Skip to content

[AMD] Add Claude skills for AMD CI workflows#8

Closed
michaelzhang-ai wants to merge 924 commits intomainfrom
amd/add-claude-skills
Closed

[AMD] Add Claude skills for AMD CI workflows#8
michaelzhang-ai wants to merge 924 commits intomainfrom
amd/add-claude-skills

Conversation

@michaelzhang-ai
Copy link
Owner

@michaelzhang-ai michaelzhang-ai commented Mar 8, 2026

Summary

Add three Claude Code skills that encode AMD-specific development workflows for SGLang, complementing the existing upstream skills (add-jit-kernel, add-sgl-kernel, sglang-bisect-ci-regression, write-sglang-test).

New Skills

Skill Lines What it does
enable-amd-nightly-model 193 End-to-end workflow: HuggingFace architecture research → AMD backend selection (aiter/triton/NSA) → accuracy test files for MI30x + MI35x → CI workflow YAML updates (2 files × 3 edit locations) → documentation → local validation
fix-amd-ci-regression 157 Debug AMD nightly CI failures caused by upstream changes. Covers the common pattern where a PR updates triton_backend.py but misses aiter_backend.py, plus classification of error types and fix patterns
write-amd-nightly-test 161 Guide for writing AMD nightly accuracy/performance tests with correct CI registration (register_amd_ci), suite naming, MI30x/MI35x platform variants, and GSM8K benchmark patterns

Motivation

These workflows have been used repeatedly for AMD model enablement (MiniMax-M2.5, GLM-5, Kimi-K2.5, DeepSeek-V3.2, Qwen-3.5, etc.) and bug fixes (PR sgl-project#19113, PR sgl-project#19736). Formalizing them as skills allows AI agents to execute them consistently without needing to rediscover the patterns each time.

Design

  • Follows the exact same SKILL.md format as existing upstream skills
  • All files under 200 lines (well within the 500-line guideline)
  • References real test files and workflow YAML patterns from the codebase
  • Includes decision trees (backend selection), file templates, and checklists

Test plan

  • Verify SKILL.md files render correctly on GitHub
  • Verify skill descriptions are specific enough for agent discovery
  • Validate referenced file paths exist in the codebase

SoluMilken and others added 30 commits March 1, 2026 14:09
@SoluMilken
[fix typo] seperated_timestep -> separated_timestep (sgl-project#19622)
0b3ddbc
@Kangyan-Zhou @claude
[Bugfix] Add missing auto_create_handle_loop to communicator methods (s…
98224de 
…gl-project#19610)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@glenliu21
[Test] add unit test for skipping already preempted request (sgl-proj…
8a0b757 
…ect#18912)
@zwang86 @gemini-code-assist @b8zhong
feat: add FA4 SM90 paged KV decode support & update attention docs (s…
f51ddba 
…gl-project#18442)

Co-authored-by: Zeyu Wang <zeyu.wang@yahooinc.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>
@SoluMilken @ishandhanani
[fix typo] expert_indicies -> expert_indices (sgl-project#19627)
20282f5 
Co-authored-by: ishandhanani <82981111+ishandhanani@users.noreply.github.com>
@alphabetc1
docs: refactor speculative decoding doc (sgl-project#19186)
e3e71f2
@wolfcomos
[WIP]enable mxfp8 on nvidia sm120 (sgl-project#19112)
e5edf22 
Co-authored-by: Your Name <you@example.com>
@ShangmingCai
[PD] Remove unused server args for disaggregation (sgl-project#19618)
0a6678b 
Signed-off-by: Shangming Cai <csmthu@gmail.com>
@yuan-luo
[JIT-kernel] Add unit test for nsa indexer fused_store_k_cache (sgl-p…
f6ee6dc 
…roject#19389)

Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
@Prozac614 @yingluosanqian @mickqian
[diffusion] model: support Hunyuan3D-2 (sgl-project#18170)
57c5c34 
Co-authored-by: yingluosanqian <yingluosanqian@gmail.com>
Co-authored-by: daiweitao <dwti614707404@163.com>
Co-authored-by: Mick <mickjagger19@icloud.com>
@Kangyan-Zhou @claude
Add bisect ci claude code skill (sgl-project#19649)
ec97754 
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@sglang-bot
[CI] Disable test_lora_tp CUDA CI during H100 to H200 transition (sgl…
0e53cee 
…-project#19654)
@Baidu-AIAK @hnyls2002
Cleanup disagg decode prebuilt flow and add cross-stream sync in merg…
922aad2 
…e_batch (sgl-project#19568)

Co-authored-by: vincent <vincent@vincentdeMacBook-Pro.local>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
@ispobock
Fix mamba2 mixer ci test (sgl-project#19658)
4726073
@YazhiGao
Remove sync points in mamba cache + prefill cudagraph plumbing for DP (…
07ef5f7 
…sgl-project#19639)
@BBuf
[Diffusion] diffusion profile and opt skills (sgl-project#19540)
e42fa00
@ishandhanani @claude
feat: TTL-based prefix pinning with refresh-on-hit for HiRadixCache (s…
f7da379 
…gl-project#18941)

Co-authored-by: Claude <noreply@anthropic.com>
@kkHuang-amd
Add aiter attention support in prefill-attention-backend of gpt-oss (s…
15af26d 
…gl-project#18282)

Co-authored-by: wunhuang <wunhuang@amd.com>
@yichiche
[AMD] Add Qwen3-Coder-Next accuracy and functionality test scripts fo…
98f47d8 
…r MI35x 8-GPU (sgl-project#18608)
@yctseng0211 @bingxche
[AMD] AMD AITER Scout Workflow (sgl-project#19467)
f2c5503 
Co-authored-by: Bingxu Chen <Bingxu.Chen@amd.com>
@mickqian
[diffusion] feat: Add --model-id for config resolution; deprecate mod…
2e15c01 
…el_detectors (sgl-project#19607)
@fzyzcjy
Support presets and arbitrary skipping keys in dump comparator (sgl-p…
ec44bc8 
…roject#19676)
@fzyzcjy
Enhance replication check, matching pattern, logging in dump comparat…
15e83ee 
…or (sgl-project#19677)
@fzyzcjy
Support flattened dims in dump comparator (sgl-project#19678)
a70dd11
@fzyzcjy
Support non orthogonal parallel axes and explicit replication annotat…
6980416 
…ion in dump comparator (sgl-project#19679)
@fzyzcjy
Support directory detection in dump comparator (sgl-project#19680)
abdc0ee
@fzyzcjy
Enhance sglang engine dumping tests in dump comparator (sgl-project#1…
3ebd85b 
…9681)
@fzyzcjy
Trace execution information in dump comparator (sgl-project#19682)
5bf3deb
@fzyzcjy
Beautify text output in dump comparator (sgl-project#19683)
3dd4649
@fzyzcjy
Support multiple verbosity in dump comparator (sgl-project#19684)
e5ef845
@michaelzhang-ai michaelzhang-ai force-pushed the amd/add-claude-skills branch 6 times, most recently from bb5f306 to 47109f9 Compare March 8, 2026 08:20
@michaelzhang-ai
[AMD] Add Claude skills for AMD CI workflows
4836f76 
Add two Claude Code skills under .claude/skills/amd/ that encode
AMD-specific development workflows, complementing the existing
upstream skills:

- enable-amd-nightly-model: End-to-end workflow for enabling a new
  model in AMD nightly CI (architecture research, backend selection
  with auto-detection logic, test files for MI30x/MI35x, CI YAML
  updates, validation)
- write-amd-nightly-test: Guide for writing AMD nightly accuracy
  and performance tests covering all test patterns (standalone GSM8K,
  shared evaluator, LMEvalMixin, VLM MMMU, NightlyBenchmarkRunner),
  CI registration, runner labels, and platform variants

All backend choices, runner labels, suite names, and test patterns
verified against the current codebase.
@michaelzhang-ai michaelzhang-ai force-pushed the amd/add-claude-skills branch from 47109f9 to 4836f76 Compare March 8, 2026 08:20
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

amd blackwell deepseek dependencies deterministic diffusion documentation Improvements or additions to documentation hicache lora model-gateway mthreads Multi-modal npu piecewise-cuda-graph quant sgl-kernel speculative-decoding

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.