Search Results for "gpu speed"

...The dashboard collects and displays a wide range of performance metrics including temperature, memory usage, power consumption, clock speeds, fan speed, and active processes. It can scale from monitoring a single GPU workstation to large distributed environments with dozens or even hundreds of GPUs by running lightweight containers on each node and aggregating the data centrally.

PowerInfer is a high-performance inference engine designed to run large language models efficiently on personal computers equipped with consumer-grade GPUs. The project focuses on improving the performance of local AI inference by optimizing how neural network computations are distributed between CPU and GPU resources. Its architecture exploits the observation that only a subset of neurons in large models are frequently activated, allowing the system to preload frequently used neurons into GPU memory while processing less common activations on the CPU. This hybrid execution strategy significantly reduces memory bottlenecks and improves overall inference speed.

llmfit is a terminal-based utility that helps developers determine which large language models can realistically run on their local hardware by analyzing system resources and model requirements. The tool automatically detects CPU, RAM, GPU, and VRAM specifications, then ranks available models based on performance factors such as speed, quality, and memory fit. It provides both an interactive terminal user interface and a traditional CLI mode, enabling flexible workflows for different user preferences. llmfit also supports advanced configurations including multi-GPU setups, mixture-of-experts architectures, and dynamic quantization recommendations. ...

...The repository also contains extensive learning notes that summarize CUDA programming concepts, GPU architecture details, and performance engineering strategies.

LightLLM is a high-performance inference and serving framework designed specifically for large language models, focusing on lightweight architecture, scalability, and efficient deployment. The framework enables developers to run and serve modern language models with significantly improved speed and resource efficiency compared to many traditional inference systems. Built primarily in Python, the project integrates optimization techniques and ideas from several leading open-source implementations, including FasterTransformer, vLLM, and FlashAttention, to accelerate token generation and reduce latency. LightLLM is designed to handle large-scale model workloads in production environments, supporting efficient batching and GPU utilization for fast inference across multiple requests. ...

ChatGLM2-6B is the second-gen Chinese-English conversational LLM from ZhipuAI/Tsinghua. It upgrades the base model with GLM’s hybrid pretraining objective, 1.4 TB bilingual data, and preference alignment—delivering big gains on MMLU, CEval, GSM8K, and BBH. The context window extends up to 32K (FlashAttention), and Multi-Query Attention improves speed and memory use. The repo includes Python APIs, CLI & web demos, OpenAI-style/FASTAPI servers, and quantized checkpoints for lightweight local...

Mosec is a high-performance and flexible model-serving framework for building ML model-enabled backend and microservices. It bridges the gap between any machine learning models you just trained and the efficient online service API.

...Its architecture centers on a high-performance transfer engine that provides unified data transfer across different storage and networking technologies. This engine enables efficient movement of tensors and model data across heterogeneous environments such as GPU memory, system memory, and distributed storage systems. Mooncake also introduces distributed key-value cache storage that allows inference systems to reuse previously computed attention states, significantly improving throughput in large-scale deployments. The system supports advanced networking technologies such as RDMA and NVMe over Fabric, enabling high-speed communication across clusters.

gpu_poor is an open-source tool designed to help developers determine whether their hardware is capable of running a specific large language model and to estimate the performance they can expect from it. The project focuses on calculating GPU memory requirements and predicted inference speed for different models, hardware configurations, and quantization strategies. By analyzing factors such as model size, context length, batch size, and GPU specifications, the system estimates how much VRAM will be required and how fast tokens can be generated during inference. The tool also provides a detailed breakdown of where GPU memory is allocated, including model weights, KV cache, activations, and other runtime overhead. ...