Nemotron 3.5 Transcribe: audio input, the 65,536-token window, an hour of audio, MTP drafts by BobClawblaw · Pull Request #1632 · ashhart/TensorFold

cuda experts: a bf16 format, a checkpoint's own weights in mma fragment blocks (relu^2 up, fp32 or bf16 down)

BobClawblaw

committed

1522104

kernel contract: the bf16 experts (experts_bf16.cu), relu^2 up, fp32 and bf16 down, gates 1-5 with two controls

BobClawblaw

committed

8fd6c58

nemotron cuda: Nemotron 3.5 Transcribe's language model (nemo_speechlm) from its bf16 weights, audio rows spliced after the embeds

BobClawblaw

committed

b346d96

nemo_audio: Nemotron 3.5 Transcribe's audio encoder on CUDA (log-mel, ASR encoder, streaming Sortformer, fusion), audio-check

BobClawblaw

committed

bca570b

kernel contract: nemo_audio.cu's encoder kernels, gates 1-5 with five controls

BobClawblaw

committed

b08f1f3

build: nemo_audio's unit tests (config, FFT, bf16 rounding, Sortformer FIFO and pooling) run under test-cuda-host

BobClawblaw

committed

f055c95

nemotron cuda: null bf16x in the hand-built load-recovery engines

BobClawblaw

committed

282c782

server: audio input for Nemotron 3.5 Transcribe (input_audio and audio_url WAV parts), encoded at each request's prompt pass, rows keyed by clip in the prompt cache

BobClawblaw

committed

4d3da78

nemotron transcribe: its trained 65536-token window, an hour of audio a request (160 MiB chat bodies when the engine takes audio), MTP drafts with the clip's rows in the head

BobClawblaw

committed

b806be2

docs: the Nemotron 3.5 Transcribe recipe (one GB10, the bf16 release, audio input), its index row and the API's audio parts

BobClawblaw

committed

6cf8925

Navigation Menu

Repository navigation

Nemotron 3.5 Transcribe: audio input, the 65,536-token window, an hour of audio, MTP drafts - #1632

Nemotron 3.5 Transcribe: audio input, the 65,536-token window, an hour of audio, MTP drafts#1632
BobClawblaw wants to merge 10 commits into
ashhart:mainfrom
BobClawblaw:nemotron-transcribe-audio

Commits on Nov 1, 2026

cuda experts: a bf16 format, a checkpoint's own weights in mma fragment blocks (relu^2 up, fp32 or bf16 down)

kernel contract: the bf16 experts (experts_bf16.cu), relu^2 up, fp32 and bf16 down, gates 1-5 with two controls

nemotron cuda: Nemotron 3.5 Transcribe's language model (nemo_speechlm) from its bf16 weights, audio rows spliced after the embeds

nemo_audio: Nemotron 3.5 Transcribe's audio encoder on CUDA (log-mel, ASR encoder, streaming Sortformer, fusion), audio-check

kernel contract: nemo_audio.cu's encoder kernels, gates 1-5 with five controls

build: nemo_audio's unit tests (config, FFT, bf16 rounding, Sortformer FIFO and pooling) run under test-cuda-host

nemotron cuda: null bf16x in the hand-built load-recovery engines

server: audio input for Nemotron 3.5 Transcribe (input_audio and audio_url WAV parts), encoded at each request's prompt pass, rows keyed by clip in the prompt cache

nemotron transcribe: its trained 65536-token window, an hour of audio a request (160 MiB chat bodies when the engine takes audio), MTP drafts with the clip's rows in the head

docs: the Nemotron 3.5 Transcribe recipe (one GB10, the bf16 release, audio input), its index row and the API's audio parts