feat(lm): sglang LM block — preflight the hardware, repair the runtime, serve from a named venv - #199

Merged

krakennetworks merged 11 commits into

mainfrom

feat/sglang-lm-block

Aug 18, 2026

Commits on Aug 18, 2026

feat(lm): declare an SGLang endpoint per graph via a top-level `lm:` block
krakennetworks
committed
334b52a
examples: sglang-qa.yaml — a graph that carries its own LM endpoint
krakennetworks
committed
4f7cc01
feat(lm): preflight hardware, sglang runtime and weights before spawning
krakennetworks
committed
aa952b1
fix(lm): repair a wrong-backend torch instead of reinstalling sglang over it
krakennetworks
committed
a9b30d0
feat(cli): --sglang-python serves from the venv that has sglang
krakennetworks
committed
4433a58
fix(lm): give the spawned server the PATH a venv activation would have
krakennetworks
committed
3b3eeda
fix(lm): keep the venv's bin on PATH by not resolving the interpreter symlink
krakennetworks
committed
928aa58
test(examples): prove the stub was called, and keep DSPy's disk cache out of it
krakennetworks
committed
2804d48
fix(cli): count LM calls off the client, and name cache hits
krakennetworks
committed
f1b7f93
chore: regenerate uv.lock for the release version bump
krakennetworks
committed
8d92770
docs(changelog): entry for the IR lm block (NFR-2 gate)
krakennetworks
committed
1d105c4