Update model workflow: local venv

Author: Mark Saroufim

.github/workflows/nvidia_model_workflow.yml

@@ -20,9 +20,6 @@ jobs:
     env:
       GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       GITHUB_REPOSITORY: ${{ github.repository }}
-      # Persistent venv with torch + vLLM deps (survives across runs).
-      # Bootstrap once, then reuse. The user's vLLM fork is installed fresh each run.
-      MODEL_VENV: /opt/model-venv
     steps:
     - uses: actions/checkout@v3
 
@@ -42,40 +39,28 @@ jobs:
     - name: Install uv
       uses: astral-sh/setup-uv@v4
 
-    - name: Setup persistent model venv
+    - name: Setup environment
       shell: bash
       run: |
-        # Create persistent venv if it doesn't exist, with torch + vLLM deps pre-installed.
-        # Mirrors the Modal model_image: install vllm to get all deps, then uninstall vllm.
-        if [ ! -f "$MODEL_VENV/bin/activate" ]; then
-          echo "Bootstrapping persistent model venv..."
-          uv venv "$MODEL_VENV" --python 3.10
-          export VIRTUAL_ENV="$MODEL_VENV"
-          export PATH="$MODEL_VENV/bin:$PATH"
-          uv pip install torch==2.9.1 --index-url https://download.pytorch.org/whl/cu130
-          uv pip install numpy transformers tokenizers huggingface_hub ray \
-            uvicorn fastapi pydantic aiohttp requests packaging ninja wheel sccache
-          uv pip install vllm && uv pip uninstall vllm
-          echo "Persistent model venv bootstrapped."
-        else
-          echo "Reusing existing persistent model venv."
-        fi
+        uv venv .venv --python 3.10
+        echo "VIRTUAL_ENV=$PWD/.venv" >> $GITHUB_ENV
+        echo "$PWD/.venv/bin" >> $GITHUB_PATH
 
-        # Activate for subsequent steps
-        echo "VIRTUAL_ENV=$MODEL_VENV" >> $GITHUB_ENV
-        echo "$MODEL_VENV/bin" >> $GITHUB_PATH
+        # Install torch first (build dep for vLLM)
+        uv pip install torch==2.9.1 --index-url https://download.pytorch.org/whl/cu130
 
-    - name: Install kernelbot
-      shell: bash
-      run: |
+        # Install vLLM to pull in all transitive deps, then remove vllm itself.
+        # The user's fork gets installed fresh by the benchmark runner.
+        uv pip install vllm && uv pip uninstall vllm
+
+        # Install kernelbot
         uv pip install -r "requirements-dev.txt"
         uv pip install -e .
 
     - name: Run model benchmark
       shell: bash
       env:
         SETUPTOOLS_SCM_PRETEND_VERSION: "0.0.1.dev0"
-        SCCACHE_DIR: /opt/sccache
       run: |
         python3 src/runners/github-runner.py