File size: 1,119 Bytes
de9df89 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 | """
Quick loader for quantized HunyuanImage-3.0 model.
Generated automatically by hunyuan_quantize_nf4.py
"""
import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
def load_quantized_hunyuan(model_path="H:\Testing\HunyuanImage-3-NF4-v2"):
"""Load the NF4 quantized HunyuanImage-3.0 model."""
quant_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
bnb_4bit_compute_dtype=torch.bfloat16,
)
model = AutoModelForCausalLM.from_pretrained(
model_path,
quantization_config=quant_config,
device_map="cuda:0", # Load entirely on GPU 0
trust_remote_code=True,
torch_dtype=torch.bfloat16,
attn_implementation="sdpa",
)
# Load tokenizer
model.load_tokenizer(model_path)
return model
if __name__ == "__main__":
print("Loading quantized model...")
model = load_quantized_hunyuan()
print("Model loaded successfully!")
print(f"Device map: {model.hf_device_map}")
|