No prebuilt wheels for sm_103 (B300) and sm_121 (GB10)

Problem

mamba-ssm does not ship prebuilt wheels for new NVIDIA GPU architectures:

• sm_103 (B300 SXM6, GB300)
• sm_121 (GB10, DGX Spark / Project DIGITS)

Users must compile from source, which requires:

1. Matching CUDA toolkit version (13.0+ for sm_121, 12.8+ for sm_103)
2. Setting TORCH_CUDA_ARCH_LIST correctly
3. Working around setuptools version conflicts

This is the same issue as #745 (RTX 5090 / sm_120) but for datacenter/devkit GPUs.

Reproduction

B300 (sm_103)

pip install mamba-ssm  # installs prebuilt wheel for sm_75-sm_90
python -c "import torch; from mamba_ssm import Mamba2"
# RuntimeError: CUDA error: no kernel image is available for execution on the device

Must compile from source:

git clone https://github.com/state-spaces/mamba.git --branch v2.3.2
cd mamba
export CUDA_HOME=/usr/local/cuda-13.1
export TORCH_CUDA_ARCH_LIST='10.0+PTX'
pip install . --no-build-isolation --no-deps --force-reinstall

GB10 (sm_121)

git clone https://github.com/state-spaces/mamba.git --branch v2.3.2
cd mamba
export CUDA_HOME=/usr/local/cuda-13.0
export TORCH_CUDA_ARCH_LIST='12.0+PTX'
pip install . --no-build-isolation --no-deps --force-reinstall

Proposed Fix

1. Add sm_103 and sm_121 to the prebuilt wheel CUDA arch list in setup.py
2. For sm_103: note that causal-conv1d issue #105 reports a numerical correctness bug when building sm_103 cubins with CUDA 13.x. The workaround is to build sm_100 + PTX (forward-compatible to sm_103). The same may apply to mamba-ssm — recommend building 10.0+PTX instead of 10.3
3. For sm_121: build with 12.0+PTX (forward-compatible from sm_120)

Environment

• mamba-ssm: 2.3.2.post1
• B300: CUDA 13.1, PyTorch 2.13.0+cu130
• GB10: CUDA 13.0, PyTorch cu130

Related

• Issue #745: RTX 5090 + CUDA 12.8 (sm_120) — same pattern, resolved by source compilation
• Issue #818: Build wheel for PyTorch 2.8 and 2.9 — mentions Blackwell support
• Issue #772: CUDA 13 support
• causal-conv1d issue #105: sm_103 numerical correctness bug (nvcc 13.x produces wrong output)