2.101.33

2.101.33

Added

• Docs: a getting-started video under docs/videos. It follows one podcast on a phone, from the Add feed search through the processing queue to a podcast app. There the served feed loads with the MinusPod badge on its cover, and the episode plays without its ads. The video is animated like the pipeline and processing-modes videos, uses placeholder data, and comes in portrait and landscape cuts. It is embedded in the README quick start and in Finding Feeds & Usage.

Fixed

• The GPU image carries NVRTC, the CUDA runtime compiler, again. It used to come in with PyTorch, so 2.101.32 shipped without it and logged Failed to load libnvrtc.so.12 each time Parakeet loaded. Transcription still worked in testing, but cuDNN uses NVRTC to compile some kernels at runtime. The image build now fails if a CUDA library that onnxruntime preloads is missing.

2.101.32

Fixed

• 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. It covers compute capability 5.0 through 12.0, so Turing through Blackwell cards see no change. The PyTorch build the image carried pulled in cuDNN 9.20, which dropped them (#988).

Changed

• PyTorch is gone from both images. It was only used to read free VRAM and the GPU name, which now come from NVML, so the readings work on every card including ones PyTorch had no kernels for. The GPU image drops from 15.5 GB to 8.1 GB.
• Dependency updates: ctranslate2 4.8.5 (fixes FP16 overflow and precision in CUDA attention, which Whisper uses on float16), anthropic 2.6.0, openai 3.46.0, anyio 4.16.0, soupsieve 4.0.1, wavesurfer.js 8.0.5, @tanstack/react-query 5.106.1, vite 8.3.6, @vitejs/plugin-react 6.2.1, @tailwindcss/vite 4.3.4 and docker/login-action 4.8.0 (#977 to #987).