Your Tesla V100 is not dead: run unmodified CUDA 13 programs (PyTorch, vLLM, SGLang, llama.cpp) on Volta (sm_70). A compatibility layer in front of NVIDIA's driver library.
// momentum
// why it's here
- Created 1 day ago and already shipping
- Built almost entirely by one person
- Health 80/100 with only 40 stars
// how to help
// readme
volta-forever
Your Tesla V100 is not dead. CUDA 13 dropped Volta. volta-forever brings it back.
Run unmodified CUDA 13 programs on a V100: todayβs PyTorch wheels, vLLM, SGLang, llama.cpp. No patch, no rebuild. Put two letters in front of your command:
vf python3 -c "import torch; print(torch.cuda.get_device_name())" # Tesla V100-SXM2-32GB
vf llama-completion -m model.gguf -p "Hello" -n 48 -ngl 99
For a model demo, see the Ubuntu guides.
What a GPU of 2017 lacks is emulated on what it has: the tensor-core instructions of Turing, Ampere and Blackwell, so your V100 now speaks BF16, FP8 and NVFP4.
Tested on Tesla V100-SXM2-32GB cards with NVIDIAβs driver 580.178.04. What it cannot run it refuses with a reason;
LIMITS.mdhas what does not work, what is slower and what was not tried. Not NVIDIAβs, and not endorsed by NVIDIA.