vllm pod: the cu129 build of vLLM — the PyPI wheel is CUDA 13 and RunPod hosts run 12.8 drivers

The pod died with 'The NVIDIA driver on your system is too old (found
version 12080)': vLLM 0.27.1 on PyPI pins torch 2.13+cu130. vLLM's own
index has a +cu129 variant with torch's cu129 wheels, and CUDA 12.x
minor-version compatibility runs a 12.9 runtime on a 12.8 driver. uv
resolves it with --index-strategy unsafe-best-match, since the torch
index also carries stale copies of ordinary packages.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FKm7Z56iKjHgNkHc6VhWKz

packaging/runpod/Dockerfile.capability

@@ -87,7 +87,10 @@ RUN set -eux; \
             # forty exact pins without installing a byte; uv resolved the same
             # set in seconds. A `<profile>.venv-<name>.uv` marker asks for it.
             python -m pip install --no-cache-dir uv; \
-            uv pip install --python "$vdir/bin/python" --no-cache -r "$extra"; \
+            # best-match across indexes: the torch index also carries stale copies
+            # of ordinary packages, and first-index-wins would pin those.
+            uv pip install --python "$vdir/bin/python" --no-cache \
+                --index-strategy unsafe-best-match -r "$extra"; \
         else \
             "$vdir/bin/pip" install --no-cache-dir -r "$extra"; \
         fi; \

packaging/runpod/profiles/vllm.venv-vllm.txt

-# The version the local vLLM node runs (requirements-vllm.txt). vLLM pins its
-# own torch and CUDA libraries, which is why it cannot share the main venv.
-vllm==0.27.1
+# The version the local vLLM node runs (requirements-vllm.txt) — but NOT the
+# PyPI wheel: that one is built against CUDA 13 (torch 2.13+cu130), and a pod
+# whose host driver is CUDA 12.8 dies at engine start with "The NVIDIA driver
+# on your system is too old (found version 12080)". vLLM publishes a +cu129
+# variant on its own index, with torch's cu129 wheels; CUDA 12.x minor-version
+# compatibility runs a 12.9 runtime on a 12.8 driver. Resolved with uv (the
+# .uv marker): pip cannot cope with the pin set. The venv is ~10.5 GB (torch's
+# CUDA libraries, of which only cusolver is unused at import) and the image
+# ~12 GB; if RunPod refuses it, run this model with venv_on_volume instead.
+--extra-index-url https://wheels.vllm.ai/0.27.1/cu129
+--extra-index-url https://download.pytorch.org/whl/cu129
+vllm==0.27.1+cu129
+torch==2.13.0+cu129
+torchvision==0.28.0+cu129
+torchaudio==2.11.0+cu129