vultorch 0.3.0

Real-time Torch visualization window with Vulkan zero-copy — built for neural textures and GPU-intensive research

🔥 Vultorch

Real-time Torch Visualization Window · Vulkan Zero-Copy

Visualize CUDA tensors at GPU speed — zero CPU readback, zero staging buffers.

🇨🇳 中文 · 🌐 Website

---

Overview

Vultorch displays CUDA tensors in a native window by sharing GPU memory between PyTorch and Vulkan. Data stays on the GPU — there is no tensor.cpu() readback and no staging buffer copy.

vultorch.show(tensor)   # zero-copy, submillisecond

Key Features

• Zero-copy display — Vulkan external memory interop, data never leaves the GPU
• True shared memory — vultorch.create_tensor() returns a torch.Tensor backed by Vulkan memory (DLPack)
• One-line API — vultorch.show(tensor) handles format conversion, upload, and display
• Built-in ImGui — Sliders, buttons, color pickers, plots, docking layout — all from Python
• 3D scene view — Map textures onto lit 3D planes with orbit camera, MSAA, Blinn-Phong shading
• Docking windows — Drag-and-drop window arrangement (ImGui docking branch)

Quick Start

pip install vultorch

import torch, vultorch
from vultorch import ui

# Your neural texture output (or any CUDA tensor)
texture = torch.rand(512, 512, 4, device="cuda")

win = vultorch.Window("Neural Texture Viewer", 800, 600)
while win.poll():
    if not win.begin_frame(): continue
    ui.begin("Output")
    vultorch.show(texture)  # zero-copy GPU → screen
    ui.end()
    win.end_frame()
win.destroy()

True Zero-Copy

# Shared GPU memory — writes are instantly visible on screen
tensor = vultorch.create_tensor(512, 512, channels=4)
tensor[:] = model(input)   # write directly, no copy needed

3D Scene

scene = vultorch.SceneView("3D", 800, 600, msaa=4)
scene.set_tensor(texture)
scene.render()  # orbit camera, Blinn-Phong lighting

Examples

Example	Description
01_hello_tensor.py	Minimal tensor display
02_imgui_controls.py	ImGui widgets: sliders, plots, colors
03_scene_3d.py	3D scene with lighting + orbit camera
04_docking_layout.py	Drag-and-drop dockable window layout
05_zero_copy.py	True zero-copy shared tensor

python examples/01_hello_tensor.py

Building from Source

Prerequisites

Component	Required	Notes
GPU	✅	Any Vulkan-capable GPU (NVIDIA, AMD, Intel)
Vulkan	Runtime	Ships with your GPU driver — no separate install needed
Vulkan SDK	Build only	lunarg.com/vulkan-sdk — only for building from source
CUDA Toolkit	Optional	For show() and create_tensor()
Python 3.8+	✅
CMake 3.25+	Build only	+ Ninja

Clone & Build

git clone --recursive https://github.com/ChenlizheMe/Vultorch.git
cd vultorch

One command — configure, compile, and produce a wheel in dist/:

# Windows
cmake --preset release-windows
cmake --build --preset release-windows

# Linux
cmake --preset release-linux
cmake --build --preset release-linux

The wheel appears in dist/. Install it:

pip install dist/vultorch-*.whl

Or use the convenience scripts:

# Windows
build.bat

# Linux / macOS
./build.sh

Architecture

vultorch/
├── src/                    # C++ core
│   ├── engine.cpp/h        # Vulkan + SDL3 + ImGui engine
│   ├── tensor_texture.*    # CUDA ↔ Vulkan zero-copy interop
│   ├── scene_renderer.*    # Offscreen 3D renderer (MSAA, Blinn-Phong)
│   ├── bindings.cpp        # pybind11 Python bindings
│   └── shaders/            # GLSL shaders → SPIR-V
├── vultorch/               # Python package
│   └── __init__.py         # High-level API (Window, show, SceneView)
├── external/               # Git submodules
│   ├── pybind11/           # C++ ↔ Python binding
│   ├── SDL/                # Window / input (SDL3)
│   └── imgui/              # Dear ImGui (docking branch)
├── examples/               # Ready-to-run demos
├── tools/                  # Build utilities
└── docs/                   # GitHub Pages website

License

MIT

---

Examples · Website · 中文文档