HF RL Explorer

TF OpenAI GPT model fails under mixed precision (AMP) and XLA

TF OpenAI GPT model fails under mixed precision (AMP) and XLA: a task in MiMo-V2.6-RL-harbor-code: MiMo-V2.6-RL Code (Harbor) (Harbor dataset). I'd like to train/run TFOpenAIGPTModel (and TFOpenAIGPTForSequenceClassification ) with mixed precision and/or XLA, both of which work fine on most other…

The task

I'd like to train/run `TFOpenAIGPTModel` (and `TFOpenAIGPTForSequenceClassification`) with mixed precision and/or XLA, both of which work fine on most other TF models in this repo. With OpenAI GPT specifically, neither works.

Part of FineEnvs/MiMo-V2.6-RL-harbor-code.