Closed SparkJiao closed 6 months ago
Hi. The forward process of llama implemented in collie is consistent with that in https://github.com/facebookresearch/llama, but it is inconsistent with the implementation and weight storage format in the Transformers repository. In order to load and store the llama weights in the transformers format in collie, we need to reshape the q and k matrices in the llama model when loading and saving the model weights.
Got it.Thanks very much!
Hi, thanks for your contribution very much!
I have a question about the following code snippet:
https://github.com/OpenLMLab/collie/blob/main/collie/models/llama/model.py#L586-L612
Why the transposition operation is requires here and while
v_proj
does not require similar operations?Thanks for your reply very much!