[CI] Add nightly Python 3.14 smoke test (#60456)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

.buildkite/test_areas/misc.yaml

@@ -446,6 +446,17 @@ steps:
       - setup.py
       - vllm/platforms/rocm.py
 
+- label: ":nvidia: (H200 MIG 18GB) Python 3.14 Smoke Nightly"
+  key: python-314-smoke-nightly
+  depends_on: ~
+  optional: true
+  timeout_in_minutes: 30
+  device: h200_18gb
+  source_file_dependencies:
+  - tests/standalone_tests/python_only_compile.sh
+  commands:
+  - PYTHON_ONLY_COMPILE_PYTHON=3.14 bash standalone_tests/python_only_compile.sh
+
 - label: ":nvidia: (H200 MIG 35GB) Async Engine, Inputs, Utils, Worker"
   device: h200_35gb
   key: async-engine-inputs-utils-worker

tests/standalone_tests/python_only_compile.sh

@@ -1,6 +1,10 @@
 #!/bin/bash
 # This script tests if the python only compilation works correctly
 # for users who do not have any compilers installed on their system
+#
+# Set PYTHON_ONLY_COMPILE_PYTHON (e.g. 3.14) to install into a fresh uv venv
+# with that interpreter and run a short generation, as a smoke test for
+# Python versions other than the one the CI image is built with.
 
 set -e
 
@@ -173,9 +177,12 @@ elif [[ ! -d vllm ]]; then
     exit 1
 fi
 
-# remove all compilers
-apt remove --purge build-essential -y
-apt autoremove -y
+# remove all compilers, except for the smoke test, whose generation needs a
+# C compiler at runtime for torch.compile and Triton
+if [[ -z "${PYTHON_ONLY_COMPILE_PYTHON:-}" ]]; then
+    apt remove --purge build-essential -y
+    apt autoremove -y
+fi
 
 rm -f /tmp/changed.file
 echo 'import os; os.system("touch /tmp/changed.file")' >> vllm/__init__.py
@@ -185,12 +192,42 @@ if [[ -n "${rocm_wheel}" ]]; then
     VLLM_PRECOMPILED_WHEEL_LOCATION="${rocm_wheel}" VLLM_USE_PRECOMPILED=1 python3 setup.py develop --no-deps
 elif [[ "${is_rocm}" == "1" ]]; then
     VLLM_PRECOMPILED_WHEEL_COMMIT=$merge_base_commit VLLM_USE_PRECOMPILED=1 python3 setup.py develop --no-deps
+elif [[ -n "${PYTHON_ONLY_COMPILE_PYTHON:-}" ]]; then
+    # Use the CUDA build the image's own torch was installed with, for both the
+    # new venv's torch and the precompiled wheel variant.
+    cuda_variant="$(python3 -c 'import torch; print("cu" + torch.version.cuda.replace(".", ""))')"
+    uv venv --python "${PYTHON_ONLY_COMPILE_PYTHON}" /tmp/python-only-venv
+    source /tmp/python-only-venv/bin/activate
+    VLLM_PRECOMPILED_WHEEL_COMMIT=$merge_base_commit VLLM_USE_PRECOMPILED=1 \
+        VLLM_PRECOMPILED_WHEEL_VARIANT="${cuda_variant}" \
+        uv pip install -e . --torch-backend="${cuda_variant}"
 else
     VLLM_PRECOMPILED_WHEEL_COMMIT=$merge_base_commit VLLM_USE_PRECOMPILED=1 pip3 install -vvv -e .
 fi
 # Run the script
 python3 -c 'import vllm'
 
+if [[ -n "${PYTHON_ONLY_COMPILE_PYTHON:-}" ]]; then
+    python3 - <<'PY'
+import os
+import sys
+
+from vllm import LLM
+
+expected = os.environ["PYTHON_ONLY_COMPILE_PYTHON"]
+actual = f"{sys.version_info.major}.{sys.version_info.minor}"
+assert actual == expected, f"expected Python {expected}, got {actual}"
+
+llm = LLM(
+    "hmellor/tiny-random-LlamaForCausalLM",
+    max_model_len=256,
+    gpu_memory_utilization=0.3,
+)
+outputs = llm.generate(["Hello, my name is"])
+assert outputs[0].outputs[0].token_ids, "no tokens generated"
+PY
+fi
+
 # Check if the clangd log file was created
 if [ ! -f /tmp/changed.file ]; then
     echo "ERROR: changed.file was not created, python only compilation failed"