fuse: mlx_lm.fuse counterpart over the public training library
(--dequantize writes dense weights; --upload-repo pushes the result like
convert's). GGUF export is refused. The merge
itself has no cancellation seam: a signal is honored before it starts; one
arriving during the merge lets it finish so the output is never half-written.
fuse: mlx_lm.fuse counterpart over the public training library (--dequantizewrites dense weights;--upload-repopushes the result like convert's). GGUF export is refused. The merge itself has no cancellation seam: a signal is honored before it starts; one arriving during the merge lets it finish so the output is never half-written.