app.py

import os, tempfile, traceback
import gradio as gr
import spaces
import requests

# ---------- Cache & HF Hub settings (before importing Docling) ----------
# Use persistent storage (/data) and disable xet (fixes PermissionDenied in some Spaces)
os.environ.setdefault("HF_HOME", "/data/.cache/huggingface")
os.environ.setdefault("HF_HUB_CACHE", "/data/.cache/huggingface/hub")
os.environ.setdefault("TRANSFORMERS_CACHE", "/data/.cache/huggingface/transformers")
os.environ.setdefault("HF_HUB_ENABLE_XET", "0")          # avoid xet write issues
os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")  # faster downloads
os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")

# Make sure the folders exist
for p in (os.environ["HF_HOME"], os.environ["HF_HUB_CACHE"], os.environ["TRANSFORMERS_CACHE"]):
    os.makedirs(p, exist_ok=True)

# ---------- Imports after env is set ----------
from docling.datamodel.base_models import InputFormat
from docling.document_converter import DocumentConverter, PdfFormatOption
from docling.pipeline.vlm_pipeline import VlmPipeline

# Detect CUDA (ZeroGPU will make this true on first decorated call)
try:
    import torch
    HAS_CUDA = torch.cuda.is_available()
    # keep threads modest on shared infra
    torch.set_num_threads(max(1, int(os.environ.get("OMP_NUM_THREADS", "2"))))
except Exception:
    HAS_CUDA = False

# Build converters once (lifetime of app)
# Standard = text-first (faster, good when PDF has text layer)
std_converter = DocumentConverter(
    format_options={InputFormat.PDF: PdfFormatOption()}
)
# VLM = Granite Docling (better for scans/tables/math)
vlm_converter = DocumentConverter(
    format_options={InputFormat.PDF: PdfFormatOption(pipeline_cls=VlmPipeline)}
)

# ---------- Helpers ----------
def _success(md: str, html: str):
    tmpdir = tempfile.gettempdir()
    md_path = os.path.join(tmpdir, "output.md")
    html_path = os.path.join(tmpdir, "output.html")
    with open(md_path, "w", encoding="utf-8") as f:
        f.write(md)
    with open(html_path, "w", encoding="utf-8") as f:
        f.write(html)
    return md, md_path, html_path

def _fail(msg: str):
    # show readable error in the preview panel
    return f"**Conversion failed**:\n```\n{msg}\n```", None, None

def _convert_local_path(path: str, use_vlm: bool):
    try:
        conv = vlm_converter if use_vlm else std_converter
        doc = conv.convert(source=path).document
        md = doc.export_to_markdown()
        html = doc.export_to_html()
        return _success(md, html)
    except Exception as e:
        return _fail(f"{e}\n\n{traceback.format_exc()}")

# ---------- GPU-decorated endpoints (ZeroGPU requirement) ----------
@spaces.GPU(duration=600)  # up to 10 minutes
def run_convert_file(file, mode):
    if file is None:
        return _fail("No file provided.")
    use_vlm = mode.startswith("VLM")
    return _convert_local_path(file.name, use_vlm)

@spaces.GPU(duration=600)
def run_convert_url(url, mode):
    if not url:
        return _fail("No URL provided.")
    # download to a temp file so Docling always reads a local path
    try:
        r = requests.get(url, stream=True, timeout=60)
        r.raise_for_status()
        fd, tmp_path = tempfile.mkstemp(suffix=".pdf")
        with os.fdopen(fd, "wb") as tmp:
            for chunk in r.iter_content(chunk_size=1 << 20):
                if chunk:
                    tmp.write(chunk)
    except Exception as e:
        return _fail(f"Failed to download URL: {e}")
    try:
        return _convert_local_path(tmp_path, mode.startswith("VLM"))
    finally:
        try:
            os.remove(tmp_path)
        except Exception:
            pass

# ---------- UI ----------
subtitle = "Device: **CUDA (ZeroGPU)**" if HAS_CUDA else "Device: **CPU** (GPU warms on first call)"

with gr.Blocks(title="Granite-Docling 258M — PDF → Markdown/HTML") as demo:
    gr.Markdown(
        f"""# Granite-Docling 258M — PDF → Markdown / HTML  
{subtitle}

**Modes**
- **Standard (faster)** → PDFs with a text layer  
- **VLM (Granite – better for complex/scanned)** → scans / heavy tables / formulas

_First call may be slow while models download and ZeroGPU warms. Cache lives in `/data`._
"""
    )

    mode = gr.Radio(
        ["Standard (faster)", "VLM (Granite – better for complex/scanned)"],
        value="Standard (faster)", label="Mode"
    )

    with gr.Tab("Upload PDF"):
        fi = gr.File(file_types=[".pdf"], label="PDF")
        out_md = gr.Markdown(label="Markdown Preview")
        dl_md = gr.File(label="Download Markdown")
        dl_html = gr.File(label="Download HTML")
        gr.Button("Convert").click(run_convert_file, [fi, mode], [out_md, dl_md, dl_html])

    with gr.Tab("Convert from URL"):
        url = gr.Textbox(label="Public PDF URL", placeholder="https://.../file.pdf")
        out2_md = gr.Markdown(label="Markdown Preview")
        dl2_md = gr.File(label="Download Markdown")
        dl2_html = gr.File(label="Download HTML")
        gr.Button("Convert").click(run_convert_url, [url, mode], [out2_md, dl2_md, dl2_html])

# Explicit bind & queue
demo.queue().launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))