Claude Code plugins

@air-gapped-marketplace · air-gapped/skills

6

vLLM operator reference suite — deployment, configuration, quantization, caching, KV, tool parsers, reasoning parsers, chat templates, benchmarking, performance tuning, observability, omni, input modalities, speculative decoding, NVIDIA hardware, and a Gemma 4 31B operating-point serve recipe.

@air-gapped-marketplace · air-gapped/skills

6

Inference-API compatibility suite — the three open LLM serving protocols as third-party surfaces: OpenAI Chat Completions + legacy /v1/completions (7 local servers, gateways, 10 cloud compat endpoints, reasoning-field schism, client wire behavior), OpenAI Responses (spec, streaming events, store/previous_response_id semantics, translation), and Anthropic Messages (7 native server implementations, thinking-signature seam, stop_reason divergences, Claude Code/opencode integration). Source-examined at pinned commits with live-verification logs.

@air-gapped-marketplace · air-gapped/skills

6

Inference KV-cache and transport suite — LMCache multiprocess (MP) standalone-server mode (DaemonSet + Deployment K8s pattern, ZMQ connector, L1 + L2 NIXL/POSIX/GDS/HF3FS/fs/s3/mooncake adapters) and NVIDIA NIXL transfer library (UCX/GDS/Mooncake/libfabric/HF3FS/S3 plugins, agent API, telemetry) used by Dynamo/vLLM/SGLang. Pairs with the vllm-caching skill in the `vllm` suite and sglang-hicache in the `sglang` suite.

@air-gapped-marketplace · air-gapped/skills

6

Inference host bring-up + tuning suite — Linux/GPU bare-metal host work that sits beneath any inference framework (vLLM, sglang, TensorRT-LLM). Two phases. Bring-up (nvidia-datacenter-bringup): Ubuntu 24.04 LTS from clean OS to gpu-operator cuda-validator passing — B300/B200/H100/A100/L40S/L4 driver + fabricmanager + NVLSM + DOCA-OFED install order, NVIDIA CUDA repo + DOCA repo + air-gap mirror, MOK + DKMS sign-on-build under Secure Boot, Dell PowerEdge XE9780/XE9785 baseboard firmware via iDRAC Redfish DellOemChassis.ExtendedReset, gpu-operator pre-installed-driver-mode integration. Tuning (gpu-host-tuning): read-only snapshot of CPU power state, C-states, NUMA topology, PCIe link state, GPU settings, kernel boot params, sysctl, ulimits, IRQ affinity, container runtime; optional pinned-host↔GPU memcpy bench (torch + numactl); per-lever cheat-sheets to flip settings (governor, EPP, cpuidle, persistence, ECC, hugepages, intel_iommu, NCCL env, tuned-adm profiles, BIOS guidance for Dell X

@air-gapped-marketplace · air-gapped/skills

6

LiteLLM proxy operator suite — multi-pod Redis/Valkey coordination (the silent per-pod-fallback failure mode, the 4-way coordination-Redis resolution order incl. the undocumented general_settings.coordination_redis block since v1.93.0, Sentinel/Cluster support matrix with fix versions, per-pod-vs-shared truth table with real staleness bounds, no-op documented knobs, open-issue catalog) and the management REST API (447-path control plane, ~63% undocumented; the four 'empty means unlimited' footguns, budget semantics traps, presence-sensitive updates that trigger spurious enterprise 403s, config-vs-DB precedence, recon-first protocol via GET /routes).

@air-gapped-marketplace · air-gapped/skills

6

SGLang operator reference suite — HiCache (three-tier hierarchical prefix cache: GPU HBM → host DRAM → distributed L3 via Mooncake/3FS/NIXL/AIBrix/EIC/SiMM/file/LMCache) and the SGLang Model Gateway (`sgl-model-gateway`, formerly `sgl-router`) — Rust router fronting vLLM/SGLang inference workers on Kubernetes with cache-aware/prefix-hash/consistent-hash routing, K8s label-selector service discovery, PD-disaggregation, and `--enable-mesh` CRDT sync between gateway replicas. Pairs with the `vllm` suite for vLLM-side reference and `inference-cache` for engine-agnostic KV-cache transport.

@air-gapped-marketplace · air-gapped/skills

6

NVIDIA AIPerf — vendor-neutral generative-AI inference benchmarking (the genai-perf successor): `aiperf profile` with concurrency / request-rate / fixed-schedule trace-replay / user-centric / multi-run-confidence load, 15 endpoint types and 6 custom dataset formats, goodput SLOs, GPU + Prometheus telemetry, and the reasoning-token TTFT/TTFO split — benchmarks vLLM / SGLang / TensorRT-LLM / Dynamo / NIM / Triton / Ollama.