This model is for debugging. It is randomly initialized using the config from mistralai/Mistral-Nemo-Instruct-2407 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 = 'mistralai/Mistral-Nemo-Instruct-2407'
repo_id = 'yujiepan/mistral-nemo-2407-tiny-random'
save_path = f'/tmp/{repo_id}'
config = transformers.AutoConfig.from_pretrained(model_id)
config.hidden_size = 8
config.intermediate_size = 32
config.num_attention_heads = 4
config.num_hidden_layers = 2
config.num_key_value_heads = 2
config.head_dim = 2
print(config)
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)
transformers.set_seed(42)
with torch.no_grad():
for _, p in sorted(model.named_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!'))
messages = [
{"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
{"role": "user", "content": "Who are you?"},
]
chatbot = pipeline("text-generation", model=save_path, max_length=1000, max_new_tokens=16)
print(chatbot(messages))
model.save_pretrained(save_path)