Update Dockerfile

Dockerfile

@@ -11,37 +11,66 @@ RUN apt-get update && apt-get install -y \
     libopenblas0 \
     && rm -rf /var/lib/apt/lists/*
 
-# Create writable directories for Feedback, Cache, and Logs
-RUN mkdir -p /app/data /app/logs /.cache /.cache/paddlepaddle && chmod -R 777 /app/data /app/logs /.cache /.cache/paddlepaddle
+# Create writable directories for model cache, feedback, and logs
+RUN mkdir -p /app/data /app/logs /.cache /.cache/paddlepaddle /.cache/huggingface \
+    && chmod -R 777 /app/data /app/logs /.cache
 
-# Copy requirements first
+# Copy requirements first (enables Docker layer caching)
 COPY requirements.txt .
 
-# Upgrade pip and install requirements with optimizations
+# Install PyTorch CPU wheel FIRST (needs --extra-index-url),
+# then install everything else from requirements.txt.
+# Splitting into two pip calls avoids conflict resolution issues.
 RUN pip install --no-cache-dir --upgrade pip setuptools wheel && \
+    pip install --no-cache-dir \
+        torch==2.4.0+cpu \
+        torchvision==0.19.0+cpu \
+        --extra-index-url https://download.pytorch.org/whl/cpu && \
     pip install --no-cache-dir -r requirements.txt
 
 # Copy the rest of the application
 COPY . .
 
-# Environment Variables (CPU optimization for 2GB RAM)
-ENV PYTHONUNBUFFERED=1 \
-    PORT=7860 \
-    TRANSFORMERS_CACHE=/.cache \
-    EASYOCR_CACHE=/.cache \
-    FEEDBACK_DATA_PATH=/app/data \
-    PADDLE_DOWNLOAD_CACHE=/.cache/paddlepaddle \
-    PADDLE_HOME=/.cache/paddlepaddle \
-    PADDLE_INFERENCE_MODEL_CACHE=/.cache/paddlepaddle/models \
-    OMP_NUM_THREADS=1 \
-    OPENBLAS_NUM_THREADS=1 \
-    MKL_NUM_THREADS=1 \
-    NUMEXPR_NUM_THREADS=1 \
-    DISABLE_TQDM=1 \
-    HF_HUB_DISABLE_PROGRESS_BARS=1
-
-# Hugging Face default port
+# ── Environment Variables ─────────────────────────────────────────────────────
+ENV PYTHONUNBUFFERED=1
+ENV PORT=7860
+
+# Shared cache dir for HF hub weights, EasyOCR models, PaddleOCR models
+ENV TRANSFORMERS_CACHE=/.cache
+ENV HF_HOME=/.cache
+ENV EASYOCR_CACHE=/.cache
+ENV FEEDBACK_DATA_PATH=/app/data
+ENV PADDLE_DOWNLOAD_CACHE=/.cache/paddlepaddle
+ENV PADDLE_HOME=/.cache/paddlepaddle
+ENV PADDLE_INFERENCE_MODEL_CACHE=/.cache/paddlepaddle/models
+
+# CPU thread caps (2 vCPU HF Spaces Basic tier)
+ENV OMP_NUM_THREADS=1
+ENV OPENBLAS_NUM_THREADS=1
+ENV MKL_NUM_THREADS=1
+ENV NUMEXPR_NUM_THREADS=1
+
+# Suppress verbose download progress in container logs
+ENV DISABLE_TQDM=1
+ENV HF_HUB_DISABLE_PROGRESS_BARS=1
+
+# ── VLM Defaults (override via HF Space Secrets/Env UI) ──────────────────────
+# Set ENABLE_VLM=0 to run EasyOCR+Paddle only (~700 MB RAM, no torch needed)
+ENV ENABLE_VLM=1
+ENV VLM_MODEL_ID=Qwen/Qwen2-VL-2B-Instruct
+ENV VLM_MAX_NEW_TOKENS=512
+# Abort VLM inference if it exceeds 60 s (falls back to EasyOCR)
+ENV VLM_TIMEOUT_SECONDS=60
+# Self-disable VLM if process RSS exceeds 12 GB (leaves 4 GB for OS/cache)
+ENV VLM_MEMORY_LIMIT_MB=12000
+
+# Hugging Face Spaces default port
 EXPOSE 7860
 
-# Fixed: Removed --timeout-notify
-CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860", "--workers", "1", "--timeout-keep-alive", "75"]
+# ── Startup command ───────────────────────────────────────────────────────────
+# workers=1: Qwen2-VL fp32 uses ~8 GB; two workers would OOM on 16 GB.
+# timeout-keep-alive=120: allows VLM cold-start (first request downloads weights).
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860", \
+     "--workers", "1", \
+     "--timeout-keep-alive", "120", \
+     "--timeout-notify", "60"]