# ktransformers — SGLang + kt-kernel serving a MoE model with experts on the
# CPU and attention on the GPU. LIGHT core: the whole stack lives in the
# `sglang` venv (engines-kt.venv-sglang.txt) because kt-kernel and sglang-kt
# pin their own torch, and the main venv only proxies to the SGLang server.
#
# THE WEIGHTS ARE THE COST: the models this exists for are 100 GB and up.
# Put them on a network volume (docs/runpod.md, "Engine pods").
