The MLX→GGUF converter in windows/tools/ cannot run from a fresh clone: three of its scripts import a module that is not in the repository.
At af5cd15 (current main):
windows/tools/repack_r3.py:28: from repack_mlx_to_gguf import (StIndex, GgufWriter, KV, T_F32, T_F16, T_Q4_1, ALIGN, QK41, ...
windows/tools/repack_r3.py:31: import repack_mlx_to_gguf as core
windows/tools/repack_r3_8b.py:38: from repack_mlx_to_gguf import (StIndex, GgufWriter, KV, T_F32, T_F16, T_Q4_1, QK41, DT_Q41, ...
windows/tools/lora_mlx_to_gguf.py:26: from repack_mlx_to_gguf import StIndex, GgufWriter, KV, T_F16, bf16_to_f32
repack_r3.py's docstring refers to it as tools/repack_mlx_to_gguf.py. No file of that name exists anywhere in the tree, and git log --all -- '*repack_mlx_to_gguf*' returns nothing, so it was never committed.
Effect. convert_mlx_to_gguf.py runs repack_r3.py / repack_r3_8b.py and then lora_mlx_to_gguf.py as subprocesses, so step 1 fails with ModuleNotFoundError. That blocks the documented path in both READMEs:
windows/README.md §1.4 (on-device conversion)android/README.md §1.4 (python ../windows/tools/convert_mlx_to_gguf.py --dir models/edge0-8b)
I found this by reading the import graph, not by running the converter, because the module the first step needs is absent.
Possible cause. windows/.gitignore has an anchored /models/ now, and the 2026-09-30 commit "windows: ship app/src/models/cards.tsx (gitignore models/ pattern was un-anchored)" fixed one file caught by an ignore rule. This may be a second file lost the same way, or simply not added.
Ask. Could you add windows/tools/repack_mlx_to_gguf.py? There is no release to take it from either.
Thanks for open-sourcing the engines. The patch-band layout with the golden tree hash is a nice piece of work.
The MLX→GGUF converter in
windows/tools/cannot run from a fresh clone: three of its scripts import a module that is not in the repository.At
af5cd15(currentmain):repack_r3.py's docstring refers to it astools/repack_mlx_to_gguf.py. No file of that name exists anywhere in the tree, andgit log --all -- '*repack_mlx_to_gguf*'returns nothing, so it was never committed.Effect.
convert_mlx_to_gguf.pyrunsrepack_r3.py/repack_r3_8b.pyand thenlora_mlx_to_gguf.pyas subprocesses, so step 1 fails withModuleNotFoundError. That blocks the documented path in both READMEs:windows/README.md§1.4 (on-device conversion)android/README.md§1.4 (python ../windows/tools/convert_mlx_to_gguf.py --dir models/edge0-8b)I found this by reading the import graph, not by running the converter, because the module the first step needs is absent.
Possible cause.
windows/.gitignorehas an anchored/models/now, and the 2026-09-30 commit "windows: ship app/src/models/cards.tsx (gitignore models/ pattern was un-anchored)" fixed one file caught by an ignore rule. This may be a second file lost the same way, or simply not added.Ask. Could you add
windows/tools/repack_mlx_to_gguf.py? There is no release to take it from either.Thanks for open-sourcing the engines. The patch-band layout with the golden tree hash is a nice piece of work.