Skip to content

Add inferencex-sync skill for ATOM benchmark comparison and InferenceX PR creation#802

Merged
valarLip merged 5 commits into
ROCm:mainfrom
seungrokj:main
May 16, 2026
Merged

Add inferencex-sync skill for ATOM benchmark comparison and InferenceX PR creation#802
valarLip merged 5 commits into
ROCm:mainfrom
seungrokj:main

Conversation

@seungrokj
Copy link
Copy Markdown
Contributor

Summary

  • Add .claude/commands/inferencex-sync.md skill that automates comparing ATOM upstream benchmark results against InferenceX published numbers and creating PRs to update
    InferenceX configs

What the skill does

  1. Fetches InferenceX benchmark data (mi355x/atom/single-node) for all tracked models
  2. Pulls throughput data from the latest successful ATOM nightly benchmark run via the gh-pages dashboard
  3. Computes per-GPU throughput comparison and shows regression/improvement %
  4. Creates PRs to SemiAnalysisAI/InferenceX with updated docker images and serve arguments when ATOM upstream shows improvements

Models tracked

  • DeepSeek-V4-Pro (fp4), DeepSeek-R1-0528 (fp8, fp4), Kimi-K2.5-MXFP4, Qwen3.5-397B-A17B (fp8, fp4), GLM-5 (fp8, fp4), MiniMax-M2.7 (fp8, fp4), gpt-oss-120b (fp4)

Usage

/inferencex-sync

seungrokj and others added 5 commits May 15, 2026 12:39
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Renames the skill to better reflect its purpose: comparing ATOM upstream
benchmark results against InferenceX and reporting regression/improvement.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: seungrokj <seungrok.jung@amd.com>
…t run

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@seungrokj
Copy link
Copy Markdown
Contributor Author

@valarLip can you plz merge this ?

@valarLip valarLip merged commit 45c31f2 into ROCm:main May 16, 2026
2 checks passed
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