quantize_int4.py

"""Quantize Z-Image-Turbo OV FP16 -> INT4 using optimum (NNCF weight-only)."""
from pathlib import Path
from optimum.intel import OVDiffusionPipeline, OVQuantizer
from optimum.intel.openvino.configuration import (
    OVConfig,
    OVWeightQuantizationConfig,
    OVPipelineQuantizationConfig,
)

fp16_dir = "./z-image-turbo-ov-fp16"
int4_dir = "./z-image-turbo-ov-int4"

# INT4 config for transformer + text_encoder (main weights)
# group_size=128, sym=False is NNCF default sweet spot for quality/size
int4_config = OVWeightQuantizationConfig(
    bits=4,
    sym=False,
    group_size=128,
    group_size_fallback="adjust",
    ratio=1.0,
)

# Keep VAE in FP16 (don't quantize) by using default FP32? Actually set bits=8 for others,
# but we will explicitly only quantize transformer+text_encoder and copy rest as-is.
# Using OVPipelineQuantizationConfig:
pipeline_config = OVPipelineQuantizationConfig(
    quantization_configs={
        "transformer": int4_config,
        "text_encoder": int4_config,
    },
    # default: 8-bit weight-only for any other submodel that gets quantized,
    # but we skip vae by not including it? OVQuantizer quantizes only listed?
    # To be safe, set default 8-bit; vae will be copied if not quantized?
    default_config=OVWeightQuantizationConfig(bits=8, sym=True),
)

print(f"Loading FP16 OV pipeline from {fp16_dir} ...")
pipe = OVDiffusionPipeline.from_pretrained(fp16_dir)
print(f"Components: {list(pipe.components.keys())}")
print(f"OV submodels: {list(pipe.ov_submodels) if hasattr(pipe, 'ov_submodels') else 'n/a'}")

quantizer = OVQuantizer.from_pretrained(pipe)
print("Quantizing transformer+text_encoder to INT4 (NNCF, optimum)...")
ov_config = OVConfig(quantization_config=pipeline_config)
quantizer.quantize(
    save_directory=int4_dir,
    ov_config=ov_config,
)
print(f"Saved INT4 model to {int4_dir}")