60 lines
2.0 KiB
Python
60 lines
2.0 KiB
Python
#!/usr/bin/env python3
|
|
"""Proper NF4 quantization using BnB during loading."""
|
|
|
|
import torch
|
|
from transformers import AutoModelForCausalLM, BitsAndBytesConfig, AutoConfig
|
|
|
|
|
|
def quantize_model(model_path, output_path):
|
|
print(f"Quantizing model from: {model_path}")
|
|
print(f"Output: {output_path}\n")
|
|
|
|
bnb_config = BitsAndBytesConfig(
|
|
load_in_4bit=True,
|
|
bnb_4bit_quant_type="nf4",
|
|
bnb_4bit_compute_dtype=torch.float16,
|
|
bnb_4bit_use_double_quant=True,
|
|
)
|
|
|
|
print("Loading model with BnB 4-bit to CPU (~70GB bf16 -> ~17GB 4-bit)...")
|
|
print(" This will use CPU RAM for loading\n")
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_path,
|
|
quantization_config=bnb_config,
|
|
device_map="cpu",
|
|
torch_dtype=torch.float16,
|
|
trust_remote_code=True,
|
|
low_cpu_mem_usage=True,
|
|
)
|
|
|
|
print("✓ Model loaded and quantized to CPU")
|
|
|
|
# Count parameters
|
|
total_params = sum(p.numel() for p in model.parameters())
|
|
bnb_params = sum(p.numel() for p in model.parameters() if hasattr(p, 'quant_state') and p.quant_state is not None)
|
|
print(f" Total: {total_params / 1e9:.2f}B parameters")
|
|
print(f" Quantized: {bnb_params / 1e9:.2f}B parameters ({bnb_params/total_params*100:.1f}%)\n")
|
|
|
|
print(f"Saving quantized model to: {output_path}")
|
|
model.save_pretrained(output_path)
|
|
|
|
# Save tokenizer
|
|
try:
|
|
tokenizer = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
|
|
# Tokenizer saving would go here if needed
|
|
except:
|
|
pass
|
|
|
|
print(f"\n✅ Quantized model saved to: {output_path}")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
import argparse
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--model-path", type=str, default="/data/models/Ornith-1.0-35B")
|
|
parser.add_argument("--output-path", type=str, default="/data/models/Ornith-1.0-35B-nf4")
|
|
args = parser.parse_args()
|
|
|
|
quantize_model(args.model_path, args.output_path)
|