Skip to content

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

C

tools: deprecate convert_olmoe.py β€” its split gate/up/down layout no longer loads in olmoe.c

Zixin Zhoud616f39e6ce2
diff --git a/.gitignore b/.gitignore
index f62b3792..08c4405d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -68,3 +68,4 @@ olmoe_merged/
olmoe_i4/
c/olmoe_merged/
c/olmoe_i4/
+.idea
diff --git a/c/tools/convert_olmoe.py b/c/tools/convert_olmoe.py
index dc25f6a7..203ac868 100644
--- a/c/tools/convert_olmoe.py
+++ b/c/tools/convert_olmoe.py
@@ -1,6 +1,10 @@
#!/usr/bin/env python3
"""Convert OLMoE HuggingFace checkpoint to colibri int4 format.
+DEPRECATED: produces the split gate_proj/up_proj/down_proj tensor layout. The
+current engine (olmoe.c) only loads the merged 'merged_weight' format β€” use
+convert_olmoe_merged.py instead.
+
Downloads or converts a local OLMoE checkpoint (e.g., allenai/OLMoE-1B-7B-0125-Instruct).
Dense weights stay as-is (engine reads BF16/F16 β†’ F32 on load).
Expert weights get row-wise symmetric quantization to --ebits bits (default 4)
@@ -16,6 +20,12 @@ Usage:
import argparse, json, math, os, struct, sys
from pathlib import Path
+sys.exit(
+ "convert_olmoe.py is deprecated: it produces the split gate_proj/up_proj/down_proj\n"
+ "tensor layout, but the current engine (olmoe.c) only loads the merged 'merged_weight'\n"
+ "format. Use convert_olmoe_merged.py instead."
+)
+
# Windows: force UTF-8 output
if sys.platform == "win32":
for s in (sys.stdout, sys.stderr):