Pre-built Linux wheels for Mamba, across Python, PyTorch, CUDA, and CPU architectures.
Following the PyTorch convention, artifacts are published to a separate index for each CUDA
version. Each wheel has a local version suffix that identifies the CUDA and PyTorch versions it was
built against, such as mamba-ssm==2.3.2.post1+cu.12.8.torch.2.11, and requires the matching PyTorch
release.
Pre-built wheels are available on Astral's GPU indexes. For example, to install a CUDA 12.8 build:
$ uv add mamba-ssm --index astral-cu128=https://wheels.astral.sh/simple/cu128/This configures the index and uses it as the source for mamba-ssm:
[tool.uv.sources]
mamba-ssm = { index = "astral-cu128" }
[[tool.uv.index]]
name = "astral-cu128"
url = "https://wheels.astral.sh/simple/cu128/"Or, with uv pip:
$ uv pip install --index https://wheels.astral.sh/simple/cu128/ mamba-ssmThe tests/ directory contains a locked uv project that installs the published
CUDA 12.8 Mamba and Causal Conv1d wheels from the Astral index alongside their
matching PyTorch build. Run the tests on a Modal GPU with:
$ modal run tests/modal_app.pyModal installs the locked dependencies in its Linux image and runs the pytest suite on an NVIDIA A10G. The CUDA wheels are not installed on the local machine.
Wheels are available for the following mamba-ssm versions:
The latest release, Mamba 2.3.2.post1, supports the following combinations:
| PyTorch | Python | x86_64 CUDA |
aarch64 CUDA |
|---|---|---|---|
| 2.4.1 | 3.9–3.12 | 12.1, 12.4 | — |
| 2.5.1 | 3.9–3.12 | 12.1, 12.4 | — |
| 2.6.0 | 3.9–3.12 | 12.4, 12.6 | 12.6 |
| 2.7.1 | 3.9–3.13 | 12.6, 12.8 | 12.8 |
| 2.8.0 | 3.9–3.13 | 12.6, 12.8, 12.9 | 12.9 |
| 2.9.1 | 3.10–3.13 | 12.6, 12.8, 12.9, 13.0 | 12.6, 12.8, 12.9, 13.0 |
| 2.10.0 | 3.10–3.14 | 12.6, 12.8, 12.9, 13.0 | 12.6, 12.8, 12.9, 13.0 |
| 2.11.0 | 3.10–3.14 | 12.6, 12.8, 12.9, 13.0 | 12.6, 12.8, 12.9, 13.0 |
| 2.12.1 | 3.10–3.14 | 12.6, 13.0, 13.2 | 12.6, 13.0, 13.2 |
build-mamba is licensed under the Apache License, Version 2.0.
