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

fixup: the tiny oracle regenerates its fixture — give macOS the pinned deps

JustVugga169bec8d4b7
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index eeabaf05..838a9bbf 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -199,6 +199,20 @@ jobs:
# check.yml, Windows was first exercised at the release tag). The tiny
# oracle is the token-exactness gate: it fails on any numerical or I/O
# divergence, which is exactly what a platform port can break.
+ #
+ # The check target REGENERATES the fixture (--force) before comparing,
+ # so it needs the same pinned torch/transformers the Linux v4-tiny job
+ # installs -- without them the step dies in generation, not in the
+ # oracle (round 2 of this PR measured exactly that).
+ - uses: actions/setup-python@v5
+ if: matrix.os == 'macos-latest'
+ with:
+ python-version: '3.13'
+ cache: pip
+ cache-dependency-path: c/tools/requirements-deepseek-v4-tiny.txt
+ - name: Install pinned tiny-fixture dependencies (macOS)
+ if: matrix.os == 'macos-latest'
+ run: python -m pip install -r c/tools/requirements-deepseek-v4-tiny.txt
- name: DeepSeek V4 tiny oracle (macOS)
if: matrix.os == 'macos-latest'
run: make -C c deepseek-v4-tiny-check