[Bugfix][MiniCPM-o] Cap offline Talker generation at remaining context

[amy-why-3459]

Summary

After #6346 made the Talker a real single-vocab codec LM, three leftover mismatches with upstream MiniCPMTTS.generate() still produce silent or stuttering tails on long answers. This PR scores the codec repetition penalty the way upstream does, restores forced EOS after vLLM's min_tokens processor blanks it, and caps offline generation at upstream's max_new_token.

Motivation

Fix: #6428
#6346 wired Talker logits into vLLM's Sampler. That was necessary, but the Sampler is not a drop-in for MiniCPMTTS.generate() / generate_chunk():

1. Whole-stream presence penalty vs 16-frame frequency penalty. vLLM taxes a code once it has ever been sampled, for the rest of the stream. A codec request is thousands of frames long, so every seen code ends up scaled by the same flat factor while unseen codes stay untouched; the tail then drifts off the speech manifold into near-silence. Upstream's CustomRepetitionPenaltyLogitsProcessorRepeat(penalty, num_code, 16) (built by gen_logits()) taxes penalty ** freq over the last 16 frames only.

2. Forced EOS vs MinTokensLogitsProcessor. Codec EOS is a stage stop_token_ids entry. When the model forces EOS (empty speech, or the recorded budget), that processor then masks EOS for the first min_tokens steps, the row becomes all -inf, and the request keeps decoding until the length cap.

3. Offline length. Without MiniCPMTTS.generate's max_new_token=2048, a request that never samples codec EOS keeps emitting for the remaining Talker context — often twice as long as upstream — which is audible as a long silent tail.

Changes

Talker (minicpmo_4_5_omni_tts.py)

• Keep a per-request recent_codes window of 16 (upstream past_window) and score _apply_batched_repetition_penalty in sample() before the Sampler, so the penalty lands ahead of top-k / top-p.
• Accept a per-row repetition_penalty (mirrors upstream's per-request sampling_params.repetition_penalty) and neutralize the Sampler's own pass (repetition_penalties = 1) so the penalty is scored once.
• Stash forced-EOS rows across MinTokensLogitsProcessor and overwrite the sampled ids afterwards, so a finished request actually emits codec EOS and releases.
• Offline budget is min(2048, remaining Talker context). Native duplex is unchanged: one generate_chunk is still 25 frames + the terminating sample.

Deploy YAML

Stage-1 default_sampling_params now track utils.TTSSamplingParams:

• temperature: 0.8 (unchanged)
• top_p: 0.8 → 0.85
• top_k: 100 → 25
• repetition_penalty: 1.02 → 1.05
• min_tokens: 50 (upstream min_new_token)
• max_tokens: 4096 (Sampler ceiling; Talker still clamps to 2048 / remaining context)

repetition_penalty: 1.05 is required: the field's semantics changed from "whole-stream presence" to "16-frame frequency", so the old 1.02 (chosen to keep the old penalty from being too harsh) is no longer a meaningful number.

top_k / top_p are not required for the code change to be correct. They were also #6346 placeholders, not a tuned baseline, and they are aligned here so the default request matches upstream. They do change acoustics and can move seed-tts / accuracy numbers; minicpmo_4_5_duplex.yaml does not need a matching edit — it only overlays min_tokens: 0 / max_tokens: 4096 and inherits the rest.

min_p / win_size / tau_r stay out: they are declared on TTSSamplingParams but never read by gen_logits().

Warper order is still not identical (upstream is top-p then top-k with min_tokens_to_keep=3; vLLM is top-k then top-p). On 0.85 / 25 that difference is small; this PR only aligns the numbers.

Tests

• Windowed penalty matches the request-local reference, forgets codes older than 16, honours no_penalties, and is consumed once so the next step cannot rescore a stale row.
• Forced EOS that min_tokens blanked is restored on the sampled ids; unforced rows are left to the Sampler.
• Offline prefill records max_tokens = 2048 when context is not the binding constraint, and still clamps to remaining context when it is.
• Config factory asserts top_k=25, top_p=0.85, repetition_penalty=1.05 on every MiniCPM-o 4.5 deploy YAML (duplex overlay is allowed to drop min_tokens to 0).

Tested

• tests/model_executor/models/minicpmo_4_5/test_talker_batching.py
• tests/config/test_config_factory.py
• offline /v1/chat/completions long-form TTS (listen for silent / stuttering tail)
• native duplex still chunks at 26 and does not hold the Talker request open

<frozen importlib._bootstrap>:488
  <frozen importlib._bootstrap>:488: DeprecationWarning: builtin type SwigPyObject has no __module__ attribute

../.venv/lib/python3.12/site-packages/torch/jit/_script.py:365: 14 warnings
  /data/why/.venv/lib/python3.12/site-packages/torch/jit/_script.py:365: DeprecationWarning: `torch.jit.script_method` is deprecated. Please switch to `torch.compile` or `torch.export`.
    warnings.warn(

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
--- Running Summary
================= 10 passed, 16 warnings in 4456.09s (1:14:16) =================
sys:1: DeprecationWarning: builtin type swigvarlink has no __module__ attribute

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

This PR appears to be related to model: minicpm.

Model owners: @y-null

@amy-why-3459, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[amy-why-3459]

@Gaohan123 @natureofnature @y-null PTAL

[yenuo26]

Does this file need to be added to the source_file_dependencies of ready & merge?

[yenuo26]

LGTM