Notably, the 345M model (1.5 GB on disk) is big enough that it's pushing the limits of conventional GPUs, and an alternative method of finetuning the model (https://github.com/nshepperd/gpt-2/commit/47df6da611716b4826...) has to be used to prevent the GPU from going OOM.
I'm working on tools to streamline GPT-2 text generation: I'm currently porting the code above to gpt-2-simple (https://github.com/minimaxir/gpt-2-simple) to allow easy finetuning/generation, and am also working on a way to quickly build an API/client for easily deploying GPT-2 to production and generating text at scale and cost effectively. Even the 117M model, managing CPU and RAM performance is tricky.
But given the incredible results of just the 117M model (e.g. Hacker News titles from a retrained 117M model: https://github.com/minimaxir/hacker-news-gpt-2), I'm eager to put the 345M model through its paces.
I'm working on tools to streamline GPT-2 text generation: I'm currently porting the code above to gpt-2-simple (https://github.com/minimaxir/gpt-2-simple) to allow easy finetuning/generation, and am also working on a way to quickly build an API/client for easily deploying GPT-2 to production and generating text at scale and cost effectively. Even the 117M model, managing CPU and RAM performance is tricky.
But given the incredible results of just the 117M model (e.g. Hacker News titles from a retrained 117M model: https://github.com/minimaxir/hacker-news-gpt-2), I'm eager to put the 345M model through its paces.