Fix TypeError in apply_min_p when min_tokens_to_keep > 1

[mloiterman]

apply_min_p's min_tokens_to_keep > 1 branch passes bare Python False as the values argument of mx.put_along_axis, whose binding requires an mx.array. The branch raises TypeError on every invocation:

>>> from mlx_lm.sample_utils import apply_min_p
>>> import mlx.core as mx
>>> logprobs = mx.log(mx.softmax(mx.random.normal((1, 100)), axis=-1))
>>> apply_min_p(logprobs, 0.05, 3)
TypeError: put_along_axis(): incompatible function arguments. ...
Invoked with types: mlx.core.array, mlx.core.array, bool, kwargs = { axis: int }

Verified the rejection is not mlx-version-dependent (probed mlx 0.31.2 and 0.32.0 — bare bool rejected on both). The scalar was introduced in #1083's refactor (the pre-refactor code passed proper arrays); it has stayed latent because tests/test_sample_utils.py has no min_tokens_to_keep coverage and mlx_lm.server doesn't expose the parameter — only direct library consumers of make_sampler(min_tokens_to_keep=...) can reach the branch. We hit it in a production server built on BatchGenerator the first time a client set the knob.

One-line fix: False → mx.array(False). Adds a regression test that fails with the TypeError pre-fix and pins the keep-floor semantics (exactly min_tokens_to_keep survivors under an aggressive min_p), single-row and batched.

[zcbenz]

Thanks for the PR but at the moment the number of PRs is way beyond our capacity to review so I'm closing the non-essential ones so we can actually work on this repo.

[mloiterman]

Could this one be reopened? It was closed in the capacity sweep rather than on the merits, and it is a hard crash, not a cleanup: any call of apply_min_p(logprobs, min_p, min_tokens_to_keep) with min_tokens_to_keep > 1 raises TypeError from mx.put_along_axis (the values argument is a bare Python False; the binding requires an mx.array). Still reproducible on current main, on mlx 0.32.0 as on 0.31.x — the branch cannot execute at all, so nobody using the keep-floor gets a result.

The fix is one token (False → mx.array(False)) plus the regression test already in the PR; the test fails on main and passes with the change. We have been carrying the same one-line change locally since July with no issues. Happy to rebase if that helps.

[zcbenz]

Sure thanks for looking into this, please excuse us there were just too many pull requests and it was easy to mistakenly close valid ones.

[zcbenz]

Thanks!