Views
No views yet
meta-llama/Llama-3.3-70B-Instruct that only has 20 layers.
The original model has 80 layers; the last 60 layers have been dropped.1import torch
2import transformers
3
4def main():
5
6 llama_ver = 3.3
7 llama_size = 70
8 n_layers_out = 20
9 pipeline = transformers.pipeline("text-generation",
10 model=f"meta-llama/Llama-{llama_ver}-{llama_size}B-Instruct",
11 model_kwargs={"torch_dtype": torch.bfloat16},
12 device_map="auto")
13
14 pipeline.model.config.num_hidden_layers = n_layers_out
15 pipeline.model.model.layers = pipeline.model.model.layers[0:n_layers_out]
16 print(pipeline.model)
17 pipeline.save_pretrained(f"llama-{llama_ver}-{n_layers_out}L",safe_serialization=True)
18
19if __name__ == '__main__': main()