lambda-inference-model-recipes

More advanced recipes for lambda inference-models — https://lambda.ai/inference-models

Reproducible, kubectl-only deployment recipes for serving open-weights models on Lambda systems with NVIDIA Blackwell GPUs. Each recipe is a self-contained set of Kubernetes manifests plus a benchmark harness. Defaults target Lambda Managed Kubernetes on a 1-Click Cluster; each recipe documents the settings to change for another cluster.

Recipes are namespaced by their Hugging Face model id, matching the model's inference page slug:

<org>/<model>/
├── README.md          # deploy + adapt + benchmark
├── manifests/         # kubectl apply, in order
├── diagnostics/       # optional bring-up / fabric tests
└── run-*-bench.sh

Recipes

Model	Hardware	Notes
moonshotai/Kimi-K3	2× NVIDIA HGX B200	2.8T MXFP4 MoE; multi-node vLLM TP=8+PP=2, no NVLink between nodes

Conventions

• kubectl only — no bespoke deployment CLI.
• Lambda-first defaults — see each recipe's adaptation table before using another cluster.
• Pinned images — recipes reference specific (often pre-release) vLLM images; check the recipe README.