The Pulse
vLLM 0.33 Drops CUDA 12.9 Builds
vLLM 0.33.0, released November 2, removes CUDA 12.9 wheels and container images, leaving CUDA 13.0 as the project’s only published CUDA build for the release. The update also moves vLLM to PyTorch 2.14 and adds model support, serving featur
AI.info Team · Written with AI · fact-checked ·
vLLM’s 0.33.0 release, published November 2, ends the project’s CUDA 12.9 builds: its NVIDIA wheels and container images now target CUDA 13.0 only. The change makes CUDA 13 the sole CUDA option in the new release, though it does not make CUDA 13 mandatory for every way of running vLLM. The project continues to publish builds for AMD ROCm, Intel XPU and CPU systems.
CUDA 13 was already the default
The shift is a removal, not a first adoption. vLLM 0.32.0, released before 0.33.0, already offered CUDA 13.0 as its default while retaining separate CUDA 12.9 wheels and Docker images. Version 0.33.0 drops those alternatives, including the image tags for CUDA 12.9 and CUDA 12.9 with Ubuntu 24.04.
For users of the published CUDA packages, the practical change is straightforward: the standard v0.33.0 Python install, pip install vllm, uses the CUDA 13.0 build, as does the default Docker image, vllm/vllm-openai:v0.33.0. The release also provides CUDA 13.0 wheels for x86_64 and Arm64. Teams that need the project’s CUDA 12.9 artifacts cannot select them from the v0.33.0 release.
PyTorch 2.14 joins the CUDA shift
The release updates vLLM to PyTorch 2.14.0, alongside torchvision 0.29.0 and Triton 3.8.0. It also describes the CUDA wheel as ABI-stable with libtorch, a change to the compatibility interface between the wheel and PyTorch’s C++ libraries. Together, those changes make this release a substantial update to the software stack, rather than a container-tag change alone.
CUDA 13.0 is not the only platform represented in the release. vLLM lists ROCm 10.0 as the default for AMD builds while continuing to publish ROCm 7.2 artifacts; CPU and XPU builds also remain available. The release therefore narrows the supported published CUDA choices without removing vLLM’s non-CUDA distribution paths.
More than a packaging change
The project says v0.33.0 contains 927 commits from 389 contributors, including 122 new contributors. Its highlights include support for new models such as Reflection Beam-1.0, Yandex AliceAI-Foundation-80B-A3B and Muse Spark Open, along with changes to speculative decoding and serving. An AWS EFA container image is also available for x86_64 and Arm64 systems.
Some performance figures in the notes are tied to specific hardware and benchmarks, not general promises. For example, the project reports that a change to DFlash and DSpark draft-slot handling increased output from 1,373 to 3,390 tokens per second in a benchmark using DeepSeek-V4.1-Flash with DSpark on GB300. A separate opt-in technique, FLy, is listed with decode-throughput gains of 1.04 to 1.31 times on MI355X and B300.
The compatibility boundary is now explicit
CUDA 13.0 had already become vLLM’s default CUDA build in version 0.32.0; v0.33.0 makes it the only CUDA version for which the project publishes release wheels and images. That distinction matters for operators planning upgrades: the release preserves several hardware platforms, but not the prior CUDA 12.9 package route. The release’s concrete dividing line is the artifact list: CUDA 13.0 for NVIDIA, with ROCm, XPU and CPU packages still listed separately.