BBuf/vllm-deepseek-v32-optimization
skills/model-optimization/vllm/vllm-deepseek-v32-optimization/SKILL.md
PR-backed optimization manual for DeepSeek V3.2 in vLLM. Use when an engineer needs to audit, debug, extend, or document DeepSeek V3.2 sparse-MLA / DSA runtime, indexer, tool parser, MTP fallback, and long-context decode kernels in vLLM.
$ install --global
skillsauthnpx skillsauth add BBuf/AI-Infra-Auto-Driven-SKILLS vllm-deepseek-v32-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 AM122.8s2 files scanned
SKILL.md
- name:
- vllm-deepseek-v32-optimization
- description:
- PR-backed optimization manual for DeepSeek V3.2 in vLLM. Use when an engineer needs to audit, debug, extend, or document DeepSeek V3.2 sparse-MLA / DSA runtime, indexer, tool parser, MTP fallback, and long-context decode kernels in vLLM.
vLLM DeepSeek V3.2 Optimization
Overview
This skill covers DeepSeek V3.2 sparse-MLA / DSA runtime, indexer, tool parser, MTP fallback, and long-context decode kernels in vLLM.
Evidence snapshot:
- vLLM mainline checked around
0f7be0f2f76814f80f9091220a5fbbb53912ad00 - Support status: supported on current mainline
- Canonical PR notes:
references/pr-history.md - History mirrors:
model-pr-optimization-history/vllm/deepseek-v32/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
vllm/vllm/model_executor/models/deepseek_v2.pyvllm/vllm/model_executor/layers/rotary_embedding/mrope.py
Current Main Summary
- DeepSeek V3.2 is the major vLLM fork of the DeepSeek line because it adds sparse MLA and indexer logic.
- Most regressions land in the indexer builder, tokenizer / parser, and specialized decode kernels rather than the generic V3 loader.
Key Landed PRs
- #25896
Support DeepSeek-V3.2: Landed the initial V3.2 model registration, sparse-attention runtime, and benchmark hooks. - #25999
Support indexer prefill chunking: Made the V3.2 sparse indexer work with chunked prefill instead of eager-only behavior. - #26670
Add AMD GPU support on DeepSeek v3.2 and SparseMLA: Opened the ROCm SparseMLA lane for V3.2 deployments. - #29848
Add DeepSeek-V3.2 tool parser: Added the parser surface that cookbook-style V3.2 reasoning deployments depend on. - #33090
Fix DeepseekV32AssertionError: num_kv_heads == 1``: Removed a hard failure triggered by newer V3.2 attention shapes. - #37421
Persistent TopK scheduler for DeepSeek-V3.2 decode: Modernized the decode scheduler with a CUDAGraph-safe persistent TopK kernel.
Open Optimization Items
- Re-run PR search before claiming new support beyond the checked mainline commit.
Validation Lanes
- SparseMLA prefill plus decode on CUDA and ROCm.
- Chunked prefill with the V3.2 indexer enabled.
- Tool parser on long reasoning responses.
- MTP / eager-fallback checks where sparse kernels are unavailable.
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