that trains a simple CNN.

We've noticed GPU 0 on our 3 GPU system is sometimes idle (which would explain performance differences). However its unclear to us why that may be. Similar to this issue

If you want a good example for reference, take a look at Pytorch Lightning's readme (https://github.com/PyTorchLightning/pytorch-lightning) It answers the 3 questions of "what is this", "why should I care", and "how do i use it" almost instantly

I skimmed through the README and paper. What does this library have that that hasn't been included in xformers or fairscale?

Things are slowly moving into PyTorch upstream such as the ZeRO redundancy optimizer but from my experience the team behind DeepSpeed just move faster. There is also fairscale from the FAIR team which seems to be a staging ground for experimental optimizations before they move into PyTorch. If you use Lightning, it's easy enough to try out these various libraries (docs here)

DeepSpeed [1] is amazing tool to enable the different kind of parallelisms and optimizations on your model. I would definitely not recommend reimplementing everything yourself.

Probably FairScale [2] too, but never tried it myself.

[1]: https://github.com/microsoft/DeepSpeed

[2]: https://github.com/facebookresearch/fairscale

It was helpful for me to see how DeepSpeed/FairScale stack up compared to vanilla PyTorch Distributed Training specifically when trying to reach larger parameter sizes, visualizing the trade off with throughput. A lot of the learnings ended up in the Lightning Documentation under the advanced GPU docs!

I created a feature request on the FairScale project so that we can track the progress on the integration: Support ZeRO-Offload · Issue #337 · facebookresearch/fairscale (github.com)