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 contains the following updates:
| Package | Update | Change |
|---|---|---|
| [ghcr.io/remsky/kokoro-fastapi-cpu](https://github.com/remsky/Kokoro-FastAPI) | minor | `v0.6.0` → `v0.7.0` |
---
### Release Notes
<details>
<summary>remsky/Kokoro-FastAPI (ghcr.io/remsky/kokoro-fastapi-cpu)</summary>
### [`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 ([#​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 (#​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 CLI](https://github.com/renovatebot/renovate).
<!--renovate-debug:eyJjcmVhdGVkSW5WZXIiOiI0NC41LjMiLCJ1cGRhdGVkSW5WZXIiOiI0NC42LjAiLCJ0YXJnZXRCcmFuY2giOiJmcmVzaC1zdGFydCIsImxhYmVscyI6W119-->
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
This PR contains the following updates:
v0.6.0→v0.7.0Release Notes
remsky/Kokoro-FastAPI (ghcr.io/remsky/kokoro-fastapi-cpu)
v0.7.0Compare Source
Added
AGENTS.mdcontributor guidelines, plusSKILL.mdnotes for the API, benchmarks, and web areas.Changed / Optimizations
requests,python-dotenv, cappedtransformers<6COPY --excludeCompressed image sizes + new bases:
python:3.10-slim->python:3.12-slimcuda:12.6.3-cudnn-runtime->cuda:12.6.3-basecuda:12.8.1-cudnn-runtime->cuda:12.8.1-baseFixed
.dockerignoreFixed pycache ignore pattern to**/to ensure nested .pyc/etc stay out of build contexts.pydubimports from the audio services._find_filerejects 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.NUMBER_PATTERNno longer swallows hyphens meant as range separators, added version-number handling (2.0.1renders as "two point zero point one" instead of being split).Configuration
📅 Schedule: (UTC)
🚦 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.
This PR has been generated by Mend Renovate CLI.
View command line instructions
Checkout
From your project repository, check out a new branch and test the changes.