Installation¶
Python Package¶
FlashInfer is available as a Python package, built on top of PyTorch to easily integrate with your Python applications.
Prerequisites¶
OS: Linux only
Python: 3.10, 3.11, 3.12, 3.13, 3.14
CUDA: 12.9, 13.0, 13.4 (PyTorch nightly)
Note
FlashInfer strives to follow PyTorch’s supported CUDA versions plus the latest CUDA release. CUDA 13.4 wheels are built with the preview toolkit and PyTorch nightly; runtime CI currently covers CUDA 12.9 and 13.0.
Quick Start¶
The easiest way to install FlashInfer is via pip. Please note that the package currently used by FlashInfer is named flashinfer-python, not flashinfer.
pip install flashinfer-python
Package Options¶
FlashInfer provides three packages:
flashinfer-python: Core package that compiles/downloads kernels on first use
Release wheels on PyPI leave out the kernel sources of the experimental features under
flashinfer.experimentalto stay under PyPI’s file size limit; those features need a nightly wheel from https://flashinfer.ai/whl or an install from source.flashinfer-cubin: Pre-compiled kernel binaries for all supported GPU architectures
flashinfer-jit-cache: CUDA-specific shim that installs architecture-specific pre-built kernel providers
For faster initialization and offline usage, install the optional packages to have most kernels pre-compiled:
pip install flashinfer-python
flashinfer download-kernels
This eliminates compilation and downloading overhead at runtime.
flashinfer download-kernels installs the flashinfer-cubin and
flashinfer-jit-cache wheels that match the installed flashinfer-python
version and the CUDA version of the installed PyTorch. By default it installs
pre-built kernels for every supported GPU architecture (several GB); pass
--mode minimal to install the jit-cache provider only for the visible GPUs,
or --sm (for example --sm sm90a, repeatable) to choose architectures
explicitly. See Command Line Interface for all options, and Upgrading before
upgrading an existing installation.
Upgrading¶
The pre-built kernel wheels are tied to one exact flashinfer-python
version, and upgrading flashinfer-python does not upgrade them. Re-install
them after every upgrade:
pip install -U flashinfer-python
flashinfer download-kernels # same --mode / --sm options as the first install
Until you do, FlashInfer logs a warning at import (Ignoring incompatible
flashinfer-cubin package / flashinfer-jit-cache package) and ignores the
old wheels: cubins are downloaded on demand and kernels are JIT-compiled, as if
the wheels were not installed. Results stay correct, but first calls are slow
again and offline machines can fail to load kernels. Set
FLASHINFER_DISABLE_VERSION_CHECK=1 only if you intentionally mix versions.
flashinfer-cubin moved off PyPI. flashinfer-cubin 0.6.13 was the last
release published to PyPI; newer releases are only on the FlashInfer wheel
index, https://flashinfer.ai/whl. Because PyPI still serves 0.6.13, these
commands silently install a stale wheel that FlashInfer then ignores:
pip install flashinfer-cubinorpip install -U flashinfer-cubinwithout--index-url https://flashinfer.ai/whla
requirements.txtor lock file that resolvesflashinfer-cubinfrom PyPI
Use flashinfer download-kernels, which picks the right index and pins the
matching version, or pin it yourself:
pip install flashinfer-cubin==<flashinfer-python version> --index-url https://flashinfer.ai/whl.
flashinfer-jit-cache was never published to PyPI; it is served from the
CUDA-specific index (for example https://flashinfer.ai/whl/cu130).
Run flashinfer show-config to compare the installed flashinfer-cubin
and flashinfer-jit-cache versions with flashinfer-python.
Install from Source¶
In certain cases, you may want to install FlashInfer from source code to try out the latest features in the main branch, or to customize the library for your specific needs.
You can follow the steps below to install FlashInfer from source code:
Clone the FlashInfer repository:
git clone https://github.com/flashinfer-ai/flashinfer.git --recursive
Make sure you have installed PyTorch with CUDA support. You can check the PyTorch version and CUDA version by running:
python -c "import torch; print(torch.__version__, torch.version.cuda)"
Install FlashInfer:
cd flashinfer python -m pip install -v .
For development, install in editable mode:
python -m pip install --no-build-isolation -e . -v
Note
When using
--no-build-isolation, pip does not automatically install build dependencies. FlashInfer requiressetuptools>=77. If you encounter an error likeAttributeError: module 'setuptools.build_meta' has no attribute 'prepare_metadata_for_build_editable', upgrade pip and setuptools first:python -m pip install --upgrade pip setuptools
(Optional) Build optional packages:
Build
flashinfer-cubin:python -m build --no-isolation --wheel flashinfer-cubin python -m pip install flashinfer-cubin/dist/*.whl
Build one JIT-cache provider for the target GPU, then build a shim that depends on that provider. The example below builds an SM90a provider; both wheels must use the same version settings:
export FLASHINFER_JIT_CACHE_PROVIDER_ARCH=9.0a python -m build --no-isolation --wheel flashinfer-jit-cache-provider export FLASHINFER_JIT_CACHE_PROVIDER_ARCHS="9.0a" python -m build --no-isolation --wheel flashinfer-jit-cache python -m pip install \ flashinfer-jit-cache-provider/dist/*.whl \ flashinfer-jit-cache/dist/*.whl
Install Nightly Build¶
Nightly builds are available for testing the latest features:
pip install -U --pre flashinfer-python --index-url https://flashinfer.ai/whl/nightly/ --no-deps # Install the nightly package from custom index, without installing dependencies
pip install flashinfer-python # Install flashinfer-python's dependencies from PyPI
flashinfer download-kernels --nightly # Optional: matching nightly flashinfer-cubin and flashinfer-jit-cache wheels
Verify Installation¶
After installation, verify that FlashInfer is correctly installed and configured:
flashinfer show-config
This command displays:
FlashInfer version and installed packages (flashinfer-python, flashinfer-cubin, flashinfer-jit-cache)
PyTorch and CUDA version information
Environment variables and artifact paths
Downloaded cubin status and module compilation status