Troubleshooting & FAQs

Resolutions for common installation, compiler, PATH, and GPU driver issues.

NextViper GPU Troubleshooting Guide

This guide details common GPU issues, error messages, and their solutions in NextViper.

NextViper could not find a Vulkan physical device with compute queue capabilities (VK_QUEUE_COMPUTE_BIT).

Check if GPU drivers are installed:

terminal

   vulkaninfo --summary

On Linux, ensure libvulkan1 and Mesa Vulkan drivers (mesa-vulkan-drivers) are installed:

terminal

   sudo apt install -y libvulkan1 mesa-vulkan-drivers

For NVIDIA GPUs, ensure the proprietary NVIDIA driver (nvidia-driver-xxx) is installed and loaded.

For headless cloud containers or CI environments without physical GPUs, install Mesa Lavapipe:

terminal

   sudo apt install -y libvulkan-dev mesa-vulkan-drivers

In user code, use device: "auto" to automatically fallback to CPU when running on machines without GPUs.

The requested tensor size exceeded the available VRAM on the target GPU.

Reduce the batch size in training (batch_size: 32 or batch_size: 16).

Free unused GPU tensors by reassigning references or calling tensor.to("cpu").

Check available GPU memory:

main.nv

   import tensor
   print("Active device:", tensor.device_name())

Inner dimensions of matrices do not match for matrix multiplication $A_{M imes K} imes B_{K imes N}$.

Ensure a.shape[1] == b.shape[0]. Use a.T() or a.reshape(...) to align dimensions before matrix multiplication.