Hugging Face's logo

nvidia
/
diar_sortformer_4spk-v1

Automatic Speech Recognition
NeMo
Safetensors
PyTorch
sortformer
speaker-diarization
speaker-recognition
speech
audio
Transformer
FastConformer
Conformer
NEST
NeMo
Eval Results (legacy)
Model card Files
xet
 Community
13

Multi-gpu usage

#3
by modjo-ai - opened
Discussion
modjo-ai

On my 24gb gpu I can only process audio files less than 10 minutes. Is it possible to run it on multiple, e.g. 4 gpus to not have CUDA OOM issues?
imedennikov
 NVIDIA org

Sorry this is not possible. But we will soon release a streaming version which will have no limits on the audio length.
modjo-ai changed discussion status to closed
leminhnguyen

@imedennikov , any update on releasing the streaming model?
imedennikov
 NVIDIA org

@leminhnguyen this work is still in progress. We hope to finalize it in Q2'25.
leminhnguyen

Hi @imedennikov , one more question: Is the streaming model just a code change, or do we need to re-train the sortformer model to incorporate the new streaming mechanism? For example cache-aware streaming ASR (https://github.com/NVIDIA/NeMo/blob/main/tutorials/asr/Online_ASR_Microphone_Demo_Cache_Aware_Streaming.ipynb) need to be trained from scratch.
imedennikov
 NVIDIA org

@leminhnguyen model retraining will be needed.

Sign up or log in to comment