NVIDIA / NVIDIA/cutlass

[QST][CuTeDSL] Is there a plan to support the tf32 type on hopper

Open
#2,887 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

? - Needs Triage inactive-30d inactive-90d question
Dominant language
C++
Stars
10.5k
Forks
2.1k
Avg merge
3d 11h
Merged PRs (30d)
7

Description

What is your question?

Is there a plan to support the tf32 type on hopper

In some scenarios, we hope to have higher precision and a wider range of numerical representation.

I see in the python/CuTeDSL/cutlass/utils/hopper helpers. Py "this directory make trivial tiled mma doesn't support TF32 type, It USES the "python/CuTeDSL/cutlass/express/nvgpu/warpgroup/mma. py" under the encapsulated mma OP. The content here seems not to allow users to develop. I noticed the balckwell has supported, hooper architecture on whether there is a support plan

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reading the referenced python/CuTeDSL/cutlass/utils/hopper helpers and python/CuTeDSL/cutlass/express/nvgpu/warpgroup/mma.py, then compare the existing Blackwell support. Determine with maintainers whether Hopper TF32 support is planned and what implementation scope is required; done would be documented support or a confirmed plan with specific next steps.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp, python
Domain
performance
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Quiet
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.