Skip to content

MoE AITER Triton Kernels Integration#1044

Open
amirumoAMD wants to merge 35 commits into
mainfrom
amemoore/gfx950-moe-triton-integration
Open

MoE AITER Triton Kernels Integration#1044
amirumoAMD wants to merge 35 commits into
mainfrom
amemoore/gfx950-moe-triton-integration

Conversation

@amirumoAMD
Copy link
Copy Markdown
Contributor

@amirumoAMD amirumoAMD commented Jun 2, 2026

Motivation

Replace triton_kernels module with aiter kernels. Add support for gpt-oss a8w4.

Technical Details

matmul_ogs now replaced by a16w4 moe gemm from aiter. custom routing redirects to updated expanded/unified aiter routing function. correlates to changes on aiter branch amemoore/gfx950-moe-triton-integration.

Test Plan

lm_eval matches expected result of non-triton run for mxfp4 weight models (DSr1 mxfp4, gpt-oss a8w4, gpt-oss regular).

Test Result

All three match

Submission Checklist

@amirumoAMD amirumoAMD force-pushed the amemoore/gfx950-moe-triton-integration branch from f18e6a3 to 1b9fd3b Compare June 2, 2026 17:30
@amirumoAMD amirumoAMD marked this pull request as ready for review June 2, 2026 17:47
@amirumoAMD amirumoAMD changed the title Amemoore/gfx950 moe triton integration MoE AITER Triton Kernels Integration Jun 3, 2026
Comment thread atom/model_ops/moe.py Outdated
@amirumoAMD amirumoAMD force-pushed the amemoore/gfx950-moe-triton-integration branch from 785a0bc to 71d4458 Compare June 5, 2026 21:27
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants