mlx_lm.generate: repetition_penalty has no default (unlike other sampling params), model can loop into repeated text with no warning

[arjunneema09]

Description
Running mlx_lm.generate with default sampling parameters, a model looped
into repeated text with no warning. Looking at generate.py on main,
other sampling parameters have explicit DEFAULT_* constants defined at
module level (DEFAULT_TEMP, DEFAULT_TOP_P, DEFAULT_MIN_P,
DEFAULT_TOP_K, etc.), but there is no DEFAULT_REPETITION_PENALTY.
SERVER.md also documents defaults for repetition_context_size (20),
presence_penalty (0.0), and frequency_penalty (0.0), but doesn't state
a default for repetition_penalty itself.

To reproduce

1. Run mlx_lm.generate --model <model> --prompt "<prompt>" with no
   --repetition-penalty flag set.
2. With some prompts/models, generation degenerates into repeated text
   or loops.

Expected behavior
Either a sensible default repetition penalty applied out of the box, or
at minimum a clear note in the docs/CLI help that repetition_penalty is
off by default and users should set it explicitly for longer generations.

Question
Is the lack of a default intentional (e.g. to match upstream sampling
behavior exactly)? If so, a one-line doc/help-text note would still help
first-time users avoid this.

Environment
mlx-lm version: [will confirm]
Python version: [will confirm]
OS: macOS [version], chip [M1/M2/M3/...]

Happy to submit a small docs PR to add this note if useful.

[arjunneema09]

Update: looking at the current main branch, mlx_lm.generate's CLI
(setup_arg_parser() in generate.py) doesn't expose --repetition-penalty
at all anymore — it's only available via logits_processors in the Python
API (sample_utils.make_logits_processors). So the gap is more specific
than I first described: CLI users have no way to set repetition penalty,
with no indication in --help that this option doesn't exist for them.