This model is randomly initialized, using the config from [https://huggingface.co/google/gemma-7b-it] but with smaller size. Note the model is in float16.
Codes:
from transformers import pipeline
from huggingface_hub import create_repo, upload_folder
import torch
import transformers
import os
model_id = 'google/gemma-7b-it'
save_path = '/tmp/yujiepan/gemma-tiny-random'
repo_id = 'yujiepan/gemma-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
print(config)
tokenizer = transformers.AutoTokenizer.from_pretrained(model_id)
tokenizer.save_pretrained(save_path)
model = transformers.AutoModelForCausalLM.from_config(config, torch_dtype=torch.float16)
model = model.half()
pipe = pipeline('text-generation', model=model, tokenizer=tokenizer, do_sample=False, device='cuda')
print(pipe('Hello World!'))
model.save_pretrained(save_path)
# ovmodel = OVModelForCausalLM.from_pretrained(save_path, export=True)
# ovmodel = ovmodel.half()
# ovmodel.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)