ops/homelab
0
0
Fork 0
Code Issues Pull requests Projects Releases Packages Activity

chore(deps): update ghcr.io/remsky/kokoro-fastapi-gpu docker tag to v0.9.0 #131

Open
renovate-bot wants to merge 1 commit from renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x into main
pull from: renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x
merge into: ops:main
Conversation 0 Commits 1 Files changed 1 +1 -1
renovate-bot commented 2026-08-04 23:14:55 +00:00
Member
Copy link

This PR contains the following updates:

Package Update Change
ghcr.io/remsky/kokoro-fastapi-gpu minor v0.6.0 → v0.9.0

Release Notes

remsky/Kokoro-FastAPI (ghcr.io/remsky/kokoro-fastapi-gpu)

v0.9.0

Compare Source

Added
  • Voice clone-tuning from a short reference clip via inno-kokoro:
    • POST /dev/tune speaks with the tuned voice (passes through to /v1/audio/speech) or returns the .pt via return_voice_pack=true,
    • When ALLOW_LOCAL_VOICE_SAVING=true, saves it as <name>_tuned with save_voice=<name>.
    • Off by default, requires ENABLE_INNO_TUNER=true. See docs/inno-tune.md.
  • Web player: Tune tab (record or upload a clip, generate, download or save the pack).
  • Four tuned voices bundled with _inno suffix:
    • af_amelia_inno, af_goodall_inno, am_price_inno, bm_atten_inno.
  • normalization_options.remove_emoji, off by default: drop emoji instead of reading their names (#​353).
  • normalization_options.caps_normalization, on by default: all-caps names and headers read as words, short acronyms (FBI) still spelled.
Changed
  • Normalizer split into per-language classes, registry keyed by lang code.
    • CONTRIBUTING.md documents standardized contract to add additional languages
  • DEFAULT_VOICE applies to requests that omit a voice, not just warmup. Reported as default_voice on /v1/audio/voices.
  • Web player:
    • normalize checkbox replaced by a menu with every normalization_options field;
    • voice list sorted by grade then name, DEFAULT_VOICE preselected.
Fixed
  • Aborting a stream mid-playback no longer segfaults the server on a follow-up request with a new lang_code engine (#​288, #​337).
  • Blank lines end a sentence, single newlines still join (#​519, #​525 by @​Christian-Sidak).
  • Very long unpunctuated non-English text no longer truncates.
  • Blank or emoji-only input is a 400 on the streaming path too, not an empty 200.
  • Web player: voice search and alias minor fixes.
  • Normalization fixes:
    • Number reading: 3,497 as thousands not a year (#​259);
    • MP3, B2B, v1.0, COVID-19 read as written;
    • .5 as zero point five; long digit runs no longer stall or 500.
    • Phone numbers, 3.5 GHz, 1 min (with unit_normalization), DVDs/DVD's, 12:30:15 pm read correctly.
    • -- reads as a dash and no longer fuses words or drops word timestamps (#​249).

v0.8.2

Compare Source

Added
  • Optional model auto-unload after an idle timeout (MODEL_AUTO_UNLOAD_TIMEOUT_SECONDS, default off) to release VRAM. Reloads on the next request. /dev/model reports load/idle state and POST /dev/reload pre-warms the model, both behind ALLOW_DEV_UNLOAD.
  • /v1/audio/voices entries carry the per-voice grades from VOICES.md, where graded; added in web player dropdown w/ hover for the target quality and training duration.
  • Web player: Normalize text toggle in settings, off sends input as written (#​391, #​523 by @​webdevsamran).
Changed
  • Sentence splitting via UAX #​29 segmentation instead of regex; e.g: etc., Dr., and CJK punctuation split more accurately (#​308, adapts #​415 by @​lionel-rowe).
  • ROCm: MIOpen off by default for better performance re: tensor shape recompilation (2.4s vs 0.27s on gfx1100). ENABLE_MIOPEN=true restores it (#​518 by @​s-kerdel).
  • Compose files no longer mount api/ or set defaults already baked into the image (DOWNLOAD_MODEL, PYTHONPATH, etc).
Fixed

v0.8.1

Compare Source

Fixed
  • /v1/audio/voices/combine accepts weighted syntax (af_bella(2)+af_sky(1)), matching the speech endpoints (#​285).
  • Oversized or pause-heavy requests now return 400 to avoid exhausting memory (reported by @​sshpie, GHSA-f64g-9jmv-pm22). Two new configurable limits:
    • MAX_INPUT_LENGTH (default 1_000_000) caps characters of text per request.
    • MAX_TOTAL_PAUSE_S (default 300) caps total [pause:Ns] / SSML <break> silence per request.
  • Native Windows installs (start-cpu.ps1 etc) no longer need a C++ toolchain: pyopenjtalk-plus (a drop-in fork with prebuilt Windows wheels) replaces pyopenjtalk on win32 only (#​508, proposed by @​siliconfps). Needs a recent uv. Linux, macOS, and Docker are unchanged.
Changed
  • Documented WEB_CONCURRENCY (uvicorn worker count) in docs/configuration.md for parallel model loads/concurrency (#​115, #​358).
  • Improved time-to-first-audio; sentence phonemization emits to avoid a first full-request pass. Some gradual latency growth at larger input sizes due to normalization pass.
Input v0.8.0 (eager) v0.8.1 (lazy)
5k chars 0.31 s 0.26 s -16%
10k chars 0.30 s 0.27 s -10%
50k chars 0.45 s 0.29 s -36%
100k chars 0.73 s 0.31 s -58%
250k chars 1.41 s 0.45 s -68%
500k chars 2.78 s 0.56 s -80%
1M chars 5.21 s 0.85 s -84%

v0.8.0

Compare Source

Added
  • Multi-speaker input on /v1/audio/speech and /dev/captioned_speech (#​294). Opt in per request with allow_voice_tags: true; disable server-wide with ENABLE_VOICE_TAGS=false.
    • Inline [voice:name] tags switch speaker mid-text.
    • voice_aliases mapping for named weighted voice mixes, with optional per-alias rate.
    • /dev/captioned_speech timestamps carry the resolved voice per word; the field is absent unless allow_voice_tags is on, so existing responses are unchanged.
  • POST /dev/dialogue for ordered multi-speaker turns.
  • SSML input (experimental). Disable server-wide with ENABLE_SSML=false.
    • ssml: true on /v1/audio/speech and /dev/captioned_speech translates and speaks in one call. Requires allow_voice_tags: true, since the translation emits [voice:] and [rate:] spans.
    • POST /dev/ssml returns the translated tokens as text instead, for inspecting them before synthesis.
  • return_timing on /v1/audio/speech: per-chunk {text, start, end} JSON sidecar next to the download (powers the web reader).
  • MAX_PAUSE_DURATION_S (default 60) caps a single [pause:Ns] tag or SSML <break>.
  • Web UI:
    • Voice alias/tag cast builder with import/export, pinning, and per-alias rate, synced with the editor (re: parallel work by @​radzrader, #​272).
    • Read-along mode: sentence highlighting synced to playback, bidirectional click to seek.
    • Find/replace across pages, direct page-number entry, download menu (audio / timings / both).
  • Wiki pages moved into docs/, versioned alongside the code.
Changed
  • Docker images compile to bytecode at build, ~40% faster startup.
  • Containers launch uvicorn directly rather than through uv run, which resolves startup permission failures on Unraid and similar hosts.
  • [rate:] tags scale the speaking voice's alias rate instead of replacing it, so a voice calibrated to 0.8 stays proportionally slower under [rate:1.1]. Matches how SSML engines treat rate.
  • Speed bounds (0.25 to 4.0) shared across speed fields and SSML.
  • Unrecognized .env keys warn at startup instead of refusing to boot.
  • README config table now covers every setting.
Fixed
  • Long generations swap from the live stream to the finished file as soon as it lands, so the scrubber shows true duration and seeking works mid-run.
  • Volume control state reconnected to the player.
Removed
  • Unused ffmpeg from all images (~600MB); audio encoding already runs through PyAV's bundled copy.
  • Dead pydub dependency.
  • Unreachable list form of voice from the speech parser and unused VoiceCombineRequest schema.
  • Legacy Gradio UI (ui/) code cruft; superseded by the web player since ~v0.2.0
  • Legacy ONNX config compose vars, endpoints e.g /debug/session_pools.
  • OUTPUT_DIR, OUTPUT_DIR_SIZE_LIMIT_MB, SAMPLE_RATE settings, never read.

v0.7.2

Compare Source

Security
  • fastapi>=0.128.8, starlette>=1.3.1 to close CVE-2025-62727 (quadratic Range header parsing in FileResponse, reachable through the audio download path) (#​500).
Changed
  • CORS allow_credentials now defaults off. Starlette 1.x echoes the caller's origin with allow-credentials: true where 0.47 returned *; nothing here uses cookies or auth, so this keeps the prior behavior.
  • Docker build cache moved from GHA to the GHCR registry so forks and local builds can pull it, plus uv cache mounts and reordered test-client layers (#​501).
  • response_format docs (correctly) now list aac as supported.
Fixed

v0.7.1

Compare Source

Added
  • /v1/download/{filename} takes an optional ?name= save-as name (sanitized, stored extension kept) and sets it in Content-Disposition. Omitting it keeps the previous name.
  • Web UI keyboard navigation and ARIA labeling across header, player controls, and editor.
Changed
  • Content-Disposition is now built by FileResponse rather than by hand, so the filename comes back quoted (filename="x.mp3") instead of bare. The name itself is unchanged when ?name= is omitted.
  • Web UI restyle: better use of space, responsive down to slim widths, playbar pinned to the bottom on narrow viewports.
  • Waveform slowed and softened, made framerate-independent, respects prefers-reduced-motion.
  • README: AMD GPU (ROCm) troubleshooting, clarified docker-compose comments.
Fixed
  • Downloads save as {voice}_{timestamp}.{format}, not the temp name (#​338). Covers right-click "Save audio as" too, since Content-Disposition outranks the link's download attribute.
  • Aborted streams no longer surface as playback failures; a user-initiated MEDIA_ERR_ABORTED is told apart from a real error.
  • Stream-to-file swap settles pending buffer operations instead of leaving the feeder awaiting forever.

v0.7.0

Compare Source

Added
  • AGENTS.md contributor guidelines, plus SKILL.md notes for the API, benchmarks, and web areas.
Changed / Optimizations
  • Docker images build on Python 3.12 (project floor stays 3.10 for local installs). Rust dropped from the CPU builder.
  • Runtime dependencies trimmed to remove deprecated imports
  • bumped requests,python-dotenv, capped transformers<6
  • Builds now explicitly require BuildKit (default since Docker 23, ~Jan 2023); utilizing COPY --exclude
  • Model bake reworked to ensure weights land exactly once (whether prexisting or downloaded at build)
  • ROCm image now bakes the model at build like CPU/GPU (instead of a first-run fetch)
  • GPU runtime now only uses torch shipped cuDNN/etc via pip wheels (#​482). (see table below for size changes)
  • Transcription benchmark reports split by device; RTF and first-token baselines refreshed.

Compressed image sizes + new bases:

Image v0.6.0 v0.7.0 Runtime base
cpu 1.66 GB 1.56 GB python:3.10-slim -> python:3.12-slim
gpu 6.81 GB 4.68 GB cuda:12.6.3-cudnn-runtime -> cuda:12.6.3-base
gpu (cu128) 8.11 GB 5.23 GB cuda:12.8.1-cudnn-runtime -> cuda:12.8.1-base
rocm 13.08 GB 13.36 GB unchanged; model now baked in
Fixed
  • Model validation checksums against downloaded release artifact to ensure consistency, downloads first to a temp dir to avoid clobbering pre-existing models in the case of network issues/corrupted downloads.
  • Model validation also rejects any custom files under 100MB to avoids false pass results (e.g. a 9-byte "Not found"), allowing a re-download instead of passing (#​301).
  • .dockerignore Fixed pycache ignore pattern to **/ to ensure nested .pyc/etc stay out of build contexts.
  • Removed dead pydub imports from the audio services.
  • _find_file rejects lookups that escape its search roots (../ sequences, absolute paths) to avoid unintentional exposure of files outside the voices/models/web dirs. Symlinks placed inside those dirs resolve as before.
  • Text normalizer: anchored decimal regex to prevent quadratic backtracking on digit floods, reordered range substitution so NUMBER_PATTERN no longer swallows hyphens meant as range separators, added version-number handling (2.0.1 renders as "two point zero point one" instead of being split).

Configuration

📅 Schedule: (UTC)

  • Branch creation
    • At any time (no schedule defined)
  • Automerge
    • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about this update again.

  • If you want to rebase/retry this PR, check this box

This PR has been generated by Mend Renovate.

This PR contains the following updates: | Package | Update | Change | |---|---|---| | [ghcr.io/remsky/kokoro-fastapi-gpu](https://github.com/remsky/Kokoro-FastAPI) | minor | `v0.6.0` → `v0.9.0` | --- ### Release Notes <details> <summary>remsky/Kokoro-FastAPI (ghcr.io/remsky/kokoro-fastapi-gpu)</summary> ### [`v0.9.0`](https://github.com/remsky/Kokoro-FastAPI/blob/HEAD/CHANGELOG.md#v090---2026-09-09) [Compare Source](https://github.com/remsky/Kokoro-FastAPI/compare/v0.8.2...v0.9.0) ##### Added - Voice clone-tuning from a short reference clip via [inno-kokoro](https://github.com/remsky/inno-kokoro): - `POST /dev/tune` speaks with the tuned voice (passes through to `/v1/audio/speech`) or returns the `.pt` via `return_voice_pack=true`, - When `ALLOW_LOCAL_VOICE_SAVING=true`, saves it as `<name>_tuned` with `save_voice=<name>`. - Off by default, requires `ENABLE_INNO_TUNER=true`. See [docs/inno-tune.md](docs/inno-tune.md). - Web player: Tune tab (record or upload a clip, generate, download or save the pack). - Four tuned voices bundled with `_inno` suffix: - `af_amelia_inno`, `af_goodall_inno`, `am_price_inno`, `bm_atten_inno`. - `normalization_options.remove_emoji`, off by default: drop emoji instead of reading their names ([#&#8203;353](https://github.com/remsky/Kokoro-FastAPI/issues/353)). - `normalization_options.caps_normalization`, on by default: all-caps names and headers read as words, short acronyms (`FBI`) still spelled. ##### Changed - Normalizer split into per-language classes, registry keyed by lang code. - CONTRIBUTING.md documents standardized contract to add additional languages - `DEFAULT_VOICE` applies to requests that omit a voice, not just warmup. Reported as `default_voice` on `/v1/audio/voices`. - Web player: - normalize checkbox replaced by a menu with every `normalization_options` field; - voice list sorted by grade then name, `DEFAULT_VOICE` preselected. ##### Fixed - Aborting a stream mid-playback no longer segfaults the server on a follow-up request with a new lang\_code engine ([#&#8203;288](https://github.com/remsky/Kokoro-FastAPI/issues/288), [#&#8203;337](https://github.com/remsky/Kokoro-FastAPI/issues/337)). - Blank lines end a sentence, single newlines still join ([#&#8203;519](https://github.com/remsky/Kokoro-FastAPI/issues/519), [#&#8203;525](https://github.com/remsky/Kokoro-FastAPI/issues/525) by [@&#8203;Christian-Sidak](https://github.com/Christian-Sidak)). - Very long unpunctuated non-English text no longer truncates. - Blank or emoji-only input is a 400 on the streaming path too, not an empty 200. - Web player: voice search and alias minor fixes. - Normalization fixes: - Number reading: `3,497` as thousands not a year ([#&#8203;259](https://github.com/remsky/Kokoro-FastAPI/issues/259)); - `MP3`, `B2B`, `v1.0`, `COVID-19` read as written; - `.5` as zero point five; long digit runs no longer stall or 500. - Phone numbers, `3.5 GHz`, `1 min` (with `unit_normalization`), `DVDs`/`DVD's`, `12:30:15 pm` read correctly. - `--` reads as a dash and no longer fuses words or drops word timestamps ([#&#8203;249](https://github.com/remsky/Kokoro-FastAPI/issues/249)). ### [`v0.8.2`](https://github.com/remsky/Kokoro-FastAPI/blob/HEAD/CHANGELOG.md#v082---2026-09-05) [Compare Source](https://github.com/remsky/Kokoro-FastAPI/compare/v0.8.1...v0.8.2) ##### Added - Optional model auto-unload after an idle timeout (`MODEL_AUTO_UNLOAD_TIMEOUT_SECONDS`, default off) to release VRAM. Reloads on the next request. `/dev/model` reports load/idle state and `POST /dev/reload` pre-warms the model, both behind `ALLOW_DEV_UNLOAD`. - `/v1/audio/voices` entries carry the per-voice grades from [VOICES.md](https://huggingface.co/hexgrad/Kokoro-82M/blob/main/VOICES.md), where graded; added in web player dropdown w/ hover for the target quality and training duration. - Web player: `Normalize text` toggle in settings, off sends input as written ([#&#8203;391](https://github.com/remsky/Kokoro-FastAPI/issues/391), [#&#8203;523](https://github.com/remsky/Kokoro-FastAPI/issues/523) by [@&#8203;webdevsamran](https://github.com/webdevsamran)). ##### Changed - Sentence splitting via UAX [#&#8203;29](https://github.com/remsky/Kokoro-FastAPI/issues/29) segmentation instead of regex; e.g: `etc.`, `Dr.`, and CJK punctuation split more accurately ([#&#8203;308](https://github.com/remsky/Kokoro-FastAPI/issues/308), adapts [#&#8203;415](https://github.com/remsky/Kokoro-FastAPI/issues/415) by [@&#8203;lionel-rowe](https://github.com/lionel-rowe)). - ROCm: MIOpen off by default for better performance re: tensor shape recompilation (2.4s vs 0.27s on gfx1100). `ENABLE_MIOPEN=true` restores it ([#&#8203;518](https://github.com/remsky/Kokoro-FastAPI/issues/518) by [@&#8203;s-kerdel](https://github.com/s-kerdel)). - Compose files no longer mount `api/` or set defaults already baked into the image (`DOWNLOAD_MODEL`, `PYTHONPATH`, etc). ##### Fixed - `DEFAULT_VOICE_CODE` now applies on the speech endpoints and `/dev/generate_from_phonemes` ([#&#8203;514](https://github.com/remsky/Kokoro-FastAPI/issues/514), [#&#8203;515](https://github.com/remsky/Kokoro-FastAPI/issues/515) by [@&#8203;Christian-Sidak](https://github.com/Christian-Sidak), [#&#8203;516](https://github.com/remsky/Kokoro-FastAPI/issues/516)). - Entrypoint and `start-*.sh` honour `HOST` / `PORT`, so `HOST=::` binds IPv6-only ([#&#8203;408](https://github.com/remsky/Kokoro-FastAPI/issues/408), reported by [@&#8203;felixls](https://github.com/felixls)). ### [`v0.8.1`](https://github.com/remsky/Kokoro-FastAPI/blob/HEAD/CHANGELOG.md#v081---2026-08-24) [Compare Source](https://github.com/remsky/Kokoro-FastAPI/compare/v0.8.0...v0.8.1) ##### Fixed - `/v1/audio/voices/combine` accepts weighted syntax (`af_bella(2)+af_sky(1)`), matching the speech endpoints ([#&#8203;285](https://github.com/remsky/Kokoro-FastAPI/issues/285)). - Oversized or pause-heavy requests now return 400 to avoid exhausting memory (reported by [@&#8203;sshpie](https://github.com/sshpie), GHSA-f64g-9jmv-pm22). Two new configurable limits: - `MAX_INPUT_LENGTH` (default 1\_000\_000) caps characters of text per request. - `MAX_TOTAL_PAUSE_S` (default 300) caps total `[pause:Ns]` / SSML `<break>` silence per request. - Native Windows installs (`start-cpu.ps1` etc) no longer need a C++ toolchain: `pyopenjtalk-plus` (a drop-in fork with prebuilt Windows wheels) replaces `pyopenjtalk` on win32 only ([#&#8203;508](https://github.com/remsky/Kokoro-FastAPI/issues/508), proposed by [@&#8203;siliconfps](https://github.com/siliconfps)). Needs a recent `uv`. Linux, macOS, and Docker are unchanged. ##### Changed - Documented `WEB_CONCURRENCY` (uvicorn worker count) in `docs/configuration.md` for parallel model loads/concurrency ([#&#8203;115](https://github.com/remsky/Kokoro-FastAPI/issues/115), [#&#8203;358](https://github.com/remsky/Kokoro-FastAPI/issues/358)). - Improved time-to-first-audio; sentence phonemization emits to avoid a first full-request pass. Some gradual latency growth at larger input sizes due to normalization pass. <div align="center"> | Input | v0.8.0 (eager) | v0.8.1 (lazy) | | | :--------- | -------------: | ------------: | ---: | | 5k chars | 0.31 s | 0.26 s | -16% | | 10k chars | 0.30 s | 0.27 s | -10% | | 50k chars | 0.45 s | 0.29 s | -36% | | 100k chars | 0.73 s | 0.31 s | -58% | | 250k chars | 1.41 s | 0.45 s | -68% | | 500k chars | 2.78 s | 0.56 s | -80% | | 1M chars | 5.21 s | 0.85 s | -84% | </div> ### [`v0.8.0`](https://github.com/remsky/Kokoro-FastAPI/blob/HEAD/CHANGELOG.md#v080---2026-08-14) [Compare Source](https://github.com/remsky/Kokoro-FastAPI/compare/v0.7.2...v0.8.0) ##### Added - Multi-speaker input on `/v1/audio/speech` and `/dev/captioned_speech` ([#&#8203;294](https://github.com/remsky/Kokoro-FastAPI/issues/294)). Opt in per request with `allow_voice_tags: true`; disable server-wide with `ENABLE_VOICE_TAGS=false`. - Inline `[voice:name]` tags switch speaker mid-text. - `voice_aliases` mapping for named weighted voice mixes, with optional per-alias `rate`. - `/dev/captioned_speech` timestamps carry the resolved `voice` per word; the field is absent unless `allow_voice_tags` is on, so existing responses are unchanged. - `POST /dev/dialogue` for ordered multi-speaker turns. - SSML input (experimental). Disable server-wide with `ENABLE_SSML=false`. - `ssml: true` on `/v1/audio/speech` and `/dev/captioned_speech` translates and speaks in one call. Requires `allow_voice_tags: true`, since the translation emits `[voice:]` and `[rate:]` spans. - `POST /dev/ssml` returns the translated tokens as text instead, for inspecting them before synthesis. - `return_timing` on `/v1/audio/speech`: per-chunk `{text, start, end}` JSON sidecar next to the download (powers the web reader). - `MAX_PAUSE_DURATION_S` (default 60) caps a single `[pause:Ns]` tag or SSML `<break>`. - Web UI: - Voice alias/tag cast builder with import/export, pinning, and per-alias rate, synced with the editor (re: parallel work by [@&#8203;radzrader](https://github.com/radzrader), [#&#8203;272](https://github.com/remsky/Kokoro-FastAPI/discussions/272)). - Read-along mode: sentence highlighting synced to playback, bidirectional click to seek. - Find/replace across pages, direct page-number entry, download menu (audio / timings / both). - Wiki pages moved into `docs/`, versioned alongside the code. ##### Changed - Docker images compile to bytecode at build, \~40% faster startup. - Containers launch uvicorn directly rather than through `uv run`, which resolves startup permission failures on Unraid and similar hosts. - `[rate:]` tags scale the speaking voice's alias rate instead of replacing it, so a voice calibrated to 0.8 stays proportionally slower under `[rate:1.1]`. Matches how SSML engines treat rate. - Speed bounds (0.25 to 4.0) shared across speed fields and SSML. - Unrecognized `.env` keys warn at startup instead of refusing to boot. - README config table now covers every setting. ##### Fixed - Long generations swap from the live stream to the finished file as soon as it lands, so the scrubber shows true duration and seeking works mid-run. - Volume control state reconnected to the player. ##### Removed - Unused `ffmpeg` from all images (\~600MB); audio encoding already runs through PyAV's bundled copy. - Dead `pydub` dependency. - Unreachable list form of `voice` from the speech parser and unused `VoiceCombineRequest` schema. - Legacy Gradio UI (`ui/`) code cruft; superseded by the web player since \~v0.2.0 - Legacy ONNX config compose vars, endpoints e.g `/debug/session_pools`. - `OUTPUT_DIR`, `OUTPUT_DIR_SIZE_LIMIT_MB`, `SAMPLE_RATE` settings, never read. ### [`v0.7.2`](https://github.com/remsky/Kokoro-FastAPI/blob/HEAD/CHANGELOG.md#v072-2026-08-06) [Compare Source](https://github.com/remsky/Kokoro-FastAPI/compare/v0.7.1...v0.7.2) ##### Security - `fastapi>=0.128.8`, `starlette>=1.3.1` to close CVE-2025-62727 (quadratic `Range` header parsing in `FileResponse`, reachable through the audio download path) ([#&#8203;500](https://github.com/remsky/Kokoro-FastAPI/issues/500)). ##### Changed - CORS `allow_credentials` now defaults off. Starlette 1.x echoes the caller's origin with `allow-credentials: true` where 0.47 returned `*`; nothing here uses cookies or auth, so this keeps the prior behavior. - Docker build cache moved from GHA to the GHCR registry so forks and local builds can pull it, plus uv cache mounts and reordered test-client layers ([#&#8203;501](https://github.com/remsky/Kokoro-FastAPI/issues/501)). - `response_format` docs (correctly) now list `aac` as supported. ##### Fixed - FLAC and WAV no longer lose the tail end of the audio; better muxer header patching at finalize ([#&#8203;497](https://github.com/remsky/Kokoro-FastAPI/issues/497), covers [#&#8203;448](https://github.com/remsky/Kokoro-FastAPI/issues/448) and [#&#8203;463](https://github.com/remsky/Kokoro-FastAPI/issues/463)). Diagnosis by [@&#8203;Technologicat](https://github.com/Technologicat). ### [`v0.7.1`](https://github.com/remsky/Kokoro-FastAPI/blob/HEAD/CHANGELOG.md#v071---2026-08-02) [Compare Source](https://github.com/remsky/Kokoro-FastAPI/compare/v0.7.0...v0.7.1) ##### Added - `/v1/download/{filename}` takes an optional `?name=` save-as name (sanitized, stored extension kept) and sets it in `Content-Disposition`. Omitting it keeps the previous name. - Web UI keyboard navigation and ARIA labeling across header, player controls, and editor. ##### Changed - `Content-Disposition` is now built by `FileResponse` rather than by hand, so the filename comes back quoted (`filename="x.mp3"`) instead of bare. The name itself is unchanged when `?name=` is omitted. - Web UI restyle: better use of space, responsive down to slim widths, playbar pinned to the bottom on narrow viewports. - Waveform slowed and softened, made framerate-independent, respects `prefers-reduced-motion`. - README: AMD GPU (ROCm) troubleshooting, clarified docker-compose comments. ##### Fixed - Downloads save as `{voice}_{timestamp}.{format}`, not the temp name ([#&#8203;338](https://github.com/remsky/Kokoro-FastAPI/issues/338)). Covers right-click "Save audio as" too, since `Content-Disposition` outranks the link's `download` attribute. - Aborted streams no longer surface as playback failures; a user-initiated `MEDIA_ERR_ABORTED` is told apart from a real error. - Stream-to-file swap settles pending buffer operations instead of leaving the feeder awaiting forever. ### [`v0.7.0`](https://github.com/remsky/Kokoro-FastAPI/blob/HEAD/CHANGELOG.md#v070---2026-07-31) [Compare Source](https://github.com/remsky/Kokoro-FastAPI/compare/v0.6.0...v0.7.0) ##### Added - `AGENTS.md` contributor guidelines, plus `SKILL.md` notes for the API, benchmarks, and web areas. ##### Changed / Optimizations - Docker images build on Python 3.12 (project floor stays 3.10 for local installs). Rust dropped from the CPU builder. - Runtime dependencies trimmed to remove deprecated imports - bumped `requests`,`python-dotenv`, capped `transformers<6` - Builds now explicitly require BuildKit (default since Docker 23, \~Jan 2023); utilizing `COPY --exclude` - Model bake reworked to ensure weights land exactly once (whether prexisting or downloaded at build) - ROCm image now bakes the model at build like CPU/GPU (instead of a first-run fetch) - GPU runtime now only uses torch shipped cuDNN/etc via pip wheels ([#&#8203;482](https://github.com/remsky/Kokoro-FastAPI/issues/482)). (see table below for size changes) - Transcription benchmark reports split by device; RTF and first-token baselines refreshed. Compressed image sizes + new bases: <div align="center"> | Image | v0.6.0 | v0.7.0 | Runtime base | | :---------- | -------: | -------: | :------------------------------------------------ | | cpu | 1.66 GB | 1.56 GB | `python:3.10-slim` -> `python:3.12-slim` | | gpu | 6.81 GB | 4.68 GB | `cuda:12.6.3-cudnn-runtime` -> `cuda:12.6.3-base` | | gpu (cu128) | 8.11 GB | 5.23 GB | `cuda:12.8.1-cudnn-runtime` -> `cuda:12.8.1-base` | | rocm | 13.08 GB | 13.36 GB | unchanged; model now baked in | </div> ##### Fixed - Model validation checksums against downloaded release artifact to ensure consistency, downloads first to a temp dir to avoid clobbering pre-existing models in the case of network issues/corrupted downloads. - Model validation also rejects any custom files under 100MB to avoids false pass results (e.g. a 9-byte "Not found"), allowing a re-download instead of passing (#&#8203;301). - `.dockerignore` Fixed pycache ignore pattern to `**/` to ensure nested .pyc/etc stay out of build contexts. - Removed dead `pydub` imports from the audio services. - `_find_file` rejects lookups that escape its search roots (`../` sequences, absolute paths) to avoid unintentional exposure of files outside the voices/models/web dirs. Symlinks placed inside those dirs resolve as before. - Text normalizer: anchored decimal regex to prevent quadratic backtracking on digit floods, reordered range substitution so `NUMBER_PATTERN` no longer swallows hyphens meant as range separators, added version-number handling (`2.0.1` renders as "two point zero point one" instead of being split). </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] <!-- rebase-check -->If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://github.com/renovatebot/renovate). <!--renovate-debug:eyJjcmVhdGVkSW5WZXIiOiI0My4yNTMuMSIsInVwZGF0ZWRJblZlciI6IjQzLjI1My4xIiwidGFyZ2V0QnJhbmNoIjoibWFpbiIsImxhYmVscyI6W119-->
renovate-bot force-pushed renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x from 8425cb520e to a1b6ceaf86 2026-09-21 19:16:47 +00:00 Compare
renovate-bot changed title from chore(deps): update ghcr.io/remsky/kokoro-fastapi-gpu docker tag to v0.7.1 to chore(deps): update ghcr.io/remsky/kokoro-fastapi-gpu docker tag to v0.9.0 2026-09-21 19:16:49 +00:00
This pull request can be merged automatically.
You are not authorized to merge this pull request.
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x:renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x
git switch renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x

Merge

Merge the changes and update on Forgejo.
git switch main
git merge --no-ff renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x
git switch renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x
git rebase main
git switch main
git merge --ff-only renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x
git switch renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x
git rebase main
git switch main
git merge --no-ff renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x
git switch main
git merge --squash renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x
git switch main
git merge --ff-only renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x
git switch main
git merge renovate/ghcr.io-remsky-kokoro-fastapi-gpu-0.x
git push origin main
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
ops/homelab!131
No description provided.