Inference engine compatibility

CUDA/ROCm support for the tools people actually use to run models locally or in production — vLLM, llama.cpp, Ollama, and ExLlama — as distinct from the training-focused PyTorch/TensorFlow tables elsewhere on this site. vLLM ships real pinned CUDA-wheel releases, so it gets a version history below like those two; the other three don't version that way (see each section's note), so they get a single current-state entry instead.

vLLM

VersionReleasedCUDAROCmStatus
0.28.02026-08-2613.07.0, 7.2.1current
0.20.02026-04-2713.0—historical
0.11.12025-11-1812.9—historical
0.9.02025-05-1512.8—historical
0.8.52025-04-2812.4—historical

llama.cpp

rolling release (no numbered versions)current

Released
N/A -- commits are tagged with incrementing build numbers (e.g. b10826 as of 2026-09-06), not dated releases
CUDA
12.4, 12.6, 12.8, 13.3, 13.4
ROCm
6.1, 7.0

Full details →

Ollama

0.33.3current

Released
2026-09-02
CUDA
—
ROCm
7.0

Full details →

ExLlamaV3

1.4.7current

Released
2026-09-05
CUDA
12.8, 13.2
ROCm
—

Full details →

ExLlamaV2

0.3.2 (last release)archived

Released
2025-07-13
CUDA
12.1
ROCm
—

Full details →

← PyTorch versions · ROCm on consumer GPUs →