github FluidInference/FluidAudio v0.17.2
v0.17.2 — Nemotron 3 output-backing fix

latest releases: v0.17.4, v0.17.3
4 days ago

Fixes a crash in Nemotron3Diarizer on v0.17.0/v0.17.1: the first prediction could throw Output backing for feature named 'speaker_preds' is not compatible with the model's output feature description.

The Nemotron 3 bundles declare their float outputs as fp32, but the diarizer preallocated fp16 output backings. Some runtimes accept the mismatch; others reject it, including an M5 on macOS 27 with the default .all, and M3 Max with .cpuAndNeuralEngine (#951). Backings are now allocated in each output's declared type. If a runtime still rejects them, the diarizer logs a warning and continues without them. Output is unchanged. No API changes.

Separately, the Nemotron 3 ANE compile still fails on M3 (ANECCompile() → InvalidMILProgram); .cpuAndNeuralEngine now runs there via CoreML's fallback. That investigation continues in #951.

What's Changed

  • fix(diarizer/nemotron3): type output backings from the model description, retry if rejected by @Alex-Wengg in #952

Full Changelog: v0.17.1...v0.17.2

Don't miss a new FluidAudio release

NewReleases is sending notifications on new releases.