fix(P0): make torch optional, add OCR deps

- Wrap 'import torch' in try/except so app starts without PyTorch
- Safe device detection with string fallback when torch missing
- Guard all device.type references with _HAS_TORCH check
- Add PaddleOCR, paddlepaddle, torch (CPU), transformers to requirements
- Update Dockerfile to install CPU-only torch before other deps
- TrOCR/Surya remain available but load lazily only when selected

Dockerfile

@@ -5,17 +5,24 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     tesseract-ocr \
     tesseract-ocr-ara \
     tesseract-ocr-deu \
+    tesseract-ocr-eng \
     libgl1 \
     libglib2.0-0 \
+    libgomp1 \
     && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
 
-# Install Python deps
+# Install CPU-only PyTorch FIRST (smaller footprint)
+RUN pip install --no-cache-dir --extra-index-url https://download.pytorch.org/whl/cpu \
+    torch
+
+# Install Python deps (OCR engines + framework)
 COPY requirements.txt .
-RUN pip install --no-cache-dir -r requirements.txt
+RUN pip install --no-cache-dir --extra-index-url https://download.pytorch.org/whl/cpu \
+    -r requirements.txt
 
-# Copy app (standalone — includes RTL fixer, field extractor, preprocessing)
+# Copy app (standalone — includes preprocessing, RTL fixer, field extractor)
 COPY app.py .
 
 # Data directory (SQLite DB + JSONL exports)

app.py

@@ -20,10 +20,28 @@ from datetime import datetime
 from pathlib import Path
 from typing import Optional, List, Tuple
 
-import torch
 import numpy as np
 from PIL import Image, ImageFilter, ImageEnhance
 
+# ============================================
+# Optional PyTorch (needed for TrOCR / Surya)
+# ============================================
+try:
+    import torch
+    _HAS_TORCH = True
+except ImportError:
+    torch = None  # type: ignore[assignment]
+    _HAS_TORCH = False
+    print("[INFO] PyTorch not installed — TrOCR and Surya will be unavailable")
+
+# Device detection (safe fallback when torch is missing)
+if _HAS_TORCH:
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    print(f"Device: {device}")
+else:
+    device = "cpu"
+    print("Device: cpu (no PyTorch)")
+
 # ============================================
 # Configuration
 # ============================================
@@ -37,12 +55,6 @@ CACHE_TTL = 1800  # 30 minutes
 MAX_IMAGE_DIMENSION = 2000  # Max width/height for OCR processing
 DPI_TARGET = 200  # Target DPI for scanned documents
 
-# ============================================
-# Device Detection
-# ============================================
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-print(f"Device: {device}")
-
 # ============================================
 # Lazy Model Loading
 # ============================================
@@ -80,7 +92,7 @@ def _load_paddleocr():
             use_angle_cls=True,
             lang='en',  # Supports Arabic through multi-language model
             show_log=False,
-            use_gpu=device.type == 'cuda',
+            use_gpu=_HAS_TORCH and hasattr(device, 'type') and device.type == 'cuda',
             max_text_length=1000,
         )
         print("PaddleOCR loaded")
@@ -91,13 +103,15 @@ def _load_easyocr():
     if _easyocr_reader is None:
         import easyocr
         print("Loading EasyOCR...")
-        _easyocr_reader = easyocr.Reader(['en', 'ar'], gpu=device.type == 'cuda')
+        _easyocr_reader = easyocr.Reader(['en', 'ar'], gpu=_HAS_TORCH and hasattr(device, 'type') and device.type == 'cuda')
         print("EasyOCR loaded")
     return _easyocr_reader
 
 def _load_trocr():
     global _trocr_processor, _trocr_model
     if _trocr_processor is None:
+        if not _HAS_TORCH:
+            raise ImportError("PyTorch is required for TrOCR")
         from transformers import TrOCRProcessor, VisionEncoderDecoderModel
         print("Loading TrOCR...")
         _trocr_processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-handwritten")
@@ -110,7 +124,8 @@ def _load_trocr():
 def _load_surya():
     global _surya_det_model, _surya_rec_model, _surya_languages, SURYA_AVAILABLE
     if not SURYA_AVAILABLE:
-        if device.type == 'cuda':
+        _device_type = device.type if _HAS_TORCH and hasattr(device, 'type') else 'cpu'
+        if _device_type == 'cuda':
             try:
                 from surya.ocr import run_ocr
                 from surya.model.detection import segformer
@@ -371,7 +386,8 @@ def ocr_ensemble(image: Image.Image, engines: list = None) -> dict:
         available.append("paddleocr")
     if EASY_AVAILABLE:
         available.append("easyocr")
-    if device.type == 'cuda' and SURYA_AVAILABLE:
+    _device_type = device.type if _HAS_TORCH and hasattr(device, 'type') else 'cpu'
+    if _device_type == 'cuda' and SURYA_AVAILABLE:
         available.append("surya")
 
     # Filter to requested engines
@@ -1039,7 +1055,7 @@ async def health():
     if EASY_AVAILABLE:
         models.append("easyocr")
     models.append("trocr")
-    if device.type == "cuda":
+    if _HAS_TORCH and hasattr(device, 'type') and device.type == "cuda":
         models.append("surya")
     return {
         "status": "healthy",
@@ -1057,7 +1073,7 @@ async def get_models():
             {"id": "paddleocr", "name": "PaddleOCR", "type": "full-page", "status": "available" if PADDLE_AVAILABLE else "not installed", "size_mb": 300, "best_for": "Complete documents, Arabic+English"},
             {"id": "easyocr", "name": "EasyOCR", "type": "multi-language", "status": "available" if EASY_AVAILABLE else "not installed", "size_mb": 500, "best_for": "Multi-language text"},
             {"id": "trocr", "name": "TrOCR", "type": "handwriting", "status": "available", "size_mb": 1500, "best_for": "Handwriting recognition"},
-            {"id": "surya", "name": "Surya OCR", "type": "layout-aware", "status": "available" if device.type == "cuda" else "disabled (CPU)", "size_mb": 800, "best_for": "Layout-aware, GPU only"},
+            {"id": "surya", "name": "Surya OCR", "type": "layout-aware", "status": "available" if (_HAS_TORCH and hasattr(device, 'type') and device.type == "cuda") else "disabled (CPU)", "size_mb": 800, "best_for": "Layout-aware, GPU only"},
             {"id": "auto", "name": "Ensemble (Auto)", "type": "ensemble", "status": "available", "size_mb": 2000, "best_for": "Best result from all engines"},
         ]
     }

requirements.txt

@@ -1,6 +1,26 @@
+# ============================================
+# HF Space — Medical OCR Demo (CPU Free Tier)
+# ============================================
+# Core framework
 gradio>=5.0.0,<7.0.0
+fastapi>=0.110.0
+uvicorn[standard]>=0.29.0
+python-multipart>=0.0.9
 Pillow>=10.0.0
 opencv-python-headless>=4.8.0
+numpy>=1.26.0,<2.0.0
 pytesseract>=0.3.10
 PyMuPDF>=1.24.0
-numpy>=1.26.0
+
+# OCR Engines — lightweight for CPU free tier
+paddleocr>=2.7.0
+paddlepaddle>=2.5.0
+
+# Optional: torch for TrOCR (CPU-only build)
+# TrOCR is available but loads lazily only when selected
+torch --extra-index-url https://download.pytorch.org/whl/cpu
+transformers>=4.36.0
+sentencepiece>=0.1.99
+
+# Utilities
+rapidfuzz>=3.0.0