Issue Station
Sign In
Tools hereToolbox

pytorch - 💡(How to fix) Fix On the future of triton-windows [4 comments, 4 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
pytorch/pytorch#178368•Fetched 2026-04-08 01:25:59
View on GitHub
Comments
4
Participants
4
Timeline
100
Reactions
3
Author
woct0rdhowoct0rdho
Participants
desertfiredesertfiremalfetmalfetnWEIdianWEIdiawoct0rdhowoct0rdho
Timeline (top)
subscribed ×54mentioned ×30labeled ×9commented ×4

Root Cause

Currently Windows users have to either manually install a triton-windows wheel that matches the torch version, or rely on third-party all-in-one distributions or installation scripts. It's proven hard because concepts like venv and compiler toolchain are unintuitive to many Windows users. Make triton-windows a dependency of torch would lower this barrier and mark a step towards the overall maturity of PyTorch ecosystem on Windows.

RAW_BUFFERClick to expand / collapse

🚀 The feature, motivation and pitch

Do you think it's time to make triton-windows a dependency of torch, and enable CI tests for Windows + Nvidia and AMD GPUs in this repo?

It's been more than a year since I started maintaining triton-windows when the torchao-optimized 'fast Flux' became popular. Nowadays it works well on Nvidia GPUs from Turing to Blackwell, and AMD GPUs with CDNA3/RDNA3/RDNA4. It has been commonly used in projects such as ComfyUI and Unsloth.

The triton-windows repo has moved to the triton-lang organization and it's no longer a one-man project of mine, thanks to AMD, Intel, and more effort in the community. We've set up auditable workflows to build wheels and run CI tests. We're working towards production readiness on AMD RDNA3 consumer GPUs, see https://github.com/triton-lang/triton-windows/issues/2 (and the backend is already well tested on CDNA3 and RDNA4). Also, Intel has announced experimental support of Triton on Windows in their fork, and has been testing the backend-agnostic part of Triton on Windows.

However, the core Triton developers are a small team focusing on the compiler, with limited bandwidth to merge Windows support, so I think triton and triton-windows will remain two packages in the foreseeable future. (Or maybe this can be resolved?)

Recently Qualcomm developers brought up this discussion again, as they're interested in Windows ARM support, see https://github.com/triton-lang/triton-windows/issues/25 .

Alternatives

Currently Windows users have to either manually install a triton-windows wheel that matches the torch version, or rely on third-party all-in-one distributions or installation scripts. It's proven hard because concepts like venv and compiler toolchain are unintuitive to many Windows users. Make triton-windows a dependency of torch would lower this barrier and mark a step towards the overall maturity of PyTorch ecosystem on Windows.

Additional context

No response

cc @seemethere @malfet @atalman @tinglvv @nWEIdia @peterjc123 @mszhanyi @skyline75489 @nbcsm @iremyux @Blackhex @pytorch/pytorch-dev-infra @chauhang @penguinwu

extent analysis

Fix Plan

To make triton-windows a dependency of torch and enable CI tests for Windows + Nvidia and AMD GPUs, follow these steps:

  • Update torch dependencies:
    • Add triton-windows to the setup.py file in the torch repository.
    • Specify the version of triton-windows that matches the torch version.
  • Modify CI configuration:
    • Update the CI workflow to include Windows tests for Nvidia and AMD GPUs.
    • Use a CI tool like GitHub Actions to run tests on Windows environments.
  • Implement Windows-specific tests:
    • Write test cases for Windows-specific functionality in triton-windows.
    • Use a testing framework like Pytest to run tests.

Example code for updating setup.py:

from setuptools import setup

setup(
    # ...
    install_requires=[
        # ...
        'triton-windows==0.1.0',  # specify the version of triton-windows
    ],
)

Example CI workflow configuration:

name: Windows Tests

on:
  push:
    branches:
      - main

jobs:
  windows-tests:
    runs-on: windows-latest
    steps:
      - name: Checkout code
        uses: actions/checkout@v2
      - name: Install dependencies
        run: |
          pip install torch triton-windows
      - name: Run tests
        run: |
          pytest tests/windows

Verification

To verify that the fix worked, check the following:

  • The triton-windows dependency is correctly installed when installing torch.
  • CI tests for Windows + Nvidia and AMD GPUs pass.
  • Windows-specific tests in triton-windows pass.

Extra Tips

  • Ensure that the triton-windows version is compatible with the torch version.
  • Use a consistent versioning scheme for triton-windows and torch.
  • Monitor CI test results to catch any issues with Windows support.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#installation#inference speed#output truncation#response parsing#generation error

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING