quantize_int4.py

"""Qwen-Image-2.1 -> OpenVINO FP16 export + NNCF weight-only INT4 quantization.

NOTE on versions (deviation from original pinned env, approved by user):
  Qwen/Qwen-Image-2.1 requires QwenImage21Pipeline (added in diffusers
  commit 6256aa766 "Add Qwen-Image 2.1 (#14804)", post-v0.40.0). Pinned
  diffusers==0.37.1 cannot load it (AttributeError). Therefore:
  - diffusers: 0.41.0.dev0 (git main, includes QwenImage21)
  - transformers: 5.10.4 (pulled by optimum-intel git; supports hub 1.x)
  - huggingface-hub: 1.33.0 (required by git diffusers >=1.31)
  - tokenizers: 0.22.2, optimum: 2.3.0
  - optimum-intel: 2.3.0.dev0+git (main, first version with QwenImage21 support)
  - openvino==2026.4.0 nncf==3.4.0 torch pillow psutil unchanged
  - transformers dependency table patched: hub cap <1.0 -> <2.0
  - optimum-intel modeling_visual_language.py patched for transformers>=5
    (VisionRotaryEmbedding alias + rot_pos_emb try/except, diffusion unused)

FP16 export CLI (same pattern as task example):
  optimum-cli export openvino -m Qwen/Qwen-Image-2.1 \
    --task text-to-image --library diffusers --weight-format fp16 \
    ./qwen-image-2.1-ov-fp16

INT4 quantization:
  OVQuantizer + OVPipelineQuantizationConfig nested in OVConfig:
  transformer + text_encoder (+ text_encoder_i2i, same Qwen3VL arch) with
  OVWeightQuantizationConfig(bits=4, sym=False, group_size=128,
    group_size_fallback="adjust", ratio=1.0),
  rest default INT8. Passed as ov_config=OVConfig(quantization_config=...).
"""
import os
import time
import json

FP16_DIR = "./qwen-image-2.1-ov-fp16"
INT4_DIR = "./qwen-image-2.1-ov-int4"
MODEL_ID = "Qwen/Qwen-Image-2.1"


def export_fp16():
    if os.path.exists(FP16_DIR) and os.listdir(FP16_DIR):
        print(f"FP16 model already exists at {FP16_DIR}, skipping export...")
        print("CLI: optimum-cli export openvino -m "
              f"{MODEL_ID} --task text-to-image --library diffusers "
              f"--weight-format fp16 {FP16_DIR}")
        return FP16_DIR, 0.0
    # Export is done via optimum-cli (see docstring); this fallback uses API.
    from optimum.intel import OVDiffusionPipeline
    print(f"Exporting {MODEL_ID} to FP16 OpenVINO...")
    start = time.time()
    pipeline = OVDiffusionPipeline.from_pretrained(
        MODEL_ID,
        export=True,
        compile=False,
        weight_format="fp16",
        token=os.environ.get("HF_TOKEN"),
    )
    pipeline.save_pretrained(FP16_DIR)
    dt = time.time() - start
    print(f"FP16 export completed in {dt:.2f}s")
    return FP16_DIR, dt


def quantize_int4(fp16_dir):
    from optimum.intel import OVDiffusionPipeline
    from optimum.intel.openvino import (
        OVQuantizer,
        OVConfig,
        OVPipelineQuantizationConfig,
        OVWeightQuantizationConfig,
    )

    print(f"Quantizing {fp16_dir} to INT4...")
    start = time.time()
    pipeline = OVDiffusionPipeline.from_pretrained(fp16_dir, compile=False)
    print("OV submodels:", pipeline._ov_model_names)
    quantizer = OVQuantizer.from_pretrained(pipeline)

    int4_cfg = OVWeightQuantizationConfig(
        bits=4,
        sym=False,
        group_size=128,
        group_size_fallback="adjust",
        ratio=1.0,
    )
    int8_default = OVWeightQuantizationConfig()  # bits=8 default INT8

    ov_config = OVConfig(
        quantization_config=OVPipelineQuantizationConfig(
            quantization_configs={
                "transformer": int4_cfg,
                "text_encoder": int4_cfg,
                "text_encoder_i2i": int4_cfg,
            },
            default_config=int8_default,
        )
    )
    quantizer.quantize(save_directory=INT4_DIR, ov_config=ov_config)
    dt = time.time() - start
    print(f"INT4 quantization completed in {dt:.2f}s")
    return INT4_DIR, dt


def main():
    print("=" * 60)
    print("Qwen-Image-2.1 OpenVINO INT4 Quantization")
    print("=" * 60)
    os.makedirs(OUTPUT_DIR, exist_ok=True)
    fp16_dir, export_time = export_fp16()
    int4_dir, quantize_time = quantize_int4(fp16_dir)
    data = {
        "model_id": MODEL_ID,
        "fp16_export_dir": fp16_dir,
        "int4_dir": int4_dir,
        "export_time_seconds": export_time,
        "quantize_time_seconds": quantize_time,
        "total_time_seconds": export_time + quantize_time,
        "quant_config": {
            "transformer": "OVWeightQuantizationConfig(bits=4,sym=False,group_size=128,group_size_fallback=adjust,ratio=1.0)",
            "text_encoder": "same INT4",
            "text_encoder_i2i": "same INT4 (Qwen-Image-2.1 editing text encoder, same arch)",
            "others": "default INT8 (OVWeightQuantizationConfig bits=8)",
            "ov_config_class": "OVPipelineQuantizationConfig nested in OVConfig",
        },
    }
    with open(os.path.join(OUTPUT_DIR, "benchmark_quantization.json"), "w") as f:
        json.dump(data, f, indent=2)
    print(f"\nDone. FP16: {fp16_dir} INT4: {int4_dir}")


if __name__ == "__main__":
    main()