fix(generate): guard per-sequence logits_processors against None in batched generation

[chronoshift]

Problem

In a mixed batch — some sequences carrying logits_processors (e.g. repetition or frequency penalties), others not — the absent entries are None. The per-sequence step loop in GenerationBatch then does:

for processor in self.logits_processors[e]:

which raises TypeError: 'NoneType' object is not iterable and kills the generation thread. The HTTP thread keeps answering GET /health with 200 while every completion hangs indefinitely, so the failure is invisible to a health check and the server has to be restarted.

The practical effect is that --decode-concurrency above 1 is unsafe whenever concurrent requests use different sampling settings, which is the normal case for a shared server.

Reproduction

Start mlx_lm.server --decode-concurrency 2 and send two concurrent chat completions where one sets repetition_penalty (or frequency_penalty) and the other does not, so they land in the same batch.

Fix

One line, mirroring the guard the sibling samplers path already uses twelve lines below in the same method:

sample_sampler = self.samplers[e] or self.fallback_sampler

-                for processor in self.logits_processors[e]:
+                for processor in (self.logits_processors[e] or []):

Relationship to existing issues and PRs

• mlx_lm.server: generation thread dies with TypeError ('NoneType' object is not iterable) when a batch mixes requests with and without logits processors — server then hangs forever #1472 and server: any uncaught exception in _generate leaves HTTP threads serving while every completion hangs forever #1505 report this crash. server: any uncaught exception in _generate leaves HTTP threads serving while every completion hangs forever #1505 generalizes it: any uncaught exception in _generate leaves the server answering 200s while generation is dead.
• fix(generate): avoid None entries in merged logits_processors #1230 proposed fixing the writer ([None] → [[]]) and was closed unmerged.
• generate.py: fix per-slot samplers/logits_processors bookkeeping in filter and extend #1225 proposes fuller per-slot samplers/logits_processors bookkeeping in filter/extend and is still open.

This patch is deliberately the reader-side guard instead. It is one line, it cannot regress a caller that legitimately passes None, it matches the established pattern immediately below it, and it is orthogonal to #1225 — if that lands, this becomes redundant rather than conflicting. It is offered as a minimal stop-gap for a crash that has been reported repeatedly and is still reproducible on main.