---
license: apache-2.0
tags:
- sageattention
- attention
- blackwell
- consumer-blackwell
- dgx-spark
- gb10
- sm120
- sm121
- sm121a
- aarch64
- cuda13
- pytorch-2.10
- pytorch-2.11
library_name: sageattention
---

# SageAttention 2.2 — DGX Spark aarch64 / CUDA 13 wheel

A community build of [thu-ml/SageAttention](https://github.com/thu-ml/SageAttention) v2.2.0 targeting **NVIDIA DGX Spark (GB10, sm_121)** on Linux aarch64 with CUDA 13. Also applies to any future aarch64 Linux host running consumer Blackwell silicon — SASS-level testing confirms the kernels emitted for this build execute byte-identically on RTX 50-series (sm_120a) and DGX Spark (sm_121a).

## What this wheel is

- SageAttention v2.2.0, unmodified upstream source (no patches, no forks)
- Built with `TORCH_CUDA_ARCH_LIST="12.0"` against:
  - Python 3.12
  - PyTorch 2.10 / 2.11 (verify your install matches)
  - CUDA 13.x toolchain
  - Linux aarch64 (Ubuntu 24.04, glibc 2.39)
- `cuobjdump --list-text` confirms kernels target `sm_120`

## Why `sm_120` and not `sm_121`?

Upstream `setup.py`'s `SUPPORTED_ARCHS` set does not include `"12.1"`, so the build accepts `"12.0"` as the closest available target. This is the correct target, not a workaround: SASS-level testing on real hardware confirms that `sm_120` (RTX 50-series, consumer Blackwell) and `sm_121` (DGX Spark, GB10) emit **byte-identical SASS** for the consumer-Blackwell tensor-core paths used by SageAttention 2.2.

Evidence: [Sggin1/DGX-SPARK/fp4_SASS](https://github.com/Sggin1/DGX-SPARK/tree/main/fp4_SASS) — a SASS-encoder-level investigation of FP4 MMA support on Spark. Test B in `findings.md` demonstrates that the consumer-Blackwell warp-level MMA path (`mma.sync.aligned.kind::mxf4.block_scale...`) produces identical encoding `0x70f00e0c0808747f` on both `sm_120a` and `sm_121a`. The same equivalence applies to the FP8 and INT8 paths exercised by SageAttention 2.2's `_qattn_sm89` kernels.

## Install

```bash
pip install https://huggingface.co/Sggin/sage-attention-2.2-dgx-spark-aarch64-cu13/resolve/main/sageattention-2.2.0-cp312-cp312-linux_aarch64.whl
```

(Adjust the repo URL after first push.)

## Verify

```bash
python -c "import sageattention; print(sageattention.__file__)"

/usr/local/cuda/bin/cuobjdump --list-text \
  $(python -c 'import sageattention,os; print(os.path.dirname(sageattention.__file__))')/_qattn_sm80*.so \
  | grep -oE "sm_[0-9]+" | sort -u
```

Expected output: `sm_120`.

## Reproduce the build

```bash
git clone https://github.com/thu-ml/SageAttention /tmp/sage22_build
cd /tmp/sage22_build
git checkout v2.2.0
TORCH_CUDA_ARCH_LIST="12.0" MAX_JOBS=8 python setup.py bdist_wheel
```

Wheel artifact lands in `dist/`.

Tested on:
- NVIDIA DGX Spark (GB10, sm_121, Ubuntu 24.04, driver 580.142)
- Python 3.12.x
- CUDA Toolkit 13.x
- PyTorch 2.10.x / 2.11.x (cu130 aarch64)

## SHA256

```
2e570e87b9f2283820ae9ee86b05d5acc468fa7bf814016863770d7264d65e05  sageattention-2.2.0-cp312-cp312-linux_aarch64.whl
```

## Source

Upstream: [github.com/thu-ml/SageAttention](https://github.com/thu-ml/SageAttention) (Apache 2.0)

This wheel is a build artifact of the upstream v2.2.0 release. No source modifications. License inherits Apache 2.0 from upstream — `LICENSE` in this repo is the upstream license file unchanged.

## See also

- [Sggin1/DGX-SPARK/fp4_SASS](https://github.com/Sggin1/DGX-SPARK/tree/main/fp4_SASS) — SASS-level investigation of consumer-Blackwell FP4 / FP8 paths on Spark
- [Sggin1/DGX-SPARK/comfyui_spark_notes](https://github.com/Sggin1/DGX-SPARK/tree/main/comfyui_spark_notes) — pip-shadowing gotchas to watch for when this wheel gets clobbered by transitive installs
- [Triplany/comfyui-dgx-spark](https://github.com/Triplany/comfyui-dgx-spark) — practical ComfyUI integration kit for DGX Spark (community kit that builds and verifies this same wheel)

## What this wheel does NOT do

- Does not enable the datacenter-Blackwell `tcgen05.mma` path (the "1 PFLOP FP4" marketing number). That instruction is `ptxas`-refused on `sm_120a` and `sm_121a` alike — it requires datacenter Blackwell silicon (B200, sm_100a). No software patch recovers it on consumer Blackwell. See `fp4_SASS/findings.md` Test A.
- Does not include any sm_121-specific patches to upstream SageAttention. The wheel is unmodified v2.2.0; the only difference from PyPI is the build target (`12.0`) and the host platform (aarch64 + CUDA 13).
- Does not address the SageAttention 3 mosaic-artifact issue on Spark ([thu-ml/SageAttention#321](https://github.com/thu-ml/SageAttention/issues/321)). That's a separate kernel-level issue in the Sage 3 Blackwell path.

## Disclaimer

Community wheel. Not an official NVIDIA, thu-ml, or HuggingFace release. Built on local hardware; verification output above is what the maintainer observed on their own Spark. Use at your own risk.