Installation

uv venv -p 3.12 && source .venv/bin/activate
uv pip install flide torchvision

Flide runs on Linux with an NVIDIA GPU, Python 3.10 or newer and PyTorch 2.12 or newer. Every family in Supported models comes inside the same package. torchvision is only needed for images and video. Install it from the same place as torch, so the two builds match.

Drivers that run CUDA 12

PyTorch from PyPI is built for CUDA 13. On a machine whose NVIDIA driver runs only CUDA 12, Flide stops at startup and says so. Install the CUDA 12.6 build of PyTorch first, then Flide, which keeps the PyTorch it finds.

uv venv -p 3.12 && source .venv/bin/activate
uv pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
uv pip install flide

Kernels

Flide's kernels are compiled for each GPU. The package carries them precompiled for sm90 GPUs (H100, H200) and sm86 GPUs (A40, A10, RTX 30 series), for PyTorch 2.13 and 2.14. On these, a server starts without compiling. On any other GPU, the first start compiles the kernels once and keeps them in the Triton cache; later starts reuse them. That first start says so in its log and reports its progress at every quarter of the CUDA graph capture.

flide kernels packs compiled kernels into a bundle and files a bundle into another machine's cache, so a fleet of the same GPU compiles once.

Command	What it does
flide kernels export --model M --out B	warms the model in a fresh cache and packs the kernels into bundle B
flide kernels seed B	files bundle B into this machine's kernel cache
flide kernels merge OUT B1 B2	unites bundles made with the same Triton on the same GPU
flide kernels info B	lists what a bundle holds

Check the install

flide formats

The command lists the installed model families: intern_decision, next_token for chat models, onejev, open_jev and torchcast.