glm-ocr-bucket.py

# /// script
# requires-python = ">=3.11"
# dependencies = [
#     "pillow",
#     "pymupdf",
#     "vllm",
#     "torch",
# ]
#
# [[tool.uv.index]]
# url = "https://wheels.vllm.ai/nightly/cu129"
#
# [tool.uv]
# prerelease = "allow"
# override-dependencies = ["transformers>=5.1.0"]
# ///

"""
OCR images and PDFs from a directory using GLM-OCR, writing markdown files.

Designed to work with HF Buckets mounted as volumes via `hf jobs uv run -v ...`
(requires huggingface_hub with PR #3936 volume mounting support).

The script reads images/PDFs from INPUT_DIR, runs GLM-OCR via vLLM, and writes
one .md file per image (or per PDF page) to OUTPUT_DIR, preserving directory structure.

Input:                          Output:
  /input/page1.png        →      /output/page1.md
  /input/report.pdf       →      /output/report/page_001.md
  (3 pages)                       /output/report/page_002.md
                                  /output/report/page_003.md
  /input/sub/photo.jpg    →      /output/sub/photo.md

Examples:

  # Local test
  uv run glm-ocr-bucket.py ./test-images ./test-output

  # HF Jobs with bucket volumes (PR #3936)
  hf jobs uv run --flavor l4x1 \\
      -s HF_TOKEN \\
      -v bucket/user/ocr-input:/input:ro \\
      -v bucket/user/ocr-output:/output \\
      glm-ocr-bucket.py /input /output

Model: zai-org/GLM-OCR (0.9B, 94.62% OmniDocBench V1.5, MIT licensed)
"""

import argparse
import base64
import io
import logging
import sys
import time
from pathlib import Path

import torch
from PIL import Image
import os
# Disable vLLM's FlashInfer sampler: it JIT-compiles a CUDA kernel needing nvcc, which the
# default uv-script image lacks (engine init then crashes). Greedy OCR doesn't use it; this
# lets the plain default-image command work. On the vllm/vllm-openai image it's a harmless no-op.
os.environ.setdefault("VLLM_USE_FLASHINFER_SAMPLER", "0")
from vllm import LLM, SamplingParams

logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
logger = logging.getLogger(__name__)

MODEL = "zai-org/GLM-OCR"

TASK_PROMPTS = {
    "ocr": "Text Recognition:",
    "formula": "Formula Recognition:",
    "table": "Table Recognition:",
}

IMAGE_EXTENSIONS = {".png", ".jpg", ".jpeg", ".tiff", ".tif", ".bmp", ".webp"}


def check_cuda_availability():
    if not torch.cuda.is_available():
        logger.error("CUDA is not available. This script requires a GPU.")
        sys.exit(1)
    logger.info(f"CUDA available. GPU: {torch.cuda.get_device_name(0)}")


def make_ocr_message(image: Image.Image, task: str = "ocr") -> list[dict]:
    """Create chat message for GLM-OCR from a PIL Image."""
    image = image.convert("RGB")
    buf = io.BytesIO()
    image.save(buf, format="PNG")
    data_uri = f"data:image/png;base64,{base64.b64encode(buf.getvalue()).decode()}"

    return [
        {
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": data_uri}},
                {"type": "text", "text": TASK_PROMPTS.get(task, TASK_PROMPTS["ocr"])},
            ],
        }
    ]


def discover_files(input_dir: Path) -> list[Path]:
    """Walk input_dir recursively, returning sorted list of image and PDF files."""
    files = []
    for path in sorted(input_dir.rglob("*")):
        if not path.is_file():
            continue
        ext = path.suffix.lower()
        if ext in IMAGE_EXTENSIONS or ext == ".pdf":
            files.append(path)
    return files


def prepare_images(
    files: list[Path], input_dir: Path, output_dir: Path, pdf_dpi: int
) -> list[tuple[Image.Image, Path]]:
    """
    Convert discovered files into (PIL.Image, output_md_path) pairs.

    Images map 1:1. PDFs expand to one image per page in a subdirectory.
    """
    import fitz  # pymupdf

    items: list[tuple[Image.Image, Path]] = []

    for file_path in files:
        rel = file_path.relative_to(input_dir)
        ext = file_path.suffix.lower()

        if ext == ".pdf":
            # PDF → one .md per page in a subdirectory named after the PDF
            pdf_output_dir = output_dir / rel.with_suffix("")
            try:
                doc = fitz.open(file_path)
                num_pages = len(doc)
                logger.info(f"PDF: {rel} ({num_pages} pages)")
                for page_num in range(num_pages):
                    page = doc[page_num]
                    # Render at specified DPI
                    zoom = pdf_dpi / 72.0
                    mat = fitz.Matrix(zoom, zoom)
                    pix = page.get_pixmap(matrix=mat)
                    img = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
                    md_path = pdf_output_dir / f"page_{page_num + 1:03d}.md"
                    items.append((img, md_path))
                doc.close()
            except Exception as e:
                logger.error(f"Failed to open PDF {rel}: {e}")
        else:
            # Image → single .md
            try:
                img = Image.open(file_path).convert("RGB")
                md_path = output_dir / rel.with_suffix(".md")
                items.append((img, md_path))
            except Exception as e:
                logger.error(f"Failed to open image {rel}: {e}")

    return items


def main():
    parser = argparse.ArgumentParser(
        description="OCR images/PDFs from a directory using GLM-OCR, output markdown files.",
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog="""
Task modes:
  ocr      Text recognition to markdown (default)
  formula  LaTeX formula recognition
  table    Table extraction (HTML)

Examples:
  uv run glm-ocr-bucket.py ./images ./output
  uv run glm-ocr-bucket.py /input /output --task table --pdf-dpi 200

HF Jobs with bucket volumes (requires huggingface_hub PR #3936):
  hf jobs uv run --flavor l4x1 -s HF_TOKEN \\
      -v bucket/user/input-bucket:/input:ro \\
      -v bucket/user/output-bucket:/output \\
      glm-ocr-bucket.py /input /output
        """,
    )
    parser.add_argument("input_dir", help="Directory containing images and/or PDFs")
    parser.add_argument("output_dir", help="Directory to write markdown output files")
    parser.add_argument(
        "--task",
        choices=["ocr", "formula", "table"],
        default="ocr",
        help="OCR task mode (default: ocr)",
    )
    parser.add_argument(
        "--batch-size", type=int, default=16, help="Batch size for vLLM (default: 16)"
    )
    parser.add_argument(
        "--max-model-len",
        type=int,
        default=8192,
        help="Max model context length (default: 8192)",
    )
    parser.add_argument(
        "--max-tokens",
        type=int,
        default=8192,
        help="Max output tokens (default: 8192)",
    )
    parser.add_argument(
        "--gpu-memory-utilization",
        type=float,
        default=0.8,
        help="GPU memory utilization (default: 0.8)",
    )
    parser.add_argument(
        "--pdf-dpi",
        type=int,
        default=300,
        help="DPI for PDF page rendering (default: 300)",
    )
    parser.add_argument(
        "--temperature",
        type=float,
        default=0.01,
        help="Sampling temperature (default: 0.01)",
    )
    parser.add_argument(
        "--top-p", type=float, default=0.00001, help="Top-p sampling (default: 0.00001)"
    )
    parser.add_argument(
        "--repetition-penalty",
        type=float,
        default=1.1,
        help="Repetition penalty (default: 1.1)",
    )
    parser.add_argument(
        "--verbose",
        action="store_true",
        help="Print resolved package versions",
    )

    args = parser.parse_args()

    check_cuda_availability()

    input_dir = Path(args.input_dir)
    output_dir = Path(args.output_dir)

    if not input_dir.is_dir():
        logger.error(f"Input directory does not exist: {input_dir}")
        sys.exit(1)

    output_dir.mkdir(parents=True, exist_ok=True)

    # Discover and prepare
    start_time = time.time()

    logger.info(f"Scanning {input_dir} for images and PDFs...")
    files = discover_files(input_dir)
    if not files:
        logger.error(f"No image or PDF files found in {input_dir}")
        sys.exit(1)

    pdf_count = sum(1 for f in files if f.suffix.lower() == ".pdf")
    img_count = len(files) - pdf_count
    logger.info(f"Found {img_count} image(s) and {pdf_count} PDF(s)")

    logger.info("Preparing images (rendering PDFs)...")
    items = prepare_images(files, input_dir, output_dir, args.pdf_dpi)
    if not items:
        logger.error("No processable images after preparation")
        sys.exit(1)

    logger.info(f"Total images to OCR: {len(items)}")

    # Init vLLM
    logger.info(f"Initializing vLLM with {MODEL}...")
    llm = LLM(
        model=MODEL,
        trust_remote_code=True,
        max_model_len=args.max_model_len,
        gpu_memory_utilization=args.gpu_memory_utilization,
        limit_mm_per_prompt={"image": 1},
    )

    sampling_params = SamplingParams(
        temperature=args.temperature,
        top_p=args.top_p,
        max_tokens=args.max_tokens,
        repetition_penalty=args.repetition_penalty,
    )

    # Process in batches
    errors = 0
    processed = 0
    total = len(items)

    for batch_start in range(0, total, args.batch_size):
        batch_end = min(batch_start + args.batch_size, total)
        batch = items[batch_start:batch_end]
        batch_num = batch_start // args.batch_size + 1
        total_batches = (total + args.batch_size - 1) // args.batch_size

        logger.info(f"Batch {batch_num}/{total_batches} ({processed}/{total} done)")

        try:
            messages = [make_ocr_message(img, task=args.task) for img, _ in batch]
            outputs = llm.chat(messages, sampling_params)

            for (_, md_path), output in zip(batch, outputs):
                text = output.outputs[0].text.strip()
                md_path.parent.mkdir(parents=True, exist_ok=True)
                md_path.write_text(text, encoding="utf-8")
                processed += 1

        except Exception as e:
            logger.error(f"Batch {batch_num} failed: {e}")
            # Write error markers for failed batch
            for _, md_path in batch:
                md_path.parent.mkdir(parents=True, exist_ok=True)
                md_path.write_text(f"[OCR ERROR: {e}]", encoding="utf-8")
            errors += len(batch)
            processed += len(batch)

    elapsed = time.time() - start_time
    elapsed_str = f"{elapsed / 60:.1f} min" if elapsed > 60 else f"{elapsed:.1f}s"

    logger.info("=" * 50)
    logger.info(f"Done! Processed {total} images in {elapsed_str}")
    logger.info(f"  Output: {output_dir}")
    logger.info(f"  Errors: {errors}")
    if total > 0:
        logger.info(f"  Speed: {total / elapsed:.2f} images/sec")

    if args.verbose:
        import importlib.metadata

        logger.info("--- Package versions ---")
        for pkg in ["vllm", "transformers", "torch", "pillow", "pymupdf"]:
            try:
                logger.info(f"  {pkg}=={importlib.metadata.version(pkg)}")
            except importlib.metadata.PackageNotFoundError:
                logger.info(f"  {pkg}: not installed")


if __name__ == "__main__":
    if len(sys.argv) == 1:
        print("=" * 60)
        print("GLM-OCR Bucket Script")
        print("=" * 60)
        print("\nOCR images/PDFs from a directory → markdown files.")
        print("Designed for HF Buckets mounted as volumes (PR #3936).")
        print()
        print("Usage:")
        print("  uv run glm-ocr-bucket.py INPUT_DIR OUTPUT_DIR")
        print()
        print("Examples:")
        print("  uv run glm-ocr-bucket.py ./images ./output")
        print("  uv run glm-ocr-bucket.py /input /output --task table")
        print()
        print("HF Jobs with bucket volumes:")
        print("  hf jobs uv run --flavor l4x1 -s HF_TOKEN \\")
        print("      -v bucket/user/ocr-input:/input:ro \\")
        print("      -v bucket/user/ocr-output:/output \\")
        print("      glm-ocr-bucket.py /input /output")
        print()
        print("For full help: uv run glm-ocr-bucket.py --help")
        sys.exit(0)

    main()