DGX Spark / GB10: PersonaPlex runs but local Web UI audio latency continuously increases

Backend impacted

The PyTorch implementation

Operating system

Linux

Hardware

GPU with CUDA

Description

Environment

• Device: NVIDIA DGX Spark / GB10
• GPU: NVIDIA GB10
• Driver: 580.142
• CUDA: 13.0
• Docker: 29.2.1
• Docker Compose: v5.0.2
• PyTorch in container: 2.10.0a0+b558c986e8.nv25.11
• CUDA available in PyTorch: True
• Device detected by PyTorch: NVIDIA GB10
• Compute capability: 12.1

Problem

PersonaPlex starts successfully and the Web UI opens, but the audio is very laggy.

Even when tested locally on the DGX using:

https://127.0.0.1:8998

the latency and missed audio keep increasing.

Example Web UI stats:

• Latency: 23s+
• Missed audio: 23s+

How I ran it

I cloned the repo and ran it with Docker Compose:

git clone https://github.com/NVIDIA/personaplex.git
cd personaplex
docker compose build
docker compose up

What I tried

The default setup had CUDA/PyTorch issues, so I changed the Docker base image to:

FROM nvcr.io/nvidia/pytorch:25.11-py3

After that, PyTorch correctly detected the GB10 GPU.

I also tested both settings:

NO_TORCH_COMPILE=1

and:

NO_TORCH_COMPILE=0

With NO_TORCH_COMPILE=0, GPU usage reaches around 90%, but the audio latency still keeps increasing.

Question

Is DGX Spark / GB10 expected to run PersonaPlex in real time?

Are there recommended Docker, PyTorch, or runtime settings for GB10 / sm_121?

Extra information

Docker GPU works PyTorch CUDA works GB10 is detected correctly PersonaPlex loads successfully Local 127.0.0.1 is still laggy Latency increases over time

Environment

Operating system version: Ubuntu 24.04 / NVIDIA DGX Spark environment

Python version: Python 3.12

PyTorch version: 2.10.0a0+b558c986e8.nv25.11

CUDA version: 13.0

GPU model and memory: NVIDIA GB10, 121 GiB unified system memory. nvidia-smi reports GPU memory as "Not Supported" on this device.