The Annotated GPT-2
amaarora.github.io
amaarora.github.io
* checkout mosaic ml, it has some recent improvements like Alibi that are important https://docs.mosaicml.com/en/v0.3.1/method_cards/alibi.html
* Huggingface trainer
* Pytorch lightning (still has some great training improvements like LR find and saving top k best checkpoints ect)
* Deepspeed/megatron/bloom
For inference:
try a few alternatives to OpenAI like https://text-generator.io/
or hosting your own network/huggingface pipelines (can use a feature extractor pipeline and a text generation pipeline to support getting embeddings and generating from the same one model on the GPU)