RIVA 2.13.1 VAD + ASR pipeline erroring

AI & Data Science Deep Learning (Training & Inference) Riva
1

After a successful riva-build with the MarbleNet VAD ,
the model is loaded with riva-deploy without any errors.

However, once an audio segment is received, the following error pops up:
cudaError_t 700 : "an illegal memory access was encountered" returned from 'cudaMemcpyAsync( this->data_.get(), &vec[0], vec.size() * sizeof(Real), cudaMemcpyHostToDevice, cudaStreamPerThread)' in fileriva/utils/matrix/cu_vector.cc line 90'

The full log from the container is provided below
riva_vad_error.log (257.0 KB)

Operating System: riva-speech:2.13.1
Riva Version: 2.13.1

2

Hi @ms.ibnseddik

Thanks for your interest in Riva

I guess you are using a VMWare Setup

I am guessing the issue to be with uvm

Can you try

pciPassthru0.cfg.enable_uvm =1

NOTE - Please use the correct GPU ID in the above command, in my above exmaple GPU ID is 0
and let me know if it helps

Thanks

3

Hello @ms.ibnseddik

How were you able to configure MarbleNet VAD in the ASR pipeline? I am unable to do so. I appreciate any information.