Repository navigation
feat(tts): auto-select NVIDIA acceleration with manual overrides - #177
Merged
Merged
Conversation
Lucas1479
added a commit
that referenced
this pull request
Oct 10, 2026
#178) ## Problem and change A completed research report could appear in full on its Slice card while Kurisu only received a short narration. WorkNote/narrator input budgets discarded the report body, and subsequent Main Chat frames projected status without the persisted result. Questions about details late in a report therefore could not be answered from the already completed Work. - Preserve complete terminal report input through WorkNote, merged narration and the existing terminal-language repair. Keep progress and spoken-output budgets. - Read the latest terminal Attempt result from the Work accepted at the receiving session/context into `current_work.result`. Retain provider attribution and Host completion state; do not expose predecessor results during a newer active or orphaned Attempt. Reuse the existing ledger after restart. - Add the requested flexible six-sentence rule to the built-in Japanese Kurisu persona: normally at most six sentences, with exceptions for requested detail or an accurate, sufficient answer. The optional full-persona override stays empty by default and still takes precedence. - Document the behavior in both READMEs. No database migration, new setting, new summarization pass, or dependency change. ## Validation - Relevant regression suites on the combined #177/#178 head `0f51b41`: **223 passed**. Command: `python -X utf8 -m pytest -q tests/test_work_terminal_report_context.py tests/test_main_chat_character_prompt.py tests/test_character_prompt_snapshots.py tests/test_character_prompt_module.py tests/test_cooperative_work_provider_facts.py tests/test_cooperative_experience.py tests/test_cooperative_shared_history.py tests/test_session_character_scope.py tests/test_work_narration_priority.py tests/test_work_narration_governor.py tests/test_work_narration_privacy.py tests/test_work_observer_role_sanitization.py tests/test_cooperative_context_recovery.py`. - `python -m ruff check .` and both architecture/README view checks passed. - All 14 GitHub CI checks passed on `0f51b41377f46585047db25578bcac6696aab9b1`, including all three isolated Python-suite shards, Electron build, Windows desktop smoke, and platform/profile checks. Maintainer manual acceptance passed on 2026-10-11 (Pacific/Auckland). The superseded partial local full-suite run is not counted as a full pass. - `npm --prefix electron run build` passed in the maintainer local checkout after synchronizing its pre-existing local voice-catalog customization into its generated catalog. Those local customization files are excluded from this PR. - Live DeepSeek main-model comparison using a synthetic 4.5 KB report, production persona/frame and transport: without the report it explicitly lacked the details; with the report it returned the late 722/372 counts and Lean caveat with `action: null`. No UI/audio run or real-news truth validation is claimed. - Regression coverage includes concise speech, terminal repair, bounded progress, error/cancelled results, session/context isolation, latest-attempt replacement, archive handling, restart recovery, and the standalone Provider Slice/narration path. Related to #176. #177 passed all 16 checks and has since merged; this branch includes the main update. PR only, per maintainer request. ## 中文说明 终态研究报告在 Slice 卡片中是完整的，但红莉栖此前只收到简短播报；后续主对话也没有读取台账中的报告正文，所以无法回答报告后半段的细节。 本次保留终态报告输入全文，并让当前会话所接收的 Work 将最新终态结果提供给后续对话；过程播报和语音输出仍保持简短。继续沿用 Provider 来源、Host 完成度及会话绑定，重启后可从现有台账读取，不新增持久化结构。 同时将“通常最多 6 句；要求详细或准确充分回答确有需要时可超过”的规则加入内置红莉栖日语人格。人格覆盖字段默认仍留空，已有自定义覆盖继续优先。 包含 #177 的最新提交已通过 223 项相关回归和静态检查；实际主模型的合成报告对照已通过。本地主线及 Electron 构建已更新，供维护者实机验收。最新提交的 14 项 CI 已全部通过，包括三组全量 Python 测试、Electron 构建和 Windows 桌面冒烟。维护者于 2026-10-11（Pacific/Auckland）确认实机验收通过。合成报告验证不代表对真实新闻的事实核验。
Lucas1479
added a commit
that referenced
this pull request
Oct 10, 2026
Prepare `v0.16.1-alpha.0` as the next source Alpha after `v0.16.0-alpha.0`, covering merged #177–#181: automatic NVIDIA TTS acceleration, complete terminal Work context, ASR-overlapped application preparation, and cold translation-client setup outside the audio event loop. Synchronize Python/Electron product versions and lock metadata, update both README release links, include the new release/acceptance documents in the source ZIP, and test native settings upgrade from 0.16.0. The native upgrade probe also resolves history from the repository root and preserves the previous tag's settings/shared-module layout: its old single-file extraction failed on 0.16.0's catalog imports. No dependency selection or production behavior changes are added by this preparation PR. Validation completed locally: 128 release/voice/entry regression tests; `uv lock --check --offline`; version/lock consistency; third-party provenance release gate; `git diff --check`; clean source-archive gate (3,891 selected files, zero errors/warnings). All 3,891 ZIP file hashes and its embedded manifest were verified. Local Electron build and native encrypted-settings upgrade from both 0.16.0 and 0.15.2 passed; the actual extracted ZIP was separately installed/built and passed the 0.16.0 native upgrade; the initial module-extraction failure and repair are recorded in acceptance. Full final-head CI and extracted-source install/build/package/startup are the merge gates. The acceptance record contains sanitized real microphone observations from #179/#180 and explicitly separates them from a controlled A/B, acoustic measurement, ASR speedup, physical barge-in acceptance or #181 VN acceptance. Existing dependency findings and experimental-platform limits remain. User configuration, raw logs and external media are excluded. After checks pass, publication uses the clean merged commit, the repository's source-Alpha Latest convention, a verified tag, final source ZIP, per-file manifest and SHA-256 checksums. Related: #176.
16 tasks done
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fresh NVIDIA GPT-SoVITS installations now choose CUDA Graph automatically and use the optional FlashAttention KV-cache extension when the actual speech device, precision and installed extension support it. Both existing switches default to
auto; explicit0/1overrides remain supported, and the desktop inference-mode control includes Automatic (recommended).Behavior and compatibility
.env.exampleexplain recommended auto settings, manual overrides and optional extension installation. Sentence/utterance grouping is unchanged.Validation
npm run buildpassed.git diff --checkpassed.tools/verify_flash_attention.py --compute --device cuda:0passed all 8 FP16/BF16 forward and KV-cache comparisons.AMADEUS_RUN_GPU_TESTS=1 python -m pytest tests/test_local_tts_device_context.py -k auto_graph_and_flash_capture_replay -q: passed automatic selection plus combined FlashAttention/Graph capture and replay on a small random-weight T2S model. This is kernel/runtime evidence, not a full speech-quality or latency benchmark.Related: #176.