Kale TTS keeps LSTM layers on CPU in two lines until transformers.js #1956 lands
NERDDISCO · x · 2026-11-20
NERDDISCO explains that on WebGPU two extra lines currently keep six small LSTM layers on the CPU; once transformers.js PR #1956 merges, a model's config.json can carry that and the lines disappear. He also asks maintainers to merge PR #1879 (onnxruntime bump to 1.31.1, already approved), which alone makes CPU inference 20% faster for his models.
Related event: Kale TTS Generates WAV in 20 Lines via transformers.js(2 posts)→