Resources

Shared scripts, recipes, mod patches, and infra docs used by the model posts.

Docs

• INFRA.md — cluster hardware, network, image distribution, common gotchas
• WHEEL-PROVENANCE.md — vLLM + flashinfer git SHAs, build steps

Bench harness

• bench.py — concurrency sweep + NIAH, streams via requests + SSE

Launch scripts

Operational relaunch scripts are not published because they contain site-local hostnames, fabric addresses, and filesystem paths. The posts include sanitized launch commands and the recipes below keep the reusable model-serving flags.

Recipes

• recipes/nemotron-3-ultra-nvfp4.yaml

Mod patches

Mod patch directories are kept in the private repo copy. The public posts describe what each mod changes without publishing operator-local wrapper scripts.

Wheel provenance

Wheels are NOT served from an upstream URL — they are built locally inside the Dockerfile from git SHAs ace95c9cf (vLLM) and 1d54c5c6 (flashinfer). See WHEEL-PROVENANCE.md for the full build recipe.