EasyContext-256K-danube2-1.8b/README.md

EasyContext

GitHub Repo

Memory optimization and training recipes to extrapolate language models' context length to 1 million tokens, with minimal hardware.

This is a context-extrapolated base model. It has not been instruct-finetuned.

This model is finetuned from h2oai/h2o-danube2-1.8b-base with EasyContext on context length 256K. Note that I keep max_position_embeddings in config.json to 4096 because HF llama will create 2D causal mask during initialization. If it is set to 256K GPU will just OOM. You can surely use this model with context length longer than 4096.