Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

fix(test): detect the CUDA recipe by fact, not by error message

JustVugg64ee54e70542
diff --git a/c/tests/test_makefile_cuda_scope.py b/c/tests/test_makefile_cuda_scope.py
index 1e95593a..eb08d2e1 100644
--- a/c/tests/test_makefile_cuda_scope.py
+++ b/c/tests/test_makefile_cuda_scope.py
@@ -32,10 +32,16 @@ def recipe(target, *variables):
def cuda_flag_is_accepted():
- """CUDA=1 is a hard error outside Linux ("supported only on Linux"), so on
- macOS and Windows there is no recipe to inspect at all — the assertions
- below would fail on an absence rather than a regression."""
- return "only on Linux" not in recipe("colibri", "CUDA=1")
+ """Whether this toolchain emits a CUDA recipe at all.
+
+ CUDA=1 is a hard error off Linux and the wording differs per platform
+ ("supported only on Linux" on macOS, "On Windows use: make CUDA_DLL=1" under
+ MSYS2), so matching the message is fragile — the Windows text was missed the
+ first time and the job failed on an absence rather than a regression. Test
+ the FACT instead: if the engine that definitely HAS a CUDA backend does not
+ receive -DCOLI_CUDA, there is no recipe here worth inspecting.
+ """
+ return "-DCOLI_CUDA" in recipe("colibri", "CUDA=1")
@unittest.skipUnless(shutil.which("make"), "make is not installed")