Volta-Forever Brings CUDA 13 Support to NVIDIA V100

That’s where the community stepped in. NVIDIA’s V100 is now nearly a decade-old GPU, and CUDA 13 has officially dropped support for it.

Someone released Volta-Forever, a CUDA compatibility layer that sits between a CUDA 13 program and NVIDIA’s driver, rewriting GPU code at load time for sm_70, without modifying the program itself or touching the driver. Result? latest versions of vLLM and SGLang, both built for CUDA 13, can reportedly run directly on a V100. So can today’s unmodified PyTorch wheels and llama.cpp builds.

GitHub address: https://github.com/nbald/volta-forever

Invoking the layer requires no recompile, no patching, and no forked dependencies. Just two letters before the 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

That’s the entire workflow change. Everything else stays as-is.

NVIDIA’s V100 has been a workhorse across data centers, university clusters, and research labs for the better part of a decade. It doesn’t retire on a hyperscaler’s upgrade schedule. For organizations still running V100 inventory — whether that’s a cost decision, a procurement delay, or simply hardware that still performs, CUDA 13 dropped Volta support at an inconvenient moment.

Practical consequence isn’t just inconvenience. CUDA 12.x remains the last toolkit with first-class V100 support, and frameworks like vLLM and TensorRT 11.4 now require compute capability 7.5, which the V100 doesn’t reach natively. Without a workaround, those organizations face a narrowing software window with no driver-level solution from NVIDIA.

Volta-Forever goes beyond basic CUDA 13 binary compatibility. It also adds support for Tensor Core instructions the V100’s hardware never natively supported, FP8 and NVFP4, along with BF16, by emulating them on what the card does have. Version 0.1.1 further ensures that CUDA driver and NVML queries report a consistent compatibility capability of 7.5, which directly affects how vLLM selects quantization methods, including AWQ.

That’s a meaningful addition. FP8 and NVFP4 are precision formats that modern inference frameworks increasingly rely on for performance. Getting those instructions to execute on a V100, even through emulation, means the latest vLLM and SGLang builds aren’t just technically loading on the card; they’re running with the precision formats they were designed to use.

Installation runs through a single curl command. It places Volta-Forever under ~/.local/share/volta-forever, drops the vf binary into ~/.local/bin, and downloads NVIDIA’s cuBLAS for CUDA 12.9.2 directly from NVIDIA’s servers — approximately 1 GB. That download is necessary because cuBLAS built for CUDA 13 dropped every sm_70 kernel variant, so the layer substitutes a CUDA 12 build behind CUDA 13’s interface. No system packages are installed, no elevated permissions required. The script will surface a clear error if curl, python3, or xz aren’t present.

Hard requirements: a V100, NVIDIA driver 580, and Linux x86-64 running glibc 2.34 or later. Ubuntu 22.04 and Debian 12 both qualify.

Project ships with a LIMITS.md documenting what doesn’t work, what runs slower, and what hasn’t been tested. When a workload exceeds what the layer supports, it refuses and states why, no silent failures. Multi-GPU tensor parallelism through CUDA 13’s NCCL runs into multimem instructions with no Volta equivalent; that use case currently requires routing NCCL through a CUDA 12 build.

This isn’t an NVIDIA product, and NVIDIA hasn’t endorsed it. It’s a third-party shim with a 302-test validation suite, defined scope, and documented edges. For V100 operators watching their software options contract with each new CUDA major release, it’s a practical bridge, not a permanent fix, but a real one.

CUDA 13 dropped Volta, but Volta-Forever makes sure V100 drops nothing.

Latest

vivo Buds Clip Review: Comfort, Sound and Features Tested

Clip-on earbud market has gotten crowded fast. What started...

Disc Image Studio Turns Blank CDs and DVDs Into Laser-Etched Artwork

There’s something oddly satisfying about finding a practical use...

REDMI Clip Neo Review: Comfort, Battery Life & Sound Quality Tested

Clip-on earbud market has been quietly heating up, and...

REDMI K100 Review: Snapdragon 8 Elite, D2 Chip and 10,000mAh Battery

Smartphone prices have been climbing steadily, and flagship hardware...

Newsletter

Don't miss

vivo Buds Clip Review: Comfort, Sound and Features Tested

Clip-on earbud market has gotten crowded fast. What started...

Disc Image Studio Turns Blank CDs and DVDs Into Laser-Etched Artwork

There’s something oddly satisfying about finding a practical use...

REDMI Clip Neo Review: Comfort, Battery Life & Sound Quality Tested

Clip-on earbud market has been quietly heating up, and...

REDMI K100 Review: Snapdragon 8 Elite, D2 Chip and 10,000mAh Battery

Smartphone prices have been climbing steadily, and flagship hardware...

OnePlus 16 Review: The Battery Beast That Also Happens to Be a Great Phone

OnePlus has never been shy about iterating fast, OnePlus...
jerrywanint
jerrywanint
Work through curiosity, & the internet.

vivo Buds Clip Review: Comfort, Sound and Features Tested

Clip-on earbud market has gotten crowded fast. What started as a niche alternative to in-ear designs has become a full-on battleground, with every major...

Disc Image Studio Turns Blank CDs and DVDs Into Laser-Etched Artwork

There’s something oddly satisfying about finding a practical use for technology you thought was obsolete. Disc Image Studio, an open-source project that’s been gaining...

REDMI Clip Neo Review: Comfort, Battery Life & Sound Quality Tested

Clip-on earbud market has been quietly heating up, and REDMI’s latest entry, REDMI Clip Neo — arrives with a clear agenda: deliver genuine comfort,...