Closed shamanez closed 1 month ago
Thank you for reporting the issue. For freezing weights, you can follow the same technique that one would follow for other torch models, need to set the requires_grad
attr to false for the params that needs to be frozen.
closing this issue, please re-open if the issue persists.
I am following the llama2 pre-training code.
https://github.com/aws-neuron/neuronx-distributed/blob/main/examples/training/llama2/tp_zero1_llama2_7b_hf_pretrain/tp_zero1_llama2_7b_hf_pretrain.py
I do not understand how to freeze some of the parameters.