From 353c219b84592cb0b7adec708e52d13909b7ffbd Mon Sep 17 00:00:00 2001 From: Taksh Date: Fri, 31 Jul 2026 13:15:31 +0300 Subject: [PATCH 1/3] ci: publish wheels for torch 2.11, 2.12.1, and 2.13.0 Extend the release matrix so PyPI/GitHub assets cover recent torch minors that currently force silent source builds (#999). --- .github/workflows/publish.yaml | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-) diff --git a/.github/workflows/publish.yaml b/.github/workflows/publish.yaml index 04440f407..9d9af2ea4 100644 --- a/.github/workflows/publish.yaml +++ b/.github/workflows/publish.yaml @@ -41,7 +41,7 @@ jobs: # manylinux docker image, but I haven't figured out how to install CUDA on manylinux. os: [ubuntu-22.04, ubuntu-22.04-arm] python-version: ["3.10", "3.11", "3.12", "3.13"] - torch-version: ["2.6.0", "2.7.1", "2.8.0", "2.9.1", "2.10.0"] + torch-version: ["2.6.0", "2.7.1", "2.8.0", "2.9.1", "2.10.0", "2.11.0", "2.12.1", "2.13.0"] cuda-version: ["11.8.0", "12.9.1", "13.0.1"] # We need separate wheels that either uses C++11 ABI (-D_GLIBCXX_USE_CXX11_ABI) or not. # Pytorch wheels currently don't use it, but nvcr images have Pytorch compiled with C++11 ABI. @@ -56,6 +56,12 @@ jobs: cuda-version: "11.8.0" - torch-version: "2.10.0" cuda-version: "11.8.0" + - torch-version: "2.11.0" + cuda-version: "11.8.0" + - torch-version: "2.12.1" + cuda-version: "11.8.0" + - torch-version: "2.13.0" + cuda-version: "11.8.0" # CUDA 13.0 is only supported by PyTorch 2.9+ - torch-version: "2.6.0" cuda-version: "13.0.1" @@ -78,6 +84,12 @@ jobs: cxx11_abi: "FALSE" - torch-version: "2.10.0" cxx11_abi: "FALSE" + - torch-version: "2.11.0" + cxx11_abi: "FALSE" + - torch-version: "2.12.1" + cxx11_abi: "FALSE" + - torch-version: "2.13.0" + cxx11_abi: "FALSE" uses: ./.github/workflows/_build.yml with: runs-on: ${{ matrix.os }} From e18802c6dd089647408b101c8e9e39a94c94aad4 Mon Sep 17 00:00:00 2001 From: Taksh Date: Fri, 31 Jul 2026 13:15:31 +0300 Subject: [PATCH 2/3] docs: note how prebuilt wheel coverage tracks the release matrix Document that missing wheels fall back to selective_scan_cuda source builds and point maintainers at publish.yaml for the supported torch versions. --- README.md | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/README.md b/README.md index 2aa92d789..885ee517d 100755 --- a/README.md +++ b/README.md @@ -36,6 +36,13 @@ Install PyTorch first. By default, `mamba-ssm` installs the core package without | CUDA selective scan opt-in | `MAMBA_KEEP_CUDA_BUILD=TRUE pip install mamba-ssm --no-build-isolation` | Installs `selective_scan_cuda`; pip first tries a matching prebuilt CUDA/HIP wheel, then compiles locally if no wheel is available. | | Force local CUDA selective scan build | `MAMBA_FORCE_BUILD=TRUE MAMBA_KEEP_CUDA_BUILD=TRUE pip install mamba-ssm --no-build-isolation` | Skips cached wheels and compiles `selective_scan_cuda` locally. | +Prebuilt CUDA wheels (when using `MAMBA_KEEP_CUDA_BUILD=TRUE`) are published on +[GitHub releases](https://github.com/state-spaces/mamba/releases) for select +torch/CUDA/Python/platform combinations. The release CI matrix in +`.github/workflows/publish.yaml` defines which torch versions are built; if your +stack is newer than the latest release (e.g. torch 2.11+ on Colab), pip falls +back to a local `selective_scan_cuda` compile with no warning beyond setup logs. + `--no-build-isolation` is required for CUDA builds so that pip uses your existing CUDA-enabled PyTorch instead of installing torch-cpu in an isolated build environment. From d7ca4de3e8d6679a521c3b549d369390104be48d Mon Sep 17 00:00:00 2001 From: Taksh Date: Fri, 31 Jul 2026 13:18:05 +0300 Subject: [PATCH 3/3] docs: note publish matrix must stay aligned with setup wheel URLs Add a maintainer comment linking the CI torch list to get_wheel_url tags. --- .github/workflows/publish.yaml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.github/workflows/publish.yaml b/.github/workflows/publish.yaml index 9d9af2ea4..b5d26817d 100644 --- a/.github/workflows/publish.yaml +++ b/.github/workflows/publish.yaml @@ -41,6 +41,8 @@ jobs: # manylinux docker image, but I haven't figured out how to install CUDA on manylinux. os: [ubuntu-22.04, ubuntu-22.04-arm] python-version: ["3.10", "3.11", "3.12", "3.13"] + # Keep in sync with setup.py get_wheel_url() torch major.minor tags. + # New torch minors require a tagged release to publish matching wheels (#999). torch-version: ["2.6.0", "2.7.1", "2.8.0", "2.9.1", "2.10.0", "2.11.0", "2.12.1", "2.13.0"] cuda-version: ["11.8.0", "12.9.1", "13.0.1"] # We need separate wheels that either uses C++11 ABI (-D_GLIBCXX_USE_CXX11_ABI) or not.