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: torch +cpu wheels are Linux-only — split the pin by platform

JustVugg59e40e26ab01
diff --git a/c/tools/requirements-deepseek-v4-tiny.txt b/c/tools/requirements-deepseek-v4-tiny.txt
index dbbfa9ac..4e4c6686 100644
--- a/c/tools/requirements-deepseek-v4-tiny.txt
+++ b/c/tools/requirements-deepseek-v4-tiny.txt
@@ -1,4 +1,8 @@
--extra-index-url https://download.pytorch.org/whl/cpu
-torch==2.13.0+cpu
+# The +cpu local-version tag exists only in the Linux wheels of the CPU index;
+# macOS wheels are CPU-only by construction and carry the plain version.
+# Same pin either way -- the environment markers pick the spelling per OS.
+torch==2.13.0+cpu; sys_platform == "linux"
+torch==2.13.0; sys_platform == "darwin"
transformers==5.14.1
safetensors==0.8.0