CUDA base image, pre-cache models, fix double loading

- Switch to nvidia/cuda:12.1.1-cudnn8-runtime-ubuntu22.04 for GPU support
- Pre-download BiRefNet and ISNet models in Docker build
- Use --noreload to prevent Django autoreload double model loading
- Fix apps.py preload condition for --noreload mode
- Update hardware options table with current HF pricing

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

DEPLOY.md

@@ -48,11 +48,17 @@ git push
 
 ## Hardware Options
 
-| Hardware | VRAM | Speed | Cost |
-|----------|------|-------|------|
-| CPU (free) | - | ~30-60s | Free |
-| T4 small | 16GB | ~2-5s | $0.40/hr |
-| A10G small | 24GB | ~1-3s | $1.05/hr |
+| Hardware | Specs | VRAM | Speed | Cost |
+|----------|-------|------|-------|------|
+| CPU basic | 2 vCPU · 16 GB RAM | - | ~30-60s | Free |
+| CPU upgrade | 8 vCPU · 32 GB RAM | - | ~15-30s | $0.03/hr |
+| T4 small | 4 vCPU · 15 GB RAM | 16GB | ~2-5s | $0.40/hr |
+| T4 medium | 8 vCPU · 30 GB RAM | 16GB | ~2-5s | $0.60/hr |
+| L4 | 8 vCPU · 30 GB RAM | 24GB | ~1-3s | $0.80/hr |
+| A10G small | 4 vCPU · 15 GB RAM | 24GB | ~1-3s | $1.00/hr |
+| A10G large | 12 vCPU · 46 GB RAM | 24GB | ~1-3s | $1.50/hr |
+| L40S | 8 vCPU · 62 GB RAM | 48GB | ~0.5-2s | $1.80/hr |
+| A100 large | 12 vCPU · 142 GB RAM | 80GB | ~0.5-1s | $2.50/hr |
 
 Enable GPU: Settings → Hardware → Select GPU → Save
 

Dockerfile

@@ -1,57 +1,55 @@
 # Hugging Face Spaces Dockerfile
 # GPU-enabled Django API for background removal
 
-FROM python:3.11-slim
+FROM nvidia/cuda:12.1.1-cudnn8-runtime-ubuntu22.04
 
 # Set environment variables
 ENV PYTHONDONTWRITEBYTECODE=1
 ENV PYTHONUNBUFFERED=1
 ENV DEBIAN_FRONTEND=noninteractive
-
-# HF Spaces runs on port 7860
 ENV PORT=7860
+ENV OMP_NUM_THREADS=4
 
-# Install system dependencies
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    libgl1 \
-    libglib2.0-0 \
-    libsm6 \
-    libxext6 \
-    libxrender1 \
-    libgomp1 \
-    curl \
-    && rm -rf /var/lib/apt/lists/*
-
-# Create app directory
-WORKDIR /app
+# Install Python 3.11 and system dependencies
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends software-properties-common && \
+    add-apt-repository ppa:deadsnakes/ppa && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends \
+    python3.11 python3.11-venv python3.11-distutils \
+    libgl1 libglib2.0-0 libsm6 libxext6 libxrender1 libgomp1 curl && \
+    rm -rf /var/lib/apt/lists/* && \
+    ln -sf /usr/bin/python3.11 /usr/bin/python && \
+    ln -sf /usr/bin/python3.11 /usr/bin/python3 && \
+    curl -sS https://bootstrap.pypa.io/get-pip.py | python
 
-# Copy requirements first for caching
-COPY requirements.txt .
+WORKDIR /app
 
 # Install Python dependencies
+COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
-# Copy application code
-COPY . .
-
 # Create non-root user (HF Spaces requirement)
 RUN useradd -m -u 1000 user
-RUN chown -R user:user /app
-USER user
-
-# Set home for model cache
 ENV HOME=/home/user
 ENV HF_HOME=/home/user/.cache/huggingface
 
-# Fix OMP_NUM_THREADS invalid value from HF
-ENV OMP_NUM_THREADS=4
+# Pre-download AI models so they don't re-download on every restart
+USER user
+RUN mkdir -p /home/user/.u2net /home/user/.config/Ultralytics && \
+    python -c "from rembg import new_session; new_session('birefnet-general', providers=['CPUExecutionProvider']); new_session('isnet-general-use', providers=['CPUExecutionProvider'])"
+
+# Copy application code
+USER root
+COPY . .
+RUN chown -R user:user /app
+USER user
 
-# Expose port
 EXPOSE 7860
 
-# Health check (longer start period for model loading)
-HEALTHCHECK --interval=30s --timeout=30s --start-period=300s --retries=5 \
+# Health check (shorter start period since models are pre-cached)
+HEALTHCHECK --interval=30s --timeout=30s --start-period=120s --retries=5 \
     CMD curl -f http://localhost:7860/api/transparent/health/ || exit 1
 
-# Start server
-CMD ["python", "manage.py", "runserver", "0.0.0.0:7860"]
+# --noreload prevents double model loading from Django's autoreload
+CMD ["python", "manage.py", "runserver", "0.0.0.0:7860", "--noreload"]

transparent/apps.py

@@ -14,11 +14,16 @@ class TransparentConfig(AppConfig):
 
     def ready(self):
         """Preload AI models when the app starts."""
-        # Only preload in the main process (not in manage.py commands)
-        if os.environ.get('RUN_MAIN') == 'true' or os.environ.get('GUNICORN_WORKER'):
+        import sys
+        is_runserver = 'runserver' in sys.argv
+        is_gunicorn = os.environ.get('GUNICORN_WORKER')
+        # With --noreload: RUN_MAIN is not set, load in main process
+        # Without --noreload: RUN_MAIN='true' in the reloader child process
+        is_reloader_child = os.environ.get('RUN_MAIN') == 'true'
+
+        if is_gunicorn or is_reloader_child or (is_runserver and '--noreload' in sys.argv):
             from transparent.controllers.style_generator import StyleGenerator
             from transparent.controllers.style_recommender import StyleRecommender
 
-            # Preload all AI models at startup
             StyleGenerator.preload_models()
             StyleRecommender.preload_models()