BBuf/sglang-mimo-v2-flash-optimization
skills/model-optimization/sglang/sglang-mimo-v2-flash-optimization/SKILL.md
PR-backed optimization manual for MiMo-V2 / MiMo-V2-Flash / MiMo-V2.5 in SGLang. Use when an engineer needs to audit, debug, extend, or document MiMo-V2 inference-centric MoE runtime, flashinfer/TRT-LLM fused all-reduce, overlap, MTP/EAGLE, multimodal/pro variants, and reasoning parser behavior.
$ install --global
skillsauthnpx skillsauth add BBuf/AI-Infra-Auto-Driven-SKILLS sglang-mimo-v2-flash-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, 2:35 AM159.5s2 files scanned
SKILL.md
- name:
- sglang-mimo-v2-flash-optimization
- description:
- PR-backed optimization manual for MiMo-V2 / MiMo-V2-Flash / MiMo-V2.5 in SGLang. Use when an engineer needs to audit, debug, extend, or document MiMo-V2 inference-centric MoE runtime, flashinfer/TRT-LLM fused all-reduce, overlap, MTP/EAGLE, multimodal/pro variants, and reasoning parser behavior.
SGLang MiMo-V2 / MiMo-V2.5 Optimization
Overview
This skill covers the MiMo-V2 family in SGLang: historical MiMo-V2-Flash,
current MiMoV2ForCausalLM compatibility, MiMo-V2.5/Pro deployment recipes,
MTP/EAGLE paths, flashinfer/TRT-LLM fused all-reduce, overlap, multimodal
variants, and reasoning parser behavior.
Evidence snapshot:
- SGLang
origin/main:6fbad22feon2026-04-28 - sgl-cookbook checked around
e88b0fd8ac5b1caa6eb42766035029220053369b - Latest family update:
#23808renamed the runtime tomimo_v2.py/mimo_v2_nextn.pywhile keeping the old Flash architecture alias loadable;#23851added the MiMo-V2.5 cookbook and command generator. - Canonical PR notes:
references/pr-history.md - History mirrors:
model-pr-optimization-history/sglang/mimo-v2-flash/README.zh.mdandREADME.en.md
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.
Runtime Surfaces
- Current runtime:
sglang/python/sglang/srt/models/mimo_v2.py - Current MTP runtime:
sglang/python/sglang/srt/models/mimo_v2_nextn.py - Historical names still relevant to older PR cards:
sglang/python/sglang/srt/models/mimo_v2_flash.py,sglang/python/sglang/srt/models/mimo_v2_flash_nextn.py - Config and server integration:
sglang/python/sglang/srt/configs/model_config.py,sglang/python/sglang/srt/server_args.py - Current docs/generator:
docs_new/cookbook/autoregressive/Xiaomi/MiMo-V2.5.mdx,docs_new/src/snippets/autoregressive/mimo-v25-deployment.jsx
Current Main Summary
- MiMo-V2-Flash is the historical throughput-oriented landing.
- Current mainline accepts both
MiMoV2ForCausalLMand the oldMiMoV2FlashForCausalLMalias through the renamedmimo_v2.pymodule. - MiMo-V2.5-Pro adds 1.02T text/pro serving with EAGLE speculative decoding;
MiMo-V2.5 covers the 310B multimodal line and TP=4-interleaved fused
qkv_projconstraints. - All-reduce fusion, overlap, hybrid SWA/full attention, MTP/EAGLE, and reasoning parser behavior matter more than generic loader work.
Key Landed PRs
- #15207
MiMo-V2-Flash day0 support: Initial MiMo-V2-Flash landing. - #15464
Optimize MiMo-V2-Flash by flashinfer fused allreduce: Targeted decode-side communication cost. - #15488
Respect--swa-full-tokens-ratio``: Fixed a concrete runtime flag integration bug. - #17634
Support two batch overlap: Added overlap / throughput optimization. - #21414
Add mimo reasoning parser: Completed the parser path for thinking outputs. - #23808
Xiaomi MiMo-V2.5-Pro day0 support: Renamed runtime files tomimo_v2.py/mimo_v2_nextn.py, addedMiMoV2ForCausalLM, and kept the Flash alias. - #23851
Add MiMo-V2.5 docs: Added current MiMo-V2.5/Pro deployment cookbook and command generator.
Validation Lanes
- Startup on MiMo-V2-Flash and current MiMo-V2.5/Pro checkpoints or cookbook routes.
- Re-run the parser, quantization, or multimodal lane that matches this family.
- Re-check MTP/EAGLE,
qkv_projloading, hybrid SWA/full attention, and registered/manual tests after touching loader or processor code.
References
references/pr-history.md: diff-reviewed MiMo-V2 family PR cards; it includes both historical Flash-file cards and currentmimo_v2.py/ MiMo-V2.5 updates.
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