BBuf/sglang-ltx23-hq-optimization
skills/model-optimization/sglang/sglang-ltx23-hq-optimization/SKILL.md
PR-backed and current-main optimization manual for LTX-2.3 High Quality pipeline in SGLang Diffusion. Use when Codex needs to audit or extend LTX2TwoStageHQPipeline, LTX-2.3 two-stage LoRA switching, HQ sigma/timestep semantics, res2s RK2 refinement, audio/video denoising, Gemma prompt trimming, low-VRAM device snapshots, or LTX-2.3 HQ sampling defaults.
$ install --global
skillsauthnpx skillsauth add BBuf/AI-Infra-Auto-Driven-SKILLS sglang-ltx23-hq-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 AM26.6s2 files scanned
SKILL.md
- name:
- sglang-ltx23-hq-optimization
- description:
- PR-backed and current-main optimization manual for LTX-2.3 High Quality pipeline in SGLang Diffusion. Use when Codex needs to audit or extend LTX2TwoStageHQPipeline, LTX-2.3 two-stage LoRA switching, HQ sigma/timestep semantics, res2s RK2 refinement, audio/video denoising, Gemma prompt trimming, low-VRAM device snapshots, or LTX-2.3 HQ sampling defaults.
SGLang LTX-2.3 HQ Optimization
Overview
LTX-2.3 HQ is a SGLang Diffusion two-stage video+audio pipeline lane. It differs from legacy LTX-2 and non-HQ LTX-2.3 because stage 1 runs at half resolution, HQ sigma shift is resolution-aware, distilled LoRA strength is stage-specific, and stage 2 uses HQ timestep semantics plus res2s refinement behavior.
Current evidence snapshot:
- SGLang
origin/main:bca3dd958on2026-04-24 - Diff-reviewed PRs: #23366 and #23624
- Main runtime files:
ltx_2_pipeline.py,ltx_2_denoising.py,configs/sample/ltx_2.py,configs/pipeline_configs/ltx_2.py
Before You Change Anything
Capture:
- pipeline class:
LTX2Pipeline,LTX2TwoStagePipeline, orLTX2TwoStageHQPipeline - whether
ltx_variant == "ltx_2_3" - stage 1 and stage 2 distilled LoRA strengths
- sigma schedule source and number of latent tokens
- video/audio CFG, STG, rescale, modality scale, and skip-step settings
- low-VRAM snapshot mode vs resident two-transformer mode
- whether batch height/width have already been halved by
LTX2HalveResolutionStage
Core Principle
Do not apply HQ semantics to every LTX-2.3 path.
- HQ uses resolution-aware sigma shift because the official HQ path passes a half-resolution latent to the scheduler.
- Non-HQ one-stage and two-stage LTX-2.3 use the constant anchor schedule.
- HQ stage-2 defaults differ from non-HQ: distilled LoRA strength, STG, rescale, and sampler choices are part of the quality contract.
PR Dossier Rule
Before adding LTX-2.3 HQ evidence, open the PR diff/source and update references/pr-history.md with motivation, implementation, code excerpts, reviewed files, and validation implications.
Validation Lanes
- LTX-2 legacy one-stage unchanged.
- LTX-2.3 one-stage unchanged.
- LTX-2.3 non-HQ two-stage unchanged.
- LTX-2.3 HQ 1088x1920 default with stage-1 halving.
- HQ stage-1/stage-2 distilled LoRA strengths from request extras.
- Audio+video denoise with fixed seed and regression artifacts.
- Low-VRAM and resident two-transformer modes.
References
references/pr-history.md: diff-reviewed LTX-2.3 HQ PR cards.
Related Skills
BBuf/sglang-humanize-review
development
VerifiedTrustedCommunity
Perform SGLang code review in the style of human maintainers by consulting the full non-agent PR review episode corpus from project start through the latest refresh (June 2026), including inline review threads, top-level PR comments, review submissions, original multilingual text, and multi-round discussions. Use when reviewing SGLang PRs, diffs, patches, or local changes for correctness, tests, performance, GPU/runtime risks, API compatibility, and maintainability.
531SKILL.mdUpdated May 21, 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.
531SKILL.mdUpdated May 17, 2026
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