Keep image generic; move startup settings into run.sh/compose

run.sh

@@ -77,6 +77,7 @@ eval "${DETECT_OUTPUT}"
 MODEL_KIND="${DETECTED_MODEL_KIND}"
 IMAGE="${IMAGE:-llama4-bi150:2026-10-26}"
 NAME="${NAME:-llama4-${MODEL_KIND}}"
+HOST="${HOST:-0.0.0.0}"
 PORT="${PORT:-${DETECTED_PORT}}"
 MODEL_MOUNT="${MODEL_MOUNT:-/models/llama4}"
 SERVED_NAME="${SERVED_NAME:-${DETECTED_SERVED_NAME}}"
@@ -122,6 +123,9 @@ fi
 
 DOCKER_RUN=(
   docker run -d --name "${NAME}"
+  --entrypoint /usr/local/bin/llama4-entrypoint.sh
+  --workdir /opt/llama4
+  --expose "${PORT}"
   --privileged
   --pid host
   --network host
@@ -134,6 +138,7 @@ DOCKER_RUN=(
   -e MODEL_PATH="${MODEL_MOUNT}"
   -e AUTO_DETECT="${AUTO_DETECT}"
   -e MODEL_KIND="${MODEL_KIND}"
+  -e HOST="${HOST}"
   -e SERVED_NAME="${SERVED_NAME}"
   -e PORT="${PORT}"
   -e TP="${TP}"
@@ -142,6 +147,7 @@ DOCKER_RUN=(
   -e MAX_NUM_SEQS="${MAX_NUM_SEQS}"
   -e GPU_MEM_UTIL="${GPU_MEM_UTIL}"
   -e ENFORCE_EAGER="${ENFORCE_EAGER}"
+  -e PYTHONUNBUFFERED=1
   -e EXTRA_ARGS="${EXTRA_ARGS}"
   "${DOCKER_EXTRA[@]}"
   --restart unless-stopped