inference/scripts/quantize_and_push.py

"""Quantize See-through pipeline models (UNet + text encoders) to NF4 or FP8.

Saves quantized components locally and optionally pushes to HuggingFace.
VAE and TransparentVAE remain in bf16.

Usage (from repo root):
    python inference/scripts/quantize_and_push.py --quant_mode nf4
    python inference/scripts/quantize_and_push.py --quant_mode fp8 --pipeline layerdiff
    python inference/scripts/quantize_and_push.py --quant_mode nf4 --push_to_hub \
        --output_repo_layerdiff layerdifforg/seethroughv0.0.2_layerdiff3d_nf4 \
        --output_repo_depth 24yearsold/seethroughv0.0.1_marigold_nf4
"""

import os.path as osp
import argparse
import sys
import os
sys.path.append(osp.dirname(osp.dirname(osp.abspath(__file__))))

import shutil
import torch
from huggingface_hub import snapshot_download, HfApi

# diffusers BitsAndBytesConfig — for UNet (inherits from diffusers ModelMixin)
from diffusers import BitsAndBytesConfig as DiffusersBnBConfig

# transformers BitsAndBytesConfig — for CLIP text encoders (transformers models)
from transformers import BitsAndBytesConfig as TransformersBnBConfig
from transformers import CLIPTextModel, CLIPTextModelWithProjection

from modules.layerdiffuse.layerdiff3d import UNetFrameConditionModel


def make_quant_configs(quant_mode: str):
    """Build diffusers and transformers BitsAndBytesConfig for the chosen mode."""
    if quant_mode == "nf4":
        diffusers_config = DiffusersBnBConfig(
            load_in_4bit=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.bfloat16,
        )
        transformers_config = TransformersBnBConfig(
            load_in_4bit=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.bfloat16,
        )
    elif quant_mode == "fp8":
        diffusers_config = DiffusersBnBConfig(load_in_8bit=True)
        transformers_config = TransformersBnBConfig(load_in_8bit=True)
    else:
        raise ValueError(f"Unknown quant_mode: {quant_mode!r}. Expected 'nf4' or 'fp8'.")
    return diffusers_config, transformers_config


def copy_subfolder(src_root: str, dst_root: str, subfolder: str):
    """Copy a subfolder from src_root to dst_root, overwriting if it exists."""
    src = osp.join(src_root, subfolder)
    dst = osp.join(dst_root, subfolder)
    if osp.isdir(src):
        if osp.exists(dst):
            shutil.rmtree(dst)
        shutil.copytree(src, dst)
        print(f"  Copied subfolder: {subfolder}")
    else:
        print(f"  WARNING: subfolder {subfolder!r} not found in {src_root}")


def copy_file(src_root: str, dst_root: str, filename: str):
    """Copy a single file from src_root to dst_root."""
    src = osp.join(src_root, filename)
    dst = osp.join(dst_root, filename)
    if osp.isfile(src):
        shutil.copy2(src, dst)
        print(f"  Copied file: {filename}")
    else:
        print(f"  WARNING: file {filename!r} not found in {src_root}")


def quantize_layerdiff(
    repo_id: str,
    output_dir: str,
    quant_mode: str,
    hf_token: str = None,
):
    """Quantize the LayerDiff3D pipeline (SDXL-based)."""
    print(f"\n{'='*60}")
    print(f"Quantizing LayerDiff3D pipeline ({quant_mode})")
    print(f"  Source: {repo_id}")
    print(f"  Output: {output_dir}")
    print(f"{'='*60}")

    diffusers_config, transformers_config = make_quant_configs(quant_mode)

    # --- Download source repo ---
    print("\n[1/6] Downloading source repo...")
    local_repo = snapshot_download(
        repo_id,
        token=hf_token,
    )
    print(f"  Downloaded to: {local_repo}")

    os.makedirs(output_dir, exist_ok=True)

    # --- Quantize UNet ---
    print("\n[2/6] Loading and quantizing UNet...")
    unet = UNetFrameConditionModel.from_pretrained(
        local_repo,
        subfolder="unet",
        quantization_config=diffusers_config,
        torch_dtype=torch.bfloat16,
    )
    unet_dir = osp.join(output_dir, "unet")
    os.makedirs(unet_dir, exist_ok=True)
    unet.save_pretrained(unet_dir)
    print(f"  Saved quantized UNet to: {unet_dir}")
    del unet
    torch.cuda.empty_cache()

    # --- Quantize text_encoder (CLIPTextModel) ---
    print("\n[3/6] Loading and quantizing text_encoder (CLIPTextModel)...")
    text_encoder = CLIPTextModel.from_pretrained(
        local_repo,
        subfolder="text_encoder",
        quantization_config=transformers_config,
        torch_dtype=torch.bfloat16,
    )
    te_dir = osp.join(output_dir, "text_encoder")
    os.makedirs(te_dir, exist_ok=True)
    text_encoder.save_pretrained(te_dir)
    print(f"  Saved quantized text_encoder to: {te_dir}")
    del text_encoder
    torch.cuda.empty_cache()

    # --- Quantize text_encoder_2 (CLIPTextModelWithProjection) ---
    print("\n[4/6] Loading and quantizing text_encoder_2 (CLIPTextModelWithProjection)...")
    text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(
        local_repo,
        subfolder="text_encoder_2",
        quantization_config=transformers_config,
        torch_dtype=torch.bfloat16,
    )
    te2_dir = osp.join(output_dir, "text_encoder_2")
    os.makedirs(te2_dir, exist_ok=True)
    text_encoder_2.save_pretrained(te2_dir)
    print(f"  Saved quantized text_encoder_2 to: {te2_dir}")
    del text_encoder_2
    torch.cuda.empty_cache()

    # --- Copy bf16 components as-is ---
    print("\n[5/6] Copying bf16 components (VAE, TransparentVAE, scheduler, tokenizers)...")
    bf16_subfolders = ["trans_vae", "vae", "scheduler", "tokenizer", "tokenizer_2"]
    for sf in bf16_subfolders:
        copy_subfolder(local_repo, output_dir, sf)

    # --- Copy root config files ---
    print("\n[6/6] Copying root config files...")
    for fname in os.listdir(local_repo):
        fpath = osp.join(local_repo, fname)
        if osp.isfile(fpath):
            copy_file(local_repo, output_dir, fname)

    print(f"\nLayerDiff3D quantization complete: {output_dir}")


def quantize_marigold(
    repo_id: str,
    output_dir: str,
    quant_mode: str,
    hf_token: str = None,
):
    """Quantize the Marigold3D pipeline (SD1.5-based)."""
    print(f"\n{'='*60}")
    print(f"Quantizing Marigold3D pipeline ({quant_mode})")
    print(f"  Source: {repo_id}")
    print(f"  Output: {output_dir}")
    print(f"{'='*60}")

    diffusers_config, transformers_config = make_quant_configs(quant_mode)

    # --- Download source repo ---
    print("\n[1/4] Downloading source repo...")
    local_repo = snapshot_download(
        repo_id,
        token=hf_token,
    )
    print(f"  Downloaded to: {local_repo}")

    os.makedirs(output_dir, exist_ok=True)

    # --- Quantize UNet ---
    print("\n[2/4] Loading and quantizing UNet...")
    unet = UNetFrameConditionModel.from_pretrained(
        local_repo,
        subfolder="unet",
        quantization_config=diffusers_config,
        torch_dtype=torch.bfloat16,
    )
    unet_dir = osp.join(output_dir, "unet")
    os.makedirs(unet_dir, exist_ok=True)
    unet.save_pretrained(unet_dir)
    print(f"  Saved quantized UNet to: {unet_dir}")
    del unet
    torch.cuda.empty_cache()

    # --- Quantize text_encoder (CLIPTextModel, single encoder for SD1.5) ---
    print("\n[3/4] Loading and quantizing text_encoder (CLIPTextModel)...")
    text_encoder = CLIPTextModel.from_pretrained(
        local_repo,
        subfolder="text_encoder",
        quantization_config=transformers_config,
        torch_dtype=torch.bfloat16,
    )
    te_dir = osp.join(output_dir, "text_encoder")
    os.makedirs(te_dir, exist_ok=True)
    text_encoder.save_pretrained(te_dir)
    print(f"  Saved quantized text_encoder to: {te_dir}")
    del text_encoder
    torch.cuda.empty_cache()

    # --- Copy bf16 components as-is ---
    print("\n[4/4] Copying bf16 components (VAE, scheduler, tokenizer)...")
    bf16_subfolders = ["vae", "scheduler", "tokenizer"]
    for sf in bf16_subfolders:
        copy_subfolder(local_repo, output_dir, sf)

    # Copy root config files
    print("  Copying root config files...")
    for fname in os.listdir(local_repo):
        fpath = osp.join(local_repo, fname)
        if osp.isfile(fpath):
            copy_file(local_repo, output_dir, fname)

    print(f"\nMarigold3D quantization complete: {output_dir}")


def push_to_hub(local_dir: str, repo_id: str, hf_token: str):
    """Upload quantized model directory to HuggingFace."""
    print(f"\nPushing {local_dir} -> {repo_id}")
    api = HfApi(token=hf_token)
    api.create_repo(repo_id, repo_type="model", exist_ok=True)
    api.upload_folder(
        folder_path=local_dir,
        repo_id=repo_id,
        repo_type="model",
    )
    print(f"  Uploaded to: https://huggingface.co/{repo_id}")


def main():
    parser = argparse.ArgumentParser(
        description="Quantize See-through models to NF4 or FP8 and optionally push to HuggingFace."
    )
    parser.add_argument(
        "--pipeline",
        type=str,
        default="both",
        choices=["layerdiff", "marigold", "both"],
        help="Which pipeline to quantize (default: both)",
    )
    parser.add_argument(
        "--repo_id_layerdiff",
        type=str,
        default="layerdifforg/seethroughv0.0.2_layerdiff3d",
        help="Source bf16 HF repo for LayerDiff3D",
    )
    parser.add_argument(
        "--repo_id_depth",
        type=str,
        default="24yearsold/seethroughv0.0.1_marigold",
        help="Source bf16 HF repo for Marigold3D",
    )
    parser.add_argument(
        "--output_repo_layerdiff",
        type=str,
        default=None,
        help="Target HF repo for quantized LayerDiff3D (required if --push_to_hub)",
    )
    parser.add_argument(
        "--output_repo_depth",
        type=str,
        default=None,
        help="Target HF repo for quantized Marigold3D (required if --push_to_hub)",
    )
    parser.add_argument(
        "--output_local",
        type=str,
        default="workspace/quantized_models/",
        help="Local save root directory (default: workspace/quantized_models/)",
    )
    parser.add_argument(
        "--quant_mode",
        type=str,
        default="nf4",
        choices=["nf4", "fp8"],
        help="Quantization mode (default: nf4)",
    )
    parser.add_argument(
        "--push_to_hub",
        action="store_true",
        help="Push quantized models to HuggingFace",
    )
    args = parser.parse_args()

    # Read HF token
    hf_token = None
    hf_credential_path = osp.join(osp.dirname(osp.dirname(osp.dirname(osp.abspath(__file__)))), "hf_credential")
    if osp.isfile(hf_credential_path):
        hf_token = open(hf_credential_path).read().strip()
        print(f"Loaded HF token from {hf_credential_path}")
    elif os.environ.get("HF_TOKEN"):
        hf_token = os.environ["HF_TOKEN"]
        print("Using HF_TOKEN from environment")
    else:
        print("WARNING: No HF token found. Private repos will fail to download.")

    # Validate push args
    if args.push_to_hub:
        if args.pipeline in ("layerdiff", "both") and not args.output_repo_layerdiff:
            parser.error("--output_repo_layerdiff is required when --push_to_hub is set for layerdiff pipeline")
        if args.pipeline in ("marigold", "both") and not args.output_repo_depth:
            parser.error("--output_repo_depth is required when --push_to_hub is set for marigold pipeline")
        if hf_token is None:
            parser.error("--push_to_hub requires an HF token (hf_credential file or HF_TOKEN env var)")

    # Build output paths
    layerdiff_dir = osp.join(args.output_local, "layerdiff")
    marigold_dir = osp.join(args.output_local, "marigold")

    print(f"\nQuantization mode: {args.quant_mode}")
    print(f"Pipeline(s): {args.pipeline}")

    # --- Quantize ---
    if args.pipeline in ("layerdiff", "both"):
        quantize_layerdiff(
            repo_id=args.repo_id_layerdiff,
            output_dir=layerdiff_dir,
            quant_mode=args.quant_mode,
            hf_token=hf_token,
        )

    if args.pipeline in ("marigold", "both"):
        quantize_marigold(
            repo_id=args.repo_id_depth,
            output_dir=marigold_dir,
            quant_mode=args.quant_mode,
            hf_token=hf_token,
        )

    # --- Push to HF ---
    if args.push_to_hub:
        if args.pipeline in ("layerdiff", "both"):
            push_to_hub(layerdiff_dir, args.output_repo_layerdiff, hf_token)
        if args.pipeline in ("marigold", "both"):
            push_to_hub(marigold_dir, args.output_repo_depth, hf_token)

    # --- Summary ---
    print(f"\n{'='*60}")
    print("Summary")
    print(f"{'='*60}")
    print(f"Quantization mode: {args.quant_mode}")
    if args.pipeline in ("layerdiff", "both"):
        print(f"LayerDiff3D:")
        print(f"  Source:  {args.repo_id_layerdiff}")
        print(f"  Local:   {osp.abspath(layerdiff_dir)}")
        print(f"  Quantized: unet, text_encoder, text_encoder_2")
        print(f"  Kept bf16: trans_vae, vae, scheduler, tokenizer, tokenizer_2")
        if args.push_to_hub:
            print(f"  Pushed to: {args.output_repo_layerdiff}")
    if args.pipeline in ("marigold", "both"):
        print(f"Marigold3D:")
        print(f"  Source:  {args.repo_id_depth}")
        print(f"  Local:   {osp.abspath(marigold_dir)}")
        print(f"  Quantized: unet, text_encoder")
        print(f"  Kept bf16: vae, scheduler, tokenizer")
        if args.push_to_hub:
            print(f"  Pushed to: {args.output_repo_depth}")
    print(f"{'='*60}")


if __name__ == "__main__":
    main()