yujiepan / meta-llama-3-tiny-random

This model is randomly initialized, using the config from meta-llama/Meta-Llama-3-8B-Instruct but with smaller size. Note the model is in bfloat16 .

"yujiepan/llama-3-tiny-random" and "yujiepan/meta-llama-3-tiny-random" shares exactly the same files except the repo name.

Codes:

import transformers
import torch
import os
from huggingface_hub import create_repo, upload_folder
import accelerate

source_model_id = 'meta-llama/Meta-Llama-3-8B-Instruct'
save_path = '/tmp/yujiepan/meta-llama-3-tiny-random'
repo_id = 'yujiepan/meta-llama-3-tiny-random'

os.system(f'rm -rf {save_path}')

config = transformers.AutoConfig.from_pretrained(
    source_model_id,
    trust_remote_code=True,
)
config._name_or_path = source_model_id
config.hidden_size = 4
config.intermediate_size = 14
config.num_attention_heads = 2
config.num_key_value_heads = 1
config.num_hidden_layers = 2
config.torch_dtype = "bfloat16"

model = transformers.AutoModelForCausalLM.from_config(
    config,
    trust_remote_code=True,
)

with accelerate.init_empty_weights():
    model.generation_config = transformers.AutoModelForCausalLM.from_pretrained(source_model_id).generation_config

model = model.to(torch.bfloat16)
model.save_pretrained(save_path)

tokenizer = transformers.AutoTokenizer.from_pretrained(
    source_model_id,
    trust_remote_code=True,
)
tokenizer.save_pretrained(save_path)

model.float().generate(torch.tensor([[1, 2, 3]]).long(), max_length=16)

os.system(f'ls -alh {save_path}')
# os.system(f'rm -rf {save_path}/model.safetensors')
create_repo(repo_id, exist_ok=True)
upload_folder(repo_id='yujiepan/meta-llama-3-tiny-random', folder_path=save_path)
upload_folder(repo_id='yujiepan/llama-3-tiny-random', folder_path=save_path)