[BugFix] qwen3_tts chunk boundary handling logic in initial chunk (IC)

[Fattysand]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

Fix the initial chunk (IC) coverage logic in qwen3_tts.py to align with the correct behavior already implemented in fish_speech.py.

Currently, qwen3_tts.py uses < and chunk_size - 1 which constrains IC coverage to strictly less than chunk_size, while fish_speech.py uses <= and no -1, allowing IC to cover up to chunk_size. This mismatch causes qwen3_tts.py to miss the last IC chunk (e.g. cs=25, ic=5: IC emits at 5, 10, 15, 20 then jumps to normal phase emitting 21–45, skipping a 1–25 emit).

Proposed fix (only two lines changed):

# line 215: < → <=
in_initial_phase = initial_chunk_size > 0 and initial_chunk_size < chunk_size and length <= chunk_size

# lines 227-229: remove -1
initial_coverage = (
    (chunk_size // initial_chunk_size) * initial_chunk_size if 0 < initial_chunk_size < chunk_size else 0
)

Reference — fish_speech.py (lines 118 & 131):

in_initial_phase = initial_chunk_size > 0 and length <= chunk_size
initial_coverage = (chunk_size // initial_chunk_size) * initial_chunk_size if initial_chunk_size > 0 else 0

Test Plan

This is a minimal two-line logic fix aligning qwen3_tts.py with the existing fish_speech.py implementation. No additional test scripts are needed — the change is self-contained and the edge cases have been manually verified (see below).

Test Result

Edge cases verified:

• Non-divisible (cs=25, ic=8): (24//8)*8 == (25//8)*8 == 24, behavior unchanged.
• ic == chunk_size: Guarded by initial_chunk_size < chunk_size, IC skipped entirely — unaffected.
• finished=True during IC: Handled by existing context_length logic.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan. Please provide the test scripts & test commands. Please state the reasons if your codes don't require additional test scripts. For test file guidelines, please check the test style doc
• The test results. Please paste the results comparison before and after, or the e2e results.
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model. Please run mkdocs serve to sync the documentation editions to ./docs.
• (Optional) Release notes update. If your change is user-facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)

cc @Sy0307

[Fattysand]

Update test_qwen3_tts_async_chunk.py to match corrected IC boundary logic

The existing test cases for the "IC evenly divides chunk_size" edge case (ic=8, cs=16) were written under the assumption that IC phase uses strict < (i.e., length < chunk_size). This contradicts the <= boundary used in both fish_speech.py and the corrected qwen3_tts.py.

What changed in tests:

1. Fixed 2 incorrect expectations for ic=8, cs=16:

case	before	after	reason
n=16, finished=False	None	(8, 16)	16<=16 → still IC phase, 16%8==0 → emit
n=24, finished=False	(8, 24)	None	normal phase, adjusted=8, 8%16!=0 → hold

2. Added 1 normal-emit verification for ic=8, cs=16:

   • n=32 → (16, 32): first normal emit at initial_coverage + chunk_size = 16+16 = 32

3. Added 5 new cases for ic=5, cs=25 (IC evenly divides chunk_size with higher multiplicity):

   • Demonstrates IC filling the entire first chunk: emit at 5, 12→hold, 25→emit, 30→hold, 50→first normal emit
   • Emit interval pattern: 5,5,5,5,5,25,25,... — smooth transition with no gap
   • This is the key scenario that exposes the bug in the old < logic: with strict <, IC would only emit at 5,10,15,20 (skipping 25), then normal phase wouldn't emit until frame 45, creating a 25-frame gap (longer than the normal chunk itself)

Updated comments clarify the IC boundary rule:

# IC phase: length <= chunk_size  (uses <=, consistent with fish_speech)
# IC emits fill the entire first chunk_size worth of frames, so the
# normal phase always starts at a clean chunk boundary.
# initial_coverage = (chunk_size // initial_chunk_size) * initial_chunk_size

---

[Sy0307]

LGTM. Nice catch.

[linyueqian]

@JuanPZuluaga please also take a look. thank you!

[linyueqian]

LGTM