Experimental Windows MSLK build for NVIDIA Blackwell (SM120) GPUs. Includes a validated Triton Split-K adjustment for head_dim=512 workloads used by xFormers.
This repository publishes documentation, validation scripts, checksums, and release metadata for an experimental Windows-compatible MSLK build validated on NVIDIA Blackwell laptop hardware.
The first release target is:
mslk-1.1.0.post1+sm120a1111-py3-none-any.whl
The wheel is not stored directly in this repository. It should be uploaded as a GitHub Release asset.
This is an experimental compatibility build. It is not official upstream MSLK Windows support and should not be treated as a general-purpose MSLK distribution.
Use it only if you understand the risks of custom GPU runtime dependencies, including unsupported kernel behavior, version mismatch, numerical instability, and app-specific validation gaps.
Windows x64
GPU: NVIDIA GeForce RTX 5070 Laptop GPU
Compute Capability: 12.0 / SM120
Python: 3.10.6
PyTorch: 2.11.0+cu128
CUDA Runtime: 12.8
CUDA Toolkit: 12.8.61
MSLK: 1.1.0.post1+sm120a1111
Primary validation app: AUTOMATIC1111 Stable Diffusion WebUI
Companion validation: xFormers 0.0.35+af84367e
Upstream MSLK does not currently provide an official Windows-supported wheel for this use case.
This repository provides a Windows-compatible MSLK build validated on NVIDIA Blackwell SM120 hardware. The build was needed to support a companion xFormers Blackwell Windows build, and validation was performed on RTX 5070 Laptop GPU hardware.
This repository does not imply official upstream support.
The functional source change is intentionally minimal.
File:
mslk/attention/fmha/triton_splitk.py
Original:
BLOCK_N: int = 64Modified:
# Lowered for Windows Blackwell laptop GPUs where head_dim=512 exceeds shared memory with 64.
BLOCK_N: int = 32This lowers the Triton Split-K block size. It reduces shared-memory pressure for the validated RTX 5070 Laptop / SM120 case. The tested workload involved head_dim=512.
The change is intentionally narrow and should be treated as experimental.
GitHub Release asset:
mslk-1.1.0.post1+sm120a1111-py3-none-any.whl
Do not commit wheel files into this repository. Release binaries should be attached to the GitHub Release.
Known SHA256 checksum:
682cd2c4b85c6ef192ff9f1009c3b2af079bdcd006c54e4348c33da0c9763e58 mslk-1.1.0.post1+sm120a1111-py3-none-any.whl
Verify on Windows with:
certutil -hashfile mslk-1.1.0.post1+sm120a1111-py3-none-any.whl SHA256The printed hash must match the value above.
Install MSLK alone:
python -m pip uninstall -y mslk
python -m pip install --no-deps mslk-1.1.0.post1+sm120a1111-py3-none-any.whlFor use with the companion xFormers wheel:
python -m pip uninstall -y xformers mslk
python -m pip install --no-deps mslk-1.1.0.post1+sm120a1111-py3-none-any.whl
python -m pip install --no-deps xformers-0.0.35+af84367e.d20260613-py39-none-win_amd64.whlMSLK was validated through the xFormers memory-efficient attention path used by A1111.
Run validation from the target virtual environment:
python tests\test_xformers_a1111_shape.py
python tests\test_xformers_a1111_tuned.pyThe required success condition is float16: PASS. The expected full-precision result is float32: FAIL.
This repository includes:
tests/test_xformers_a1111_shape.py
tests/test_xformers_a1111_tuned.py
These scripts are included because MSLK was validated through the companion xFormers attention path. The tuned script applies the known Split-K settings used for the RTX 5070 Laptop / SM120 / head_dim=512 validation case.
Torch: 2.11.0+cu128
CUDA: 12.8
GPU: NVIDIA GeForce RTX 5070 Laptop GPU
Compute Capability: (12, 0)
Testing torch.float16
PASS torch.Size([1, 9600, 1, 512]) torch.float16 finite=True
Testing torch.float32
FAIL
Summary:
float16: PASS
float32: FAIL
float16: PASS is the required success condition. float32: FAIL is expected. The validated path is intended for half precision attention. Full precision float32 attention is not supported by the validated route.
A1111 runtime validation:
AUTOMATIC1111 was launched with --xformers.
WebUI launched successfully.
Model loaded successfully.
Generation completed successfully.
No xFormers runtime errors were observed.
Fast iteration testing completed.
This was validation of the full MSLK + xFormers stack, not MSLK alone.
MSLK is a required companion dependency for the validated xFormers Blackwell Windows build.
This repository publishes the MSLK side separately because users may search specifically for Windows MSLK support.
Companion repository:
xformers-blackwell-windows
https://github.com/kittensx/xformers-blackwell-windows
Install this MSLK wheel before installing the companion xFormers Blackwell wheel.
Recommended install order:
1. mslk
2. xformers
3. application validation
- This is not official upstream MSLK Windows support.
- RTX 5070 Laptop GPU only validated.
- Windows only validated.
- Float32 attention unsupported in the validated path.
- No Linux testing.
- No multi-GPU testing.
- No training validation.
- No guarantee for all RTX 50-series cards.
- No guarantee for all xFormers workloads.
- RTX 5060 validation.
- RTX 5070 desktop validation.
- RTX 5080 validation.
- RTX 5090 validation.
- Additional CUDA versions.
- Additional PyTorch versions.
- Upstream Windows compatibility investigation.
- Performance benchmarking.
- VRAM benchmarking.
Credit belongs to the upstream MSLK, PyTorch, Triton, CUDA, and xFormers maintainers. This repository documents an experimental Windows packaging and validation path for a specific Blackwell laptop environment.
This repository distributes documentation, validation scripts, and release metadata for an experimental Windows MSLK build. It does not relicense upstream MSLK or xFormers projects. See LICENSE_NOTES.md.