chore: vllm update v0.19.0

docker/vllm.Dockerfile

@@ -14,14 +14,22 @@ COPY --link . /daemon/
 
 WORKDIR /daemon/nilai-models/
 
+# Install daemon dependencies into an isolated venv, separate from vLLM's /opt/venv.
+# vLLM base image sets VIRTUAL_ENV=/opt/venv and puts it on PATH.
+# UV_PROJECT_ENVIRONMENT ensures uv sync targets the daemon's own .venv.
 RUN apt-get update && \
     apt-get install build-essential -y && \
     pip install uv && \
-    uv sync && \
+    UV_PROJECT_ENVIRONMENT=/daemon/nilai-models/.venv uv sync && \
     apt-get clean && \
     apt-get autoremove && \
     rm -rf /var/lib/apt/lists/*
 
+# Upgrade transformers in vLLM's /opt/venv (where pip/python3 resolve via PATH).
+# Gemma 4 architecture requires transformers>=5.5.0.
+# --no-deps avoids pulling transitive deps that conflict with vLLM's pins.
+RUN pip install 'transformers>=5.5.0' --no-deps && pip install 'huggingface-hub>=1.5.0,<2.0'
+
 # Expose port 8000 for incoming requests
 EXPOSE 8000
 

nilai-models/run.sh

@@ -10,7 +10,7 @@ start_primary_process() {
 
 start_secondary_process() {
     echo "Starting the secondary process"
-    uv run python3 -m nilai_models.daemon
+    UV_PROJECT_ENVIRONMENT=/daemon/nilai-models/.venv uv run python3 -m nilai_models.daemon
 }
 
 main() {