Tracking issue for Apple M-series GPU support

This issue documents the current state of support for Apple M-series (M1/M2) support in Thinc and spacy-transformers. Subscribe to this issue if you'd like to receive status updates on support for M-series GPUs.

## Thinc layers

Apple M-series (M1/M2) GPUs are currently not used by Thinc models. The matrix multiplication (AMX) units of the M-series CPUs will be used when `thinc-apple-ops` is installed.

## PyTorch layers

M-series GPUs are supported by the PyTorch wrapper in Thinc. Calling `require_gpu` or `prefer_gpu` changes the default PyTorch device to `mps`. As a result, Metal Performance Shaders will be used by PyTorch to run compute kernels on the GPU.

Keep in mind though that the Metal Performance Shader support is still fairly limited, especially so in PyTorch 1.12.x, as a consequence your models may not work correctly.

## spacy-transformers

`spacy-transformers` 1.1.8 adds experimental support for M1 GPUs through Metal Performance Shaders. However, this support comes with some limitations:

- Only inference is supported.
- PyTorch 1.13.x is required. Even though PyTorch 1.12.x adds support for MPS, it is too incomplete to use with `spacy-transformers`.
- We have not tested MPS support on Intel Macs with Metal-capable GPUs.

We have tested the following models with MPS, PyTorch 1.13.0 and `spacy-transformers` 1.1.8. The models for which _Requires fallback_ is ticked only work with the `PYTORCH_ENABLE_MPS_FALLBACK=1` environment variable set.

| Model | Supported | Requires fallback |
| :--------------- | :-------: | :---------------: |
| ca_core_news_trf | ✅ | ✅ |
| da_core_news_trf | ✅ | |
| de_dep_news_trf | ✅ | |
| en_core_web_trf | ✅ | ✅ |
| es_dep_news_trf | ✅ | |
| fr_dep_news_trf | ✅ | ✅ |
| ja_core_news_trf | ✅ | |
| uk_core_news_trf | ✅ | ✅ |
| zh_core_web_trf | ✅ | |

You can find more information about using `spacy-transformers` with M-series GPUs in the [GPU Support Troubleshooting FAQ](https://github.com/explosion/spaCy/discussions/11436).

## Benchmarks

More background about the MPS support, including benchmarks, can be found in our blog post [Fast transformer inference with Metal Performance Shaders](https://explosion.ai/blog/metal-performance-shaders).