amalxloop/OpenMythos — reverse-engineered prompt

Build me a Python research package called OpenMythos. I want it to be a theoretical, independent reconstruction of a Claude Mythos style language model, not connected to Anthropic and not pretending to include trained weights.

The package should let someone create a PyTorch model with a simple config, run a forward pass on token ids, generate new token ids, and print the parameter count. The model should have three main stages, a first transformer section, a reusable internal reasoning block that can loop several times, and a final transformer section. Please include switchable MLA and GQA attention, sparse expert feed forward layers, shared experts, and configurable loop counts.

Add easy preset configs for sizes from around 1B up to 1T parameters. Include a small example script, basic tests, clear docs, and a training script example for FineWeb Edu using multiple GPUs when available. Make the package installable with pip, and keep the code clean enough for researchers to read and modify.