Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Merge pull request #1063 from terrizoaguimor/refactor/family-registry
Vincenzo Fornaro12368b7aa19f
diff --cc c/resource_plan.pyindex ae8a62f3,ad70504e..8e7a324a--- a/c/resource_plan.py+++ b/c/resource_plan.py@@@ -10,9 -10,11 +10,11 @@@ import subprocesimport sysfrom pathlib import Path+ from family_registry import expert_contributions, planner_geometry, resolve_model+GB = 1_000_000_000-EXPERT_RE = re.compile(r"model\.layers\.(\d+)\.mlp\.experts\.(\d+)\.")+EXPERT_RE = re.compile(r"(?:model\.)?layers\.(\d+)\.(?:mlp|ffn)\.experts\.(\d+)\.")def _tensor_sizes(path):