Support CONTEXT_LENGTH env var (default 4096) for ctx-size

start.sh

@@ -10,8 +10,23 @@ export NUMEXPR_NUM_THREADS="$THREADS"
 export HF_HOME="${HF_HOME:-/data/hf-cache}"
 export MODEL_DIR="${MODEL_DIR:-/data/models}"
 export MODEL_PATH="${MODEL_PATH:-/data/models/model.gguf}"
+export CTX_SIZE="${CONTEXT_LENGTH:-${CTX_SIZE:-4096}}"
 
-python3 /app/download_model.py
+# Ensure data directories exist and are writable. Run as root in container startup.
+mkdir -p "$HF_HOME" "$MODEL_DIR"
+# try to chown to appuser; if that fails (readonly mount), relax permissions
+if chown -R 65532:65532 "$HF_HOME" "$MODEL_DIR" 2>/dev/null; then
+  true
+else
+  chmod -R a+rwx "$HF_HOME" "$MODEL_DIR" || true
+fi
+
+# Run downloader as the non-root `appuser` using the venv python when available.
+if [ -x "/opt/venv/bin/python" ]; then
+  su -s /bin/sh appuser -c "/opt/venv/bin/python /app/download_model.py"
+else
+  su -s /bin/sh appuser -c "python3 /app/download_model.py"
+fi
 
 LLAMA_SERVER_BIN="${LLAMA_SERVER_BIN:-/usr/local/bin/llama-server}"
 
@@ -36,4 +51,5 @@ if [[ -n "${PARALLEL:-}" ]]; then
   server_args+=(--parallel "$PARALLEL")
 fi
 
-exec "$LLAMA_SERVER_BIN" "${server_args[@]}"
+server_cmd=("$LLAMA_SERVER_BIN" "${server_args[@]}")
+exec su -s /bin/sh appuser -c "$(printf '%q ' "${server_cmd[@]}")"