Hello,
The vLLM source at present lists a hard dependency on torch==2.9.0 for CUDA: vllm/requirements/cuda.txt at main · vllm-project/vllm · GitHub
I have a working image where I built a few wheels and it has Torch 2.9.1. Is it possible to build vllm for it by simply editing that requirement (and the corresponding torchaudio and torchvision), or will that break things?
Simply editing the requirement in requirements/cuda.txt from torch==2.9.0 to torch==2.9.1 (and updating torchaudio/torchvision) is not guaranteed to work, as vLLM tightly couples its CUDA/C++ kernels and dependencies to specific PyTorch versions for binary compatibility. There is no official documentation or guarantee that vLLM built for torch==2.9.0 will work with torch==2.9.1, and attempts to update PyTorch minor versions have previously caused CI failures and required further dependency updates and testing before merging, as seen in recent PRs and issues. Compatibility is only ensured for the versions explicitly listed in the requirements files and tested in CI/CD pipelines; using unlisted versions may result in build or runtime errors due to ABI or API changes in PyTorch or its dependencies.
If you want to attempt this, you must update all related dependencies (torch, torchvision, torchaudio, and possibly xformers) to versions compatible with torch==2.9.1, and be prepared to resolve any build or runtime issues that arise. However, this is not officially supported or recommended unless the vLLM maintainers have merged and tested such a change in the main branch. For production or reliable use, it is best to stick with the officially supported torch version specified in requirements/cuda.txt or wait for an official update to torch==2.9.1 support in vLLM. See the discussion and failed attempts in update to 2.9.1 PyTorch release and Install pre-built xformers-0.0.32.post2 built with pt-2.9.0.
Would you like a step-by-step breakdown of what would be required to attempt this update?
Sources:
Hint: Mention @RunLLM in the post for followups.