Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Point the efficiency tests at the engine's actual name
ZacharyZcR38fd944e990a
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.mdindex a37fd176..6ce28cb7 100644--- a/CONTRIBUTING.md+++ b/CONTRIBUTING.md@@ -5,7 +5,9 @@ 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`).+ (engine always passes the token-exact oracle: `SNAP=./glm_tiny TF=1 ./colibri 64 16 16`,+ run from `c/`). The `glm_tiny` fixture is generated, not committed --+ `python3 c/tools/make_glm_oracle.py` builds it (needs torch).- **`dev`** is the integration branch. **Open your PR against `dev`.** Reviewed PRsland there first; once a batch is tested and stable, the maintainer fast-forwardsit into `main`. This keeps `main` clean instead of taking every PR one at a time.diff --git a/c/tests/test_inefficiency.py b/c/tests/test_inefficiency.pyindex 00d8b58b..70fadb72 100644--- a/c/tests/test_inefficiency.py+++ b/c/tests/test_inefficiency.py@@ -2,7 +2,7 @@These run against the bundled glm_tiny model (~0.6 MB resident, ~0.1s/run) andgate CI: a regression here means something broke. They run on a plain CPU-only-`glm.exe` build; the CUDA_* tests auto-skip if the engine wasn't built with+`colibri` build; the CUDA_* tests auto-skip if the engine wasn't built withCUDA_DLL=1 (see tests/README_efficiency.md for the build command).The signals under test, and the inefficiency each catches:@@ -26,18 +26,24 @@ from tools.efficiency import (HERE = Path(__file__).resolve().parentC_DIR = HERE.parent-ENGINE = C_DIR / "glm.exe"+# The engine was renamed glm -> colibri, and this file kept pointing at+# "glm.exe" -- a name that does not exist on Linux or macOS (no .exe) and no+# longer exists on Windows either. _engine_present() was therefore always+# False, so every test in this file has been skipping on every platform+# instead of running. Resolve the real name, both extensions.+ENGINE = next((p for p in (C_DIR / "colibri", C_DIR / "colibri.exe") if p.exists()),+ C_DIR / "colibri")TINY = C_DIR / "glm_tiny"def _engine_present() -> bool:"""True iff BOTH the built engine AND the tiny fixture are available.- These tests need glm.exe (a build artifact) AND glm_tiny/ (a generated+ These tests need the colibri engine (a build artifact) AND glm_tiny/ (a generatedfixture, gitignored). CI runs `make check` = "dependency-free tests, nomodel downloads" (workflow .github/workflows/check.yml, by design #140), soneither is present there and these tests must SKIP rather than fail. They- run locally after `make glm.exe` (glm_tiny ships alongside the source, or+ run locally after `make colibri` (glm_tiny ships alongside the source, oris regenerated by tools/make_glm_oracle.py)."""return ENGINE.exists() and (TINY / "config.json").exists()@@ -46,7 +52,7 @@ def _engine_present() -> bool:def _skip_reason() -> str:"""Name exactly which prerequisite is missing, so the skip is actionable."""if not ENGINE.exists():- return "glm.exe not built (run: make glm.exe)"+ return "colibri not built (run: make -C c colibri)"if not (TINY / "config.json").exists():return "glm_tiny fixture absent (gitignored; ship it locally or run tools/make_glm_oracle.py)"return ""@@ -84,7 +90,7 @@ def _cuda_available() -> bool:return False[email protected](_engine_present(), _skip_reason() or "glm.exe + glm_tiny required")[email protected](_engine_present(), _skip_reason() or "colibri + glm_tiny required")class TinyEfficiencyTest(unittest.TestCase):"""Asserted regression tests on the resident tiny model. Gates CI."""@@ -174,7 +180,7 @@ class TinyEfficiencyTest(unittest.TestCase):@unittest.skipUnless(_cuda_available(),- _skip_reason() or "CUDA build not present (run: make clean && make glm.exe CUDA_DLL=1 && make cuda-dll)")+ _skip_reason() or "CUDA build not present (run: make clean && make colibri CUDA_DLL=1 && make cuda-dll)")class TinyCudaEfficiencyTest(unittest.TestCase):"""CUDA-path regression tests on the tiny model. Skip unless CUDA built.diff --git a/c/tools/efficiency.py b/c/tools/efficiency.pyindex 775e8cf0..1e21e51a 100644--- a/c/tools/efficiency.py+++ b/c/tools/efficiency.py@@ -389,12 +389,15 @@ def run_engine() -> tuple[dict, subprocess.CompletedProcess]:"""Run the engine with an env overlay; return (parsed_telemetry, proc).- `engine` defaults to ./glm.exe (colibri's Windows host). `snap` defaults to+ `engine` defaults to the built ./colibri (or .exe). `snap` defaults tothe bundled tiny model (glm_tiny) so callers can omit it for fast tests.The positional argv is `cap ebits dbits`, matching the engine's main()."""if engine is None:- engine = str(Path(__file__).resolve().parent.parent / "glm.exe")+ # glm was renamed colibri; prefer whichever of the two names exists so+ # a stale checkout still resolves, and fall back to the current one.+ _c = Path(__file__).resolve().parent.parent+ engine = str(next((p for p in (_c / "colibri", _c / "colibri.exe") if p.exists()), _c / "colibri"))env = os.environ.copy()# Strip CUDA vars by default so a CPU run isn't accidentally GPU-accelerated# by a leftover env; callers opt in by passing them in env_overlay.