BBuf/sglang-qwen-image-optimization
skills/model-optimization/sglang/sglang-qwen-image-optimization/SKILL.md
PR-backed and current-main optimization manual for Qwen-Image and Qwen-Image-Edit in SGLang Diffusion. Use when Codex needs to recover, extend, or audit diffusion transformer loading, layer serving, CUDA graph, TeaCache, IMA, ModelOpt FP8, AMD kernels, Qwen-Image detectors, or cookbook diffusion recipes.
$ install --global
skillsauthnpx skillsauth add BBuf/AI-Infra-Auto-Driven-SKILLS sglang-qwen-image-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 AM26.6s4 files scanned
SKILL.md
- name:
- sglang-qwen-image-optimization
- description:
- PR-backed and current-main optimization manual for Qwen-Image and Qwen-Image-Edit in SGLang Diffusion. Use when Codex needs to recover, extend, or audit diffusion transformer loading, layer serving, CUDA graph, TeaCache, IMA, ModelOpt FP8, AMD kernels, Qwen-Image detectors, or cookbook diffusion recipes.
SGLang Qwen-Image Optimization
Overview
Qwen-Image is the diffusion Qwen lane. It is intentionally split from autoregressive Qwen because the runtime surfaces are SGLang Diffusion pipelines, transformer layers, ModelOpt export/loading, image-edit conditioning, CUDA graph, TeaCache, and AMD diffusion kernels.
Current evidence snapshot:
- SGLang
origin/main:bca3dd958on2026-04-24 - sgl-cookbook
origin/main:816bad5on2026-04-21 - Manual PR diff review date:
2026-04-23 - Docs/snippets:
docs_new/cookbook/diffusion/Qwen-Image/Qwen-Image.mdx,Qwen-Image-Edit.mdx, anddocs_new/src/snippets/diffusion/qwen-image*.jsx - Relevant skill: use the local SGLang diffusion/modelopt skills for implementation-level quantization or profiling work.
PR Dossier Rule
When writing or updating Qwen-Image optimization docs, manually inspect every cited PR diff. Include PR link, state, diff stats, motivation, implementation path, key code excerpt, reviewed files, and validation implications. Use model-pr-diff-dossier for new PR-history production.
Before You Change Anything
Capture:
- model: Qwen-Image or Qwen-Image-Edit
- mode: text-to-image, edit, multi-output, or layer serving
- quantization: BF16, FP8 ModelOpt, or backend-native path
- CUDA graph enabled or disabled
- TeaCache enabled or disabled
- IMA or conditional batching enabled
- backend: NVIDIA or AMD
- whether any external Diffusers pipeline assumptions are being imported
Core Principle
Keep diffusion pipeline changes separate from autoregressive Qwen.
- Qwen-Image issues usually involve denoise-step scheduling, transformer/layer loading, conditional batches, or CUDA graph capture.
- ModelOpt FP8 support must validate image quality and layer-loading compatibility, not just server startup.
- AMD kernels need diffusion-specific microbench and denoise validation.
Main Runtime Surfaces
docs_new/cookbook/diffusion/Qwen-Image/Qwen-Image.mdxdocs_new/cookbook/diffusion/Qwen-Image/Qwen-Image-Edit.mdxdocs_new/src/snippets/diffusion/qwen-image-deployment.jsxdocs_new/src/snippets/diffusion/qwen-image-edit-deployment.jsxpython/sglang/multimodal_gen/configs/pipeline_configs/qwen_image.pypython/sglang/multimodal_gen/configs/models/dits/qwenimage.pypython/sglang/multimodal_gen/runtime/models/dits/qwen_image.pypython/sglang/multimodal_gen/runtime/models/encoders/qwen2_5vl.pypython/sglang/multimodal_gen/runtime/cache/teacache.pypython/sglang/multimodal_gen/runtime/utils/cuda_graph.pypython/sglang/multimodal_gen/tools/build_modelopt_fp8_transformer.py
Optimization Order
- Validate BF16 image generation with fixed seed and prompt.
- Validate edit mode separately from text-to-image.
- Enable CUDA graph and compare image hash/metrics or visual regression outputs.
- Enable TeaCache or IMA after baseline quality is stable.
- Add ModelOpt FP8 and compare BF16-vs-FP8 output.
- Run AMD kernel benchmarks only after e2e denoise is correct.
Open PRs to Track
- #18530: AMD norm/RoPE fusion.
- #19066: Qwen2.5-VL ViT/text encoder optimization for diffusion-adjacent path.
- #19516: CUDA graph.
- #19521: detectors.
- #20429: LN/scale-shift-gate kernel fusion.
- #20432: dual-stream forward.
- #20447: TeaCache.
- #20810: reland CUDA graph.
- #21988: conditional batch multi-output.
- #22362: layer-serving fix.
- #22397: weight-name mapping.
- #22953: merged Qwen-Image RoPE overflow guard.
- #23155: ModelOpt FP8.
Read references/pr-history.md for the diff-reviewed card of each PR. Most entries are currently open; #22953 is merged in current main and can be treated as current behavior.
Validation Lanes
- fixed-seed BF16 generation
- fixed-seed edit generation
- CUDA graph on/off comparison
- TeaCache on/off comparison
- ModelOpt FP8 image quality comparison
- AMD denoise latency and kernel profile
References
references/pr-history.md: manually diff-reviewed PR cards.references/playbook.md: symptom map, toggle matrix, and validation rules.
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