Memory-Efficient LLM Training by Various-Grained Low-Rank Projection of Gradients

Re-assign community