BBuf/sglang-moss-vl-optimization
skills/model-optimization/sglang/sglang-moss-vl-optimization/SKILL.md
PR-backed and current-main optimization manual for Moss-VL in SGLang. Use when an engineer needs to audit or extend Moss-VL multimodal runtime support, Qwen3VL-like vision encoder plumbing, cross-attention custom masks, vision position ids, image/video processor behavior, conversation template registration, flashinfer prefill requirements, or Moss-VL weight loading.
$ install --global
skillsauthnpx skillsauth add BBuf/AI-Infra-Auto-Driven-SKILLS sglang-moss-vl-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: May 2, 2026, 2:37 AM232.5s2 files scanned
SKILL.md
- name:
- sglang-moss-vl-optimization
- description:
- PR-backed and current-main optimization manual for Moss-VL in SGLang. Use when an engineer needs to audit or extend Moss-VL multimodal runtime support, Qwen3VL-like vision encoder plumbing, cross-attention custom masks, vision position ids, image/video processor behavior, conversation template registration, flashinfer prefill requirements, or Moss-VL weight loading.
SGLang Moss-VL Optimization
Overview
Moss-VL landed as a native SGLang runtime path in #23454. It is not a thin alias: the PR added a dedicated model file, processor, multimodal scheduling fields, prompt template, encoder-prefix handling, and a flashinfer prefill requirement for cross-attention custom masks.
Current evidence snapshot:
- SGLang
origin/main:bca3dd958on2026-04-24 - Primary PR: #23454
- Runtime files:
moss_vl.py,multimodal/processors/moss_vl.py,schedule_batch.py,conversation.py,server_args.py
Non-Negotiable Evidence Rule
Use skills/model-optimization/model-pr-diff-dossier/SKILL.md as the production bar.
Every PR cited for this family must be based on diff reading, not only PR titles.
Before You Change Anything
Capture:
- image vs video input count and shape
grid_thw,media_nums_per_sample,visible_frame_counts, andvision_position_ids- whether encoder KV is cached
- prefill backend; Moss-VL requires flashinfer prefill
- conversation template selection (
moss-vl) - MRoPE enabled or disabled
- cross-attention custom mask behavior for extend vs decode
Core Principle
Debug Moss-VL at the multimodal boundary.
- Encoder-prefix placeholder tokens are stripped before text extend.
- Vision tokens include per-frame separator tokens.
- Frame-level visibility comes from processor
cross_attention_maskand becomes a packed flashinfer custom mask. - Heavy vision tensors should be released after encoder KV is produced.
PR Dossier Rule
Before adding Moss-VL evidence, open the PR diff/source and update references/pr-history.md with motivation, key implementation, real code excerpts, reviewed files, and validation implications.
Validation Lanes
- text-only Moss-VL request, to verify cross-attention skip behavior
- single image and multi-image prefill
- video input from URL/path/data URI
- prefix-cache reuse after encoder KV is cached
- extend-stage cross-attention custom mask with frame-level visibility
--prefill-attention-backend flashinferand failure when another backend is forced
References
references/pr-history.md: diff-reviewed Moss-VL 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