scripts/ckpt2hf.py

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

print("Loading checkpoint...")

# Define the paths to your checkpoint files
checkpoint_paths = [
    './llama3-5b/model-00001-of-00003.pt',
    './llama3-5b/model-00002-of-00003.pt',
    './llama3-5b/model-00003-of-00003.pt'
]

# Initialize an empty state dictionary
merged_state_dict = {}

# Load each checkpoint and merge them
for checkpoint_path in checkpoint_paths:
    checkpoint = torch.load(checkpoint_path, map_location='cpu')
    merged_state_dict.update(checkpoint)

print("Loading original model...")

# Define the original model name or path
original_model_name = "../../slice_with_mergekit/merged/"

# Load the model configuration and create a new model instance
model = AutoModelForCausalLM.from_pretrained(original_model_name, state_dict=merged_state_dict)

print("Converting to fp16...")

# Convert model parameters to float16
model.half()

print("Saving model...")

# Save the model in the safetensors format
output_dir = './llama3-5b/hf/'
model.save_pretrained(output_dir, safe_serialization=True)

print("Saving tokenizer...")

# Save the tokenizer as well
tokenizer = AutoTokenizer.from_pretrained(original_model_name)
tokenizer.save_pretrained(output_dir)

print(f"Merged model saved to {output_dir}")