[AMD] Add Claude skills for AMD CI workflows

[michaelzhang-ai]

Motivation

AMD model enablement and CI debugging workflows follow repeatable patterns that have been applied to 15+ models (MiniMax-M2.5, GLM-5, Kimi-K2.5, DeepSeek-V3.2, Qwen-3.5, etc.). Formalizing them as Claude Code skills lets AI agents execute these workflows consistently without rediscovering patterns each time.

Modifications

Add two Claude Code skills under .claude/skills/amd/, complementing the existing upstream skills (add-jit-kernel, add-sgl-kernel, sglang-bisect-ci-regression, write-sglang-test):

Skill	Lines	What it does
enable-amd-model	219	End-to-end: HF architecture research, AMD backend auto-selection (aiter/triton/wave/NSA with head-count logic from _get_default_attn_backend), accuracy test files for MI30x + MI35x, CI YAML updates (2 files x 3 edit locations), docs, local validation
write-amd-nightly-test	294	Guide for all AMD test patterns: standalone GSM8K, shared few_shot_gsm8k evaluator, LMEvalMixin, VLM MMMU, NightlyBenchmarkRunner perf tests. Covers register_amd_ci suites, all 6 runner labels, MI30x/MI35x variants, and backend args by architecture (MHA/MLA/MoE/NSA/mixed prefill-decode)

Key details verified against codebase

• Backends: Full ATTENTION_BACKEND_CHOICES list from server_args.py; auto-selection logic for MHA (aiter on HIP), MLA (depends on num_kv_heads / tp_size being 16 or 128), NSA (tilelang default on HIP)
• Runner labels: All 6 labels from workflow YAML (linux-mi325-{1,2,8}gpu-sglang, linux-mi35x-gpu-{1,8}, linux-mi35x-gpu-8.fabric)
• AttentionArch: Only MLA and MHA exist (from model_config.py); NSA is a backend, not an architecture
• Suite names: Documented actual naming patterns from 50+ register_amd_ci() calls across the codebase
• Test patterns: All 6 types (standalone GSM8K, shared evaluator, LMEvalMixin, VLM MMMU, NightlyBenchmarkRunner, diffusion) with reference template files

Files

.claude/skills/amd/enable-amd-model/SKILL.md        # NEW (219 lines)
.claude/skills/amd/write-amd-nightly-test/SKILL.md   # NEW (294 lines)

Checklist

• All backend choices verified against server_args.py (ATTENTION_BACKEND_CHOICES, NSA_CHOICES)
• Runner labels verified against workflow YAML files
• Suite names verified against register_amd_ci() calls in test files
• Test patterns verified against actual test file templates
• Both skills under 500 lines (219 + 294 = 513 total)
• Follows existing .claude/skills/ SKILL.md format

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces three new Claude Code skills specifically designed to streamline and support AMD-related continuous integration (CI) workflows. These skills aim to empower developers to efficiently enable new models for AMD nightly testing, diagnose and resolve CI regressions, and author robust accuracy and performance tests for AMD MI30x and MI35x GPUs. The addition of these skills enhances the maintainability and reliability of AMD CI by providing structured, automated guidance for common development tasks.

Highlights

• New Claude Skill: Enable AMD Nightly Model: A new skill has been added that provides an end-to-end workflow for enabling new models in AMD nightly CI. This includes steps for architecture research, AMD backend selection (aiter/triton/NSA), creating accuracy test files for MI30x and MI35x, updating CI workflow YAMLs, and documenting the process.
• New Claude Skill: Fix AMD CI Regression: A skill dedicated to debugging and fixing AMD-specific CI failures caused by upstream changes has been introduced. It outlines how to trace errors in aiter/triton backends, cross-reference upstream PRs, and generate targeted fixes, especially when one backend is updated but another is missed.
• New Claude Skill: Write AMD Nightly Test: This new skill guides users on writing AMD nightly accuracy and performance tests for MI30x and MI35x platforms. It covers various test types, CI suite registration using register_amd_ci, suite naming conventions, patterns for GSM8K accuracy tests, and considerations for MI30x vs MI35x differences.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Changelog

• .claude/skills/amd/enable-amd-nightly-model/SKILL.md
  • Added a new skill documentation for enabling models in AMD nightly CI.
• .claude/skills/amd/fix-amd-ci-regression/SKILL.md
  • Added a new skill documentation for debugging and fixing AMD CI regressions.
• .claude/skills/amd/write-amd-nightly-test/SKILL.md
  • Added a new skill documentation for writing AMD nightly tests.

Activity

• The pull request was created by michaelzhang-ai.
• No additional comments or reviews have been posted yet.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces three comprehensive Claude skills designed to streamline AMD-specific CI workflows. These skills provide detailed, step-by-step guides for enabling new models, debugging CI regressions, and authoring new tests, complete with code snippets and best practices. The documentation is well-structured and will be a valuable resource for developers working on AMD integration. I've identified a minor inconsistency in one of the new skill files regarding the setup of the HuggingFace cache environment, and I've provided a suggestion to align it with the other documentation and existing code.

_{Note: Security Review has been skipped due to the limited scope of the PR.}