vultorch 0.2.0

Vulkan + PyTorch zero-copy tensor display with ImGui — one line to visualize any CUDA tensor

Vultorch

One line to visualize any CUDA tensor.

Vultorch is a Vulkan-based GPU tensor viewer for PyTorch. It opens a native window and displays CUDA tensors with zero CPU involvement — the data stays on the GPU the entire time.

Quick Start

pip install vultorch

import torch, vultorch

tensor = torch.rand(256, 256, 3, device="cuda")

win = vultorch.Window("Vultorch", 512, 512)
while win.poll():
    if not win.begin_frame(): continue
    vultorch.show(tensor)
    win.end_frame()
win.destroy()

Features

• One-line display: vultorch.show(tensor) — that's it
• GPU-GPU transfer: Vulkan external memory interop, no CPU bounce
• ImGui built in: Sliders, buttons, plots — all from Python
• 3D scene view: Map a tensor onto a lit 3D plane with orbit camera
• MSAA: 2x/4x/8x anti-aliasing for the 3D viewer

Requirements

• GPU with Vulkan support (any modern NVIDIA, AMD, or Intel GPU)
• Up-to-date GPU drivers (Vulkan runtime ships with drivers)
• CUDA (optional, for vultorch.show() with CUDA tensors)
• No Vulkan SDK needed — it's only required at build time

License

MIT