Hugging Face·1yVisualize and understand GPU memory in PyTorch
This tutorial explains how to visualize and understand GPU memory usage in PyTorch during model training. It provides step-by-step instructions on generating and interpreting memory profiles using PyTorch's built-in tools. The tutorial also covers how to estimate and optimize memory requirements for training large models, offering practical tips to manage GPU memory efficiently.
Community Picks·1yUniversal, Pure-GPU HTML Renderer
Ultralight is a GPU-accelerated toolkit designed to embed modern HTML in games and native applications. It offers ultra-fast rendering either directly on the GPU or on the CPU for ease of integration. Ultralight is highly portable, with support for multiple platforms including Windows, macOS, Linux, PlayStation, Xbox, and ARM64 devices. The toolkit is built in collaboration with leading game studios and provides deep GPU integration, transparent rendering, custom image compositing, and more. It is also optimized for native app developers, offering consistent performance across platforms, automatic window management, and seamless JavaScript-native code integration.
ThePrimeTime·1ySo I Tried To Learn Shaders...
The writer shares their journey of learning shaders, starting from a point of confusion back in college to making a renewed effort with modern tools like the Book of Shaders. They explain basic concepts such as what shaders are, how they function in parallel on GPUs, and the importance of uniforms in passing consistent input data to shaders. The post highlights the challenge of understanding shader syntax and execution but encourages continuous learning through practical experimentation.
Hacker News·1yhao-ai-lab/FastVideo: FastVideo is an open-source framework for accelerating large video diffusion model.
FastVideo is an open-source framework designed to accelerate large video diffusion models. It features FastHunyuan and FastMochi for consistent video diffusion model speedup, providing 8x inference acceleration. The framework supports scalable training across multiple GPUs and offers memory-efficient finetuning options. FastVideo includes distillation recipes and techniques based on the Phased Consistency Model, with additional support for preprocessing and finetuning using various datasets.