AMD port of TurboDiffusion - Working on gfx1151 on Windows - #66

couldnt git clone this branch, so i modified the files by hand
git clone --branch jam/windows_amd https://github.com/woct0rdho/triton-windows.git triton-windows-patch Cloning into 'triton-windows-patch'... fatal: Remote branch jam/windows_amd not found in upstream origin

@jammm

jammm commented Jan 5, 2026

Copy link

Copy Markdown

Author

3 Install Dependencies
pip install -r requirements.txt in which repo?

In Turbodiffusion repo (on this PR branch)

couldnt git clone this branch, so i modified the files by hand git clone --branch jam/windows_amd https://github.com/woct0rdho/triton-windows.git triton-windows-patch Cloning into 'triton-windows-patch'... fatal: Remote branch jam/windows_amd not found in upstream origin

ah, for triton-windows you just need pip install triton-windows. I'll update the README to mention this. My bad

@jammm

jammm force-pushed the jam/amd_windows branch from 0ae67c3 to d947d6d Compare

January 5, 2026 21:23

@githust66

githust66 commented Jan 6, 2026

Copy link

Copy Markdown

The step "pip install -r requirements.txt" should not be necessary, as the TurboDiffusion project does not contain a requirements.txt file; only the SpargeAttn project includes such a file.

@0xDELUXA

0xDELUXA commented Jan 6, 2026

Copy link

Copy Markdown

...though can possibly work on RDNA4 with minor modificaitons.

Sounds promising 👀

@jammm

jammm force-pushed the jam/amd_windows branch from d947d6d to 55a7ae6 Compare

January 6, 2026 11:20

@jammm

jammm commented Jan 6, 2026

Copy link

Copy Markdown

Author

The step "pip install -r requirements.txt" should not be necessary, as the TurboDiffusion project does not contain a requirements.txt file; only the SpargeAttn project includes such a file.

Fixed, thanks!

...though can possibly work on RDNA4 with minor modificaitons.

Sounds promising 👀

Yes, it's just a matter of refactoring the rocWMMA code to not assume that the per-thread matrix fragments are replicated across the half-waves. It's just another prompt to claude actually.

jammm and others added 4 commits

January 6, 2026 17:05

@jammm


 Add ROCm/HIP support for TurboDiffusion ops

24f0325

- Add HIP kernels for GEMM, LayerNorm, RMSNorm, and quantization ops
- Integrate rocWMMA for matrix operations on AMD GPUs
- Update setup.py for Windows ROCm builds with clang-cl
- Add platform detection (CUDA/HIP) with common abstractions
- Optimize SLA kernel config for ROCm (BLKK=16)
- Update .gitignore to exclude build artifacts and IDE files
- Fix distributed utils and network files for ROCm compatibility

@jammm


 Add AMD Windows setup guide

ffce8b9

@jammm


 Add numeric_conversion_hip header

4b7674c

@jammm


 Update README_AMD_WINDOWS.md

6199e67

@jammm

jammm force-pushed the jam/amd_windows branch from 55a7ae6 to 6199e67 Compare

January 6, 2026 11:42

@0xDELUXA

0xDELUXA commented Jan 7, 2026

Copy link

Copy Markdown

Yes, it's just a matter of refactoring the rocWMMA code to not assume that the per-thread matrix fragments are replicated across the half-waves. It's just another prompt to claude actually.

I'll be the 1st tester though, when the time comes ^^

Navigation Menu

Uh oh!

AMD port of TurboDiffusion - Working on gfx1151 on Windows - #66