NEWS
n3d 0.1.0 (2026-10-14)
n3d 0.0.1.1
- On CUDA, encoder layers run as one traced TorchScript function shared by
all 31 layers, and each chunk ends with a minor garbage collection. Peak
device memory drops from 1.7-2.1 GiB to 0.47-0.55 GiB (weights are
0.37 GiB), and a streaming step from 55 ms to about 18 ms on an RTX 5060
Ti.
load_n3d() traces and warms up the layer once (about 1 s).
options(n3d.jit = FALSE) runs the plain R functions; the CPU always
does.
- Log-mel features and logits for the whole input stay on the CPU, and
one chunk at a time moves to the device, so device memory no longer
grows with the length of the audio (0.52 GiB for 9.8 min).
n3d 0.0.1
- Initial version: native R 'torch' port of NVIDIA Nemotron 3 Diarization
with offline diarization (
diarize()), streaming sessions
(n3d_stream(), n3d_stream_push(), n3d_stream_finish()), log-mel
features and the arrival-order speaker cache implemented in R, and
pinned-revision weight download through 'hfhub'.