chore(build): remove vllm_version envvar and lock into templates

Signed-off-by: paperspace <29749331+aarnphm@users.noreply.github.com>

Author: aarnphm

openllm-python/src/_openllm_tiny/Dockerfile.j2

@@ -1,5 +1,5 @@
 {% extends bento_base_template %}
 {% block SETUP_BENTO_COMPONENTS %}
-{% call common.RUN(__enable_buildkit__) -%} {{ __pip_cache__ }} {% endcall -%} bash -c 'pip install --no-color --progress-bar off "vllm==${VLLM_VERSION}" || true'
+{% call common.RUN(__enable_buildkit__) -%} {{ __pip_cache__ }} {% endcall -%} bash -c 'pip install --no-color --progress-bar off "vllm==0.4.2" || true'
 {{ super() }}
 {% endblock %}

openllm-python/src/_openllm_tiny/_entrypoint.py

@@ -527,7 +527,6 @@ def build_command(
           else [],
           envs=[
             EnvironmentEntry(name='NVIDIA_DRIVER_CAPABILITIES', value='compute,utility'),
-            EnvironmentEntry(name='VLLM_VERSION', value='0.4.2'),
             EnvironmentEntry(name=HF_HUB_DISABLE_PROGRESS_BARS, value='TRUE'),
           ],
           description=service_readme,