BBuf/sglang-qwen36-optimization
skills/model-optimization/sglang/sglang-qwen36-optimization/SKILL.md
PR-backed and current-main optimization manual for Qwen3.6 in SGLang. Use when an engineer needs to recover, extend, or audit Qwen3.6-35B-A3B/27B dense deployment, hybrid Gated Delta Network behavior, multimodal inputs, thinking preservation, Qwen3 reasoning plus Qwen3-Coder tool parser, MTP, Mamba scheduler strategy, FP8/BF16 commands, CPU offload, or cookbook parity.
$ install --global
skillsauthnpx skillsauth add BBuf/AI-Infra-Auto-Driven-SKILLS sglang-qwen36-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, 2:34 PM149.9s1 file scanned
SKILL.md
- name:
- sglang-qwen36-optimization
- description:
- PR-backed and current-main optimization manual for Qwen3.6 in SGLang. Use when an engineer needs to recover, extend, or audit Qwen3.6-35B-A3B/27B dense deployment, hybrid Gated Delta Network behavior, multimodal inputs, thinking preservation, Qwen3 reasoning plus Qwen3-Coder tool parser, MTP, Mamba scheduler strategy, FP8/BF16 commands, CPU offload, or cookbook parity.
SGLang Qwen3.6 Optimization
Overview
Qwen3.6 is split out because the current SGLang support is mostly a deployment/cookbook layer over hybrid Qwen infrastructure, with unique user-facing behavior: thinking preservation, multimodal input, Qwen3 reasoning parser, Qwen3-Coder tool parser, MTP, and Mamba scheduler strategy.
Current evidence snapshot:
- SGLang
origin/main:b3e6cf60aon2026-04-22 - sgl-cookbook
origin/main:816bad5on2026-04-21 - Docs/snippet files:
docs_new/cookbook/autoregressive/Qwen/Qwen3.6.mdx,docs_new/src/snippets/autoregressive/qwen36-deployment.jsx - Adjacent runtime files:
qwen3_next.py,qwen3_5.py, Qwen VLM processors, and Qwen3-Coder parser depending on the checkpoint path
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:
- checkpoint: Qwen3.6-35B-A3B BF16/FP8 or Qwen3.6-27B dense
- whether multimodal input is used
- reasoning parser: usually
qwen3 - tool parser: usually
qwen3_coder - MTP enabled or disabled
--mamba-scheduler-strategy: defaultno_bufferorextra_buffer- whether
SGLANG_ENABLE_SPEC_V2=1is set - attention backend on B200: current docs add
--attention-backend trtllm_mha
Core Principle
Qwen3.6 work should preserve deployment semantics while reusing lower-level Qwen3-Next/Qwen3.5 machinery.
- Do not add Qwen3.6-specific runtime forks until the actual architecture mismatch is proven.
- Keep docs and command generators accurate; this is currently the primary source of Qwen3.6 behavior in the repo.
- Treat CPU offload and hybrid linear-attention bugs as shared hybrid-model issues.
Main Runtime Surfaces
docs_new/cookbook/autoregressive/Qwen/Qwen3.6.mdxdocs_new/src/snippets/autoregressive/qwen36-deployment.jsxdocs_new/docs.jsonpython/sglang/srt/models/qwen3_next.pypython/sglang/srt/models/qwen3_5.pypython/sglang/srt/function_call/qwen3_coder_detector.pypython/sglang/srt/multimodal/processors/qwen_vl.py
Optimization Order
- First verify that the desired behavior is absent from Qwen3-Next/Qwen3.5 shared code.
- Keep Qwen3.6 docs and command generator in sync with any runtime flag change.
- Validate reasoning and tool calling together because docs recommend both.
- Validate multimodal requests separately from text-only MTP.
- Only split runtime code if the checkpoint config requires a new class or loader rule.
PR Dossier Rule
Before adding Qwen3.6 PR evidence, open the PR diff/source and write a full card in references/pr-history.md with motivation, key implementation, code excerpt, reviewed files, and validation implications.
Do not add title-only open PR lists. If an open PR matters for Qwen3.6, keep it in a clearly marked open optimization card only after reviewing the diff, as done for #23474.
Validation Lanes
- Text-only BF16/FP8 server command from
qwen36-deployment.jsx. - Multimodal image and video request examples from
Qwen3.6.mdx. - Streaming reasoning plus tool-call parser together.
- MTP with
extra_bufferand no-MTP with default scheduler.
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