Skip to content

PyPI wheels broken on aarch64 (DGX Spark / ARM Grace) #891

@Sggin1

Description

@Sggin1

pip install mamba-ssm on aarch64 + PyTorch 2.10+ fails for two reasons:

  1. Wrong architecture. The wheel ships selective_scan_cuda.cpython-312-x86_64-linux-gnu.so — can't load on aarch64.
  2. ABI mismatch. Even with MAMBA_FORCE_BUILD=TRUE, the PyPI release (v2.3.1) compiles against c10::cuda::CUDAStream::query(), a symbol removed in PyTorch 2.10+. Same issue affects causal-conv1d from PyPI.

Working fix: Install from GitHub main branch which has the ABI fixes:

export MAMBA_FORCE_BUILD=TRUE
export CAUSAL_CONV1D_FORCE_BUILD=TRUE
pip install --no-build-isolation "causal-conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git"
pip install --no-build-isolation "mamba-ssm @ git+https://github.com/state-spaces/mamba.git"

Tested on: DGX Spark GB10 (aarch64, SM121), nvcr.io/nvidia/pytorch:25.11-py3, PyTorch 2.10, CUDA 13.0. Successfully loads Nemotron-Nano-12B-v2-VL-BF16 (13.2B params) on GPU.

Dockerfile with working build: https://github.com/Sggin1/spark-ai-containers/tree/main/mamba-dev

Note: Tested on one configuration only (DGX Spark, pytorch:25.11-py3). I'm a hobbyist sharing what worked for me — there may be better approaches I'm not aware of.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions