task: import Qwen3-TTS Ascend optimization techniques and evidence #35

Current state: PR #36 merged on 2026-10-19 with full-check and lint-and-smoke passing but no recorded review; its three peer Qwen3-TTS Ascend records remain on main, and the imported performance stays INCONCLUSIVE.
Decision needed: Whether the merged PR #36 import meets the completion criteria so this Issue can be accepted, from the requester (amy-why-3459) or the repository maintainers.
Next action: No assignee; the requester or a maintainer records the acceptance decision.

Goal and acceptance

Import source-grounded Qwen3-TTS Ascend optimization knowledge from the public InferMatrixLab/qwen3-tts-engine project into InferRSI. Completion requires searchable peer-technique records, immutable source attribution, a campaign evidence/navigation note, and passing catalog/CLI/package checks. Imported performance remains INCONCLUSIVE for InferRSI; this task does not authorize a new performance PASS.

Scope and constraints

Preserve the existing inference runtime in its owning repository. Add codec constant/state reuse, graph/batching ownership, and fixed CustomVoice prompt preparation lessons to the existing knowledge catalog. Describe validation requirements and transfer risks without assuming vLLM-Omni obtains the standalone engine's gains. No new GPU/NPU dependency, CI lane, default performance threshold or automatic launcher.

Ownership and authorization

The requesting contributor explicitly asked Codex in this session to submit the Qwen3-TTS optimizations from InferMatrixLab/qwen3-tts-engine to InferMatrixLab/InferRSI. This records that scoped implementation and PR-publication delegation; it does not claim maintainer acceptance, merge permission or goal closure. Acting implementer: OpenAI Codex. Goal acceptance and issue closure remain with the requester/repository maintainers.

Inputs and deliverables

Source snapshot: https://github.com/InferMatrixLab/qwen3-tts-engine/tree/9f190577aa2a147c6840560a56c5996a6c4879a4 . This is the documentation/source attribution snapshot, not the measured engine revision. The original measured revision, baseline revisions, workload and gates remain in that snapshot's evidence/campaign-20261017/plan/ and linked historical reports.

Deliver catalog records, source hashes, explicit import-scope evidence and navigation, plus CPU retrieval and installed-package validation. No accelerator rerun, audio-quality reproduction or independent review has yet been performed. Retain INCONCLUSIVE and leave this issue open for acceptance.