Repository navigation
Stable release covering 2.101.32 through 2.101.34. The previous stable was 2.101.31.
New features
- Processing schedule: Queue > Queue Control > Processing schedule limits when new runs start to a weekly window, set by From and Until times and the days the window starts on. Times use the instance time zone. Outside the window nothing new starts, including Reprocess, bulk actions and play-triggered starts, which wait in the queue until it opens. Runs already in progress finish, and the idle local transcription model is unloaded to free GPU memory. It is off by default and available through
GET/PUT /api/v1/settings/processing-schedule(#993, #992). - A getting-started video in the README quick start and in Finding Feeds & Usage follows one podcast from the Add feed search to a podcast app (#991).
Fixes
- Parakeet loads on GTX 10-series and other Maxwell, Pascal and Volta cards in the GPU image. The image now pins cuDNN 9.10.2, the last release that supports those cards. Turing through Blackwell cards see no change (#991, #988).
- On a host with more than one GPU, free VRAM and the GPU name now come from the card transcription runs on, so chunk sizes are no longer taken from the other card. A shortened GPU UUID in
CUDA_VISIBLE_DEVICESworks too. Hosts with one GPU see no change (#993). - Opening a deep link from another app, such as a link to an episode's review page, could log you out of every open tab. Anonymous responses no longer set cookies, so the logged-in session is kept (#993).
- With feed authentication on, the transcript and chapters links of episodes outside the feed's episode cap now carry the feed key. Podcast apps got a 401 for them before (#993).
- Rebuilding the Recents feed or a feed's served RSS no longer reads every episode's transcript and detection data. In a benchmark on a 2.2 GB synthetic database, the Recents read went from 1.15 s to 0.16 s (#993).
- Deleting a feed or episode while it was queued or processing no longer logs a misleading "orphaned" warning or sends the episode through the retry path (#993).
- Hosts without a GPU no longer log
NVML unavailable, GPU memory and name will not be reported(#993). - ctranslate2 4.8.5 fixes FP16 overflow and precision in CUDA attention, which Whisper uses on float16 (#991).