chore: set HF caches to /data/hf and install accelerate for faster model load

Dockerfile

@@ -5,7 +5,11 @@ FROM python:3.10-slim
 ENV PIP_NO_CACHE_DIR=1 \
     PYTHONDONTWRITEBYTECODE=1 \
     PYTHONUNBUFFERED=1 \
-    HF_HUB_ENABLE_HF_TRANSFER=1
+    HF_HUB_ENABLE_HF_TRANSFER=1 \
+    HF_HOME=/data/hf \
+    TRANSFORMERS_CACHE=/data/hf \
+    DIFFUSERS_CACHE=/data/hf \
+    HUGGINGFACE_HUB_CACHE=/data/hf
 
 WORKDIR /app
 
@@ -29,10 +33,12 @@ RUN python -m pip install --upgrade pip && \
     # CPU-only torch to avoid CUDA/xformers issues on Spaces CPU images
     python -m pip install --index-url https://download.pytorch.org/whl/cpu torch torchvision torchaudio --upgrade --force-reinstall && \
     # Ensure xformers is not installed
-    python -m pip uninstall -y xformers || true
+    python -m pip uninstall -y xformers || true && \
+    # Speed up and lower RAM during model load
+    python -m pip install accelerate
 
 EXPOSE 7860
 
 ENV DATA_DIR=/data
-RUN mkdir -p /data && chmod -R 777 /data
+RUN mkdir -p /data /data/hf && chmod -R 777 /data /data/hf
 CMD ["python", "-m", "uvicorn", "server:app", "--host", "0.0.0.0", "--port", "7860"]