yujiepan / gemma-2-tiny-random

This model is randomly initialized, using the config from google/gemma-2-27b-it but with smaller size.

Codes:

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

model_id = 'google/gemma-2-27b-it'
save_path = '/tmp/yujiepan/gemma-2-tiny-random'
repo_id = 'yujiepan/gemma-2-tiny-random'

config = transformers.AutoConfig.from_pretrained(model_id)
config.hidden_size = 8
config.head_dim = 2
config.intermediate_size = 16
config.num_attention_heads = 4
config.num_hidden_layers = 2
config.num_key_value_heads = 2

tokenizer = transformers.AutoTokenizer.from_pretrained(model_id)
tokenizer.save_pretrained(save_path)

model = transformers.AutoModelForCausalLM.from_config(config, torch_dtype=torch.bfloat16)
model.generation_config = transformers.GenerationConfig.from_pretrained(model_id)
with torch.no_grad():
    for p in model.parameters():
        torch.nn.init.uniform_(p, -0.1, 0.1)

pipe = pipeline('text-generation', model=model, tokenizer=tokenizer, do_sample=False, device='cuda')
print(pipe('Hello World!'))

model.save_pretrained(save_path)

os.system(f'ls -alh {save_path}')
create_repo(repo_id, exist_ok=True)
upload_folder(repo_id=repo_id, folder_path=save_path)