Docker file for model up

Dockerfile

@@ -0,0 +1,34 @@
+# Use a standard Python 3.12 image
+FROM python:3.12-slim
+
+WORKDIR /app
+
+# Install all dependencies
+RUN pip install "llama-cpp-python[server]" huggingface_hub
+
+# --- Model Download ---
+# Copy the download script into the container
+COPY download_model.py .
+
+# Make the HF_TOKEN secret available as an argument
+# This will be passed in by the HF Spaces platform
+ARG HF_TOKEN
+# Run the script to download the model
+RUN --mount=type=secret,id=HF_TOKEN \
+    python download_model.py
+
+# --- Server Runtime ---
+# Expose port 8000 (which we defined in README.md)
+EXPOSE 8000
+
+# This is the command that will run when the container starts
+# It reads the API_KEY secret from the environment
+CMD [ \
+    "python", \
+    "-m", "llama_cpp.server", \
+    "--model", "prem-1B-SQL.Q8_0.gguf", \
+    "--n_gpu_layers", "0", \
+    "--port", "8000", \
+    "--host", "0.0.0.0", \
+    "--api_key_env_var", "API_KEY" \
+]