ROCm / ROCm/aiter

[Feature]: Add head_dim=256 support to fmha_v3_fwd kernels

Open
#2,013 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
565
Forks
585
Avg merge
3d 4h
Merged PRs (30d)
366

Description

Suggestion Description

Current fmha v3 kernel only supports head_dim=128 or 192. We have observed significant performance gains when using the v3 compared to v2 kernel. But models like Qwen3-Next with head_dim=256 are not able to benefit from the perf improvement.

Operating System

No response

GPU

No response

ROCm Component

No response

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 locating the fmha_v3_fwd kernels and reviewing how head_dim=128 and 192 are currently handled. The work is complete when head_dim=256 is supported by the v3 forward kernels and its correctness and performance can be verified against the existing v2 and v3 paths.

Written by the indexing model from the issue text.

Assessment

Domain
machine-learning, performance
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
30/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.