Run frontier MoE models on hardware you already own β pure C, zero deps, experts streamed from disk. Tiny engine, immense model. π¦
C
docs: adopt dev integration branch β PRs target dev, batched into stable main (ZacharyZcR's suggestion)
JustVuggf38e74efe7b1
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.mdindex a85b8602..0755056b 100644--- a/CONTRIBUTING.md+++ b/CONTRIBUTING.md@@ -2,6 +2,17 @@Keep changes focused and preserve Colibri's dependency-free default CPU path.+## Branches++- **`main`** is the stable branch. It's what users clone, and it stays known-good+ (engine always passes the token-exact oracle: `SNAP=./glm_tiny TF=1 ./glm 64 16 16`).+- **`dev`** is the integration branch. **Open your PR against `dev`.** Reviewed PRs+ land there first; once a batch is tested and stable, the maintainer fast-forwards+ it into `main`. This keeps `main` clean instead of taking every PR one at a time.++Every PR β on either branch β is reviewed for a clean build (0 warnings), the oracle+(32/32 TF + 20/20 greedy), and its own targeted validation before merge.+## Local checksRun the lightweight checks locally: