[Accelerator refactor] Move Accelerator into Strategy

[four4fish]

Proposed refactor

Part 3 of Accelerator and Plugin refactor #10416

Motivation

Moving towards stable version

After step 2 Move Precision Plugin into TTP Precision Plugins should be part of Training Type Plugins #7324
Accelerator is not the routing layer for strategy and precision anymore, optimizer related logic, steps, hooks all moved in to strategy.

Accelerator only have device information - strategy should own accelerator. Reduce the code complexity and improve code maintainability

Pitch

move accelerator into ttp/strategy as device_plugin (similar to checkpoint_io) and updating logic in accelerator-connector, training, loops accordingly

[RFC] Should we have a new name for Accelerator?

Additional context

---

If you enjoy Lightning, check out our other projects! ⚡

• Metrics: Machine learning metrics for distributed, scalable PyTorch applications.

• Lite: enables pure PyTorch users to scale their existing code on any kind of device while retaining full control over their own loops and optimization logic.

• Flash: The fastest way to get a Lightning baseline! A collection of tasks for fast prototyping, baselining, fine-tuning, and solving problems with deep learning.

• Bolts: Pretrained SOTA Deep Learning models, callbacks, and more for research and production with PyTorch Lightning and PyTorch.

• Lightning Transformers: Flexible interface for high-performance research using SOTA Transformers leveraging Pytorch Lightning, Transformers, and Hydra.

cc @justusschock @awaelchli @akihironitta @kaushikb11 @ananthsub

[ananthsub]

move accelerator into ttp/strategy as device_plugin (similar to checkpoint_io)

I would prefer keeping the Accelerator name. Other than that this looks good to me.

[awaelchli]

Definitely 100% the Accelerator name should stay, and also its core responsibilities. It should remain the main component to access hardware. For example, all the *_to_device methods and similar currently on the strategy should eventually call in the Accelerator to perform the action IMO. Does that make sense?

[ananthsub]

For example, all the *_to_device methods and similar currently on the strategy should eventually call in the Accelerator to perform the action IMO. Does that make sense?

Even today, the accelerator doesn't handle moving the batch to device since the LightningModule's hooks can customize this. Moving the model to device also sits closer to the strategy IMO than the accelerator.

I do think the accelerator implementations could be useful for storing specific properties like:

• Whether to enable CUDA graphs (on the GPU/CUDA accelerator)
• Whether to enable intra-device inter-batch parallelism (for example, using CUDA streams & events: https://github.com/PyTorchLightning/pytorch-lightning/blob/619ef7a665244eab2c47c892f7d821360597da9e/pytorch_lightning/utilities/fetching.py#L334)