BBuf/sglang-z-image-turbo-optimization
skills/model-optimization/sglang/sglang-z-image-turbo-optimization/SKILL.md
PR-backed optimization manual for Z-Image and Z-Image-Turbo in SGLang Diffusion. Use when Codex needs to audit or extend Z-Image registry entries, Turbo/base sampling defaults, CFG normalization, sequence-parallel latent sharding, Cache-DiT/TeaCache behavior, LoRA/FP8 coverage, or AMD nightly validation.
$ install --global
skillsauthnpx skillsauth add BBuf/AI-Infra-Auto-Driven-SKILLS sglang-z-image-turbo-optimizationInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: Apr 25, 2026, 2:40 AM24.0s2 files scanned
SKILL.md
- name:
- sglang-z-image-turbo-optimization
- description:
- PR-backed optimization manual for Z-Image and Z-Image-Turbo in SGLang Diffusion. Use when Codex needs to audit or extend Z-Image registry entries, Turbo/base sampling defaults, CFG normalization, sequence-parallel latent sharding, Cache-DiT/TeaCache behavior, LoRA/FP8 coverage, or AMD nightly validation.
SGLang Z-Image-Turbo Optimization
Overview
Z-Image and Z-Image-Turbo are native SGLang Diffusion model families, not generic Diffusers-only fallbacks. Current main registers both Tongyi-MAI/Z-Image-Turbo and Tongyi-MAI/Z-Image, uses ZImagePipelineConfig, and serves through ZImagePipeline or ComfyUIZImagePipeline.
Current evidence snapshot:
- SGLang
origin/main:bca3dd958on2026-04-24 - Core runtime:
python/sglang/multimodal_gen/configs/pipeline_configs/zimage.py - DiT runtime:
python/sglang/multimodal_gen/runtime/models/dits/zimage.py - Registry:
python/sglang/multimodal_gen/registry.py - Cookbook:
docs_new/cookbook/diffusion/Z-Image/Z-Image-Turbo.mdx - Diff-reviewed PRs: #17822, #19733, #23455
Before You Change Anything
Capture:
- model id:
Tongyi-MAI/Z-Image-TurboorTongyi-MAI/Z-Image - sampling class:
ZImageTurboSamplingParamsorZImageSamplingParams cfg_normalization,guidance_scale, andguidance_rescale- sequence parallelism:
--sp-degree,--ulysses-degree,--ring-degree - Cache-DiT/TeaCache env settings
- LoRA/FP8 transformer path, if used
- AMD CI coverage status for
nightly-amd-1-gpu-zimage-turbo
Core Principles
- Keep Turbo and base Z-Image sampling defaults separate. Turbo defaults to few-step distilled behavior, while base Z-Image keeps CFG enabled.
- Do not route
z-imagedetectors through the Turbo sampling params unless the model id containsz-image-turbo. - Preserve Z-Image's native H/W sequence-parallel sharding plan; it shards latent spatial axes before patchification-sensitive gather.
cfg_normalizationis a quality-control knob, not a generic sampling parameter to silently drop.- AMD CI coverage is part of the support surface because #23455 restored a test that had been removed for missing
__main__.
Validation Lanes
- Registry resolves
Tongyi-MAI/Z-Image-TurbotoZImageTurboSamplingParamsandZImagePipelineConfig. - Registry resolves
Tongyi-MAI/Z-ImagetoZImageSamplingParamsand excludes Turbo. test/registered/amd/test_zimage_turbo.pyruns viapython test/registered/amd/test_zimage_turbo.py -v.- Diffusion server image generation returns bytes and passes the CLIP threshold when CLIP dependencies are available.
- Multi-GPU SP tests preserve non-square latent shapes and gather/crop correctly.
PR Dossier Rule
Before adding Z-Image evidence, open the PR diff/source and update references/pr-history.md with motivation, key implementation, short code excerpts, reviewed files, and validation implications.
References
references/pr-history.md: diff-reviewed Z-Image PR cards.
Related Skills
BBuf/vllm-sota-humanize-loop
development
VerifiedTrustedCommunity
Run an autonomous Humanize-governed vLLM SOTA performance loop for one LLM model: first perform the fixed fair vLLM/SGLang/TensorRT-LLM deployment search and benchmark, then start one RLCR loop that repeatedly decides the gap, profiles the current bottleneck, runs layer/kernel pipeline analysis, patches vLLM code, optionally uses ncu-report-skill for kernel evidence, and revalidates until vLLM matches or beats the best observed framework under the same workload and SLA.
423SKILL.mdUpdated May 27, 2026
BBuf/llm-pipeline-analysis
devops
VerifiedTrustedCommunity
Inspect LLM torch profiler traces at forward-pass, layer, and kernel level. Use when you need layer timings, anchor-kernel boundaries, representative kernel flows, or Perfetto time ranges.
423SKILL.mdUpdated May 21, 2026
BBuf/sglang-sota-humanize-loop
development
VerifiedTrustedCommunity
Run an autonomous Humanize-governed SGLang SOTA performance loop for one LLM model: first perform the fixed fair SGLang/vLLM/TensorRT-LLM deployment search and benchmark, then start one RLCR loop that repeatedly decides the gap, profiles the current bottleneck, runs layer/kernel pipeline analysis, patches SGLang code, optionally uses ncu-report-skill for kernel evidence, and revalidates until SGLang matches or beats the best observed framework under the same workload and SLA.
423SKILL.mdUpdated May 17, 2026
BBuf/model-pr-history-knowledge
documentation
VerifiedTrustedCommunity
Use when an SGLang, vLLM, or TensorRT-LLM serving/model optimization task needs prior model-family PR evidence. Query and read the PR-driven history docs under model-pr-optimization-history before choosing source paths, fast paths, kernel/fusion ideas, regression risks, or validation lanes.
423SKILL.mdUpdated May 17, 2026