|
|
|
While the `convert2llama.py` tool is available, we still advise opting for InternLM2 when practical, chiefly due to its superior efficiency. InternLM2, which is adapted from LLaMA, streamlines the process by integrating the `Wq`, `Wk`, `Wv` weight matrices into a single matrix `Wqkv`. This integration leads to approximately a **5%** speed increase during training. Given the substantial costs associated with pre-training, this efficiency boost can result in significant savings.
|