mlx_lm.convert counterpart: main's flags, messages, and check order; the
output defaults to the app's models directory instead of ./mlx_model.
Uniform affine 4/8-bit or the OptiQ mixed path via --target-bpw, or without
-q a dtype cast and/or dequantization (--dtype, -d), through the
same producer as the web quantize job in an owned child. Cancellation
terminates and joins that child; the atomic writer never publishes a partial
output and the parent removes the child's staging.
mlx_lm.convert counterpart: main's flags, messages, and check order; the output defaults to the app's models directory instead of
./mlx_model. Uniform affine 4/8-bit or the OptiQ mixed path via --target-bpw, or without -q a dtype cast and/or dequantization (--dtype, -d), through the same producer as the web quantize job in an owned child. Cancellation terminates and joins that child; the atomic writer never publishes a partial output and the parent removes the child's staging.