Skip to content

Integrate DS R1 GroupedTopk + Sigmoid Routing Into DS Routing#3522

Merged
k50112113 merged 18 commits into
mainfrom
amemoore/deepseek-R1-fp4-routing
Jun 6, 2026
Merged

Integrate DS R1 GroupedTopk + Sigmoid Routing Into DS Routing#3522
k50112113 merged 18 commits into
mainfrom
amemoore/deepseek-R1-fp4-routing

Conversation

@amirumoAMD
Copy link
Copy Markdown
Contributor

@amirumoAMD amirumoAMD commented Jun 3, 2026

Motivation

Integrate new GroupedTopk kernel with support for score_mode sigmoid into routing + unified routing function, renamed function to reflect change. For changes done in ROCm/ATOM#1044.

Technical Details

grouped_topk kernel added to topk files, integrated into grouped topk.

Test Plan

Pytest for kernel included in routing pytest file, lm_eval on atom deepseek r1 mxfp4

Test Result

Kernel pytests passing, lm_eval 98% for atom integration for deepseek r1 mxfp4.

Submission Checklist

@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jun 3, 2026

🏷️ CI Guide

Runs automatically on every PR:

  • ✅ Pre-checks (submodule verification, code formatting)
  • ✅ Aiter op tests (gfx942 + gfx950)
  • ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label Tests
ci:triton-300x Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 3522 --add-label <label>

@amirumoAMD amirumoAMD marked this pull request as ready for review June 3, 2026 16:21
@amirumoAMD amirumoAMD requested a review from a team June 3, 2026 16:21
@amirumoAMD amirumoAMD force-pushed the amemoore/deepseek-R1-fp4-routing branch from 3ecd064 to 63bca7d Compare June 3, 2026 21:23
@vgokhale vgokhale requested review from lburzawa and nsusanto June 4, 2026 02:05
Comment thread aiter/ops/triton/moe/moe_routing/routing.py Outdated
Copy link
Copy Markdown
Contributor

@k50112113 k50112113 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you also help check if the _grouped_topk kernel you added here is functionally the same as the one on the HIP side? no need to add unit test

@amirumoAMD amirumoAMD requested a review from k50112113 June 4, 2026 17:48
@amirumoAMD amirumoAMD force-pushed the amemoore/deepseek-R1-fp4-routing branch from 740a970 to 682213e Compare June 5, 2026 02:12
@amirumoAMD amirumoAMD dismissed k50112113’s stale review June 5, 2026 15:39

fixed recommended issues, dismissing stale review

@k50112113 k50112113 merged commit 3ffb5e4 into main Jun 6, 2026
55 checks passed
@k50112113 k50112113 deleted the amemoore/deepseek-R1-fp4-routing branch June 6, 2026 02:27
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.

3 participants