Long speech, bounded memory

The question

Long input needed a bounded rendering path. The live service split an eight-sentence request into chunks and checked the resulting audio, while documenting the separate CLI run that remained deferred.

From the original notebook

Sentence-aware auto-chunking for the live MOSS-TTS v1.5 HQ4-AWQ stack (daemon moss54d on :8894 + golden CLI infer_hybrid_3060.py). Long inputs split on sentence boundaries (dots, ?, !, …, newlines; abbreviations like Sr./EE.UU. protected); each chunk renders separately with bounded VRAM and joins with 150 ms of silence.

Snapshot of the exact files deployed to WORKSTATION/moss54-staging on host research workstation (RTX 3060 12 GB). Source of truth stays in staging; this dir is the reproducible-experiment copy (NOT mirrored to vladgateway).

Files

• chunking.py — pure-python splitter: split_sentences, chunk_text (default 300 chars/chunk, hard-split fallback for dotless text).
• moss54d.py — warm daemon; speech() dispatches to _render_single per chunk. New request fields: auto_chunk (default true), chunk_chars, chunk_gap_ms. Response adds "chunks".
• infer_hybrid_3060.py — golden CLI with the same chunk loop inside run_single (model stays loaded across chunks). Flags: --no-auto-chunk, --chunk-chars, --chunk-gap-ms.

Semantics

• 1 chunk → byte-identical to the old single-pass path (same seed).
• N chunks → chunk i renders with seed+i (deterministic on both paths).
• Per-chunk audio-health gate (error names the failing chunk).
• Reference voice (--reference / voice) reused for every chunk.

Verification (live daemon, 2026-09-21)

Case	chunks	Audio	Tokens	Total	Health
short (Hola, esta es una prueba corta…)	1	4.8 s	94	12.9 s	rms 0.095, peak 0.86, not silent
long (551 chars, 8 sentences)	2 ([282, 269] chars)	48.0 s	666	92.5 s	rms 0.078, peak 0.84, not silent

Daemon log: [moss54d] auto-chunk: 2 chunks ([282, 269] chars). GPU steady at ~9.6 GiB (vLLM 2.5 + daemon 6.4 + headroom) — no VRAM explosion. samples/ holds both wavs rendered with seed=1234, language=Spanish.

Existing suite tests/test_runtime_guards.py: 5/5 pass after the change.

Note: the golden CLI chunk loop is code-complete and shares the unit-tested splitter, but a live CLI long-render was not run (it would exceed 12 GB alongside the warm daemon — needs a maintenance window).

Keep exploring

Original experiment record. Workstation paths have been generalized. Detailed measurements below retain their original workload and validation boundaries.