ExLlamaV2 vs Vast.ai — Comparison

ExLlamaV2

A fast inference library for running LLMs locally on modern consumer-class GPUs - turboderp-org/exllamav2

There was an error while loading. Please reload this page. ExLlamaV2 is an inference library for running local LLMs on modern consumer GPUs. The official and recommended backend server for ExLlamaV2 is TabbyAPI, which provides an OpenAI-compatible API for local or remote inference, with extended features like HF model downloading, embedding model support and support for HF Jinja2 chat templates. See the wiki for help getting started. The dynamic generator supports all inference, sampling and speculative decoding features of the previous two generators, consolidated into one API (with the exception of FP8 cache, though the Q4 cache mode is supported and performs better anyway, see here.) The generator is explained in detail here. See the full, updated examples here. Some quick tests to compare performance with ExLlama V1. There may be more performance optimizations in the future, and speeds will vary across GPUs, with slow CPUs still being a potential bottleneck: To install from the repo you'll need the CUDA Toolkit and either gcc on Linux or (Build Tools for) Visual Studio on Windows). Also make sure you have an appropriate version of PyTorch, then run: A simple console chatbot is included. Run it with: To install the current dev version, clone the repo and run the setup script: This will install the "JIT version" of the package, i.e. it will install the Python components without building the C++ extension in the process. Instead, the extension will be built the first time the library is used, then cached in ~/.cache/torch_extensions for subsequent use. Either download an appropriate wheel or install directly from the appropriate URL: A PyPI package is available as well. This is the same as the JIT version (see above). It can be installed with: ExLlamaV2 supports the same 4-bit GPTQ models as V1, but also a new "EXL2" format. EXL2 is based on the same optimization method as GPTQ and supports 2, 3, 4, 5, 6 and 8-bit quantization. The format allows for mixing quantization levels within a model to achieve any average bitrate between 2 and 8 bits per weight. Moreover, it's possible to apply multiple quantization levels to each linear layer, producing something akin to sparse quantization wherein more important weights (columns) are quantized with more bits. The same remapping trick that lets ExLlama work efficiently with act-order models allows this mixing of formats to happen with little to no impact on performance. Parameter selection is done automatically by quantizing each matrix multiple times, measuring the quantization error (with respect to the chosen calibration data) for each of a number of possible settings, per layer. Finally, a combination is chosen that minimizes the maximum quantization error over the entire model while meeting a target average bitrate. In my tests, this scheme allows Llama2 70B to run on a single 24 GB GPU with a 2048-token context, producing coherent and mostly stable output with 2.55 bits per weight

Vast.ai

Real-Time GPU Pricing

Vast.ai is a GPU compute marketplace founded on one idea: whoever controls compute controls AI. We exist to make sure that power stays distributed. Christian Horne — a fellow thinker and builder who also published on LessWrong — shared Jake's view that the compute scaling thesis had profound implications, not just for AI development, but for who would control it. Both saw the same thing: if whoever controlled the most compute controlled the most powerful AI, then the future of artificial general intelligence would be determined by who had the deepest pockets, not who had the best ideas. On June 28, 2016, they incorporated Vast.ai. The founding thesis fit on a napkin: the world was full of underutilized GPU hardware — in gaming rigs, mining farms, research labs, and small data centers — and the people who needed that compute most couldn't afford the hyperscaler rates. But the motivation was never purely commercial. A world where compute flows freely to thousands of independent researchers is a fundamentally different world than one where it is locked behind the pricing walls of a few incumbents. “A world where compute flows freely to thousands of independent researchers is a fundamentally different world than one where it is locked behind the pricing walls of AWS, GCP, and Azure.” What Jake predicted. What the team built. How the field caught up. Jake Cannell publishes a series of essays on LessWrong arguing that intelligence is fundamentally a function of compute — not clever algorithms or hand-engineered modules. Christian Horne (lahwran), a fellow LessWrong contributor, shares the same conviction. The two become collaborators. AlexNet breaks ImageNet benchmarks by scaling a known neural network architecture on GPUs — exactly as the scaling hypothesis predicted. The deep learning revolution begins. Jake publishes his landmark essay arguing that the human brain is a single, general-purpose learning algorithm — not a zoo of specialized circuits. He predicts AlphaGo two years before it happens and forecasts human-level vision (~2024±3) and language via scaled deep learning. Jake Cannell and Christian Horne incorporate Vast.ai as a Delaware C Corporation. The founding thesis: the world is full of underutilized GPU hardware, and the people who need that compute most can’t afford hyperscaler rates. The market needs a two-sided platform. For two years, Jake and Christian build the marketplace platform end-to-end: host onboarding, search interface, pricing engine, Docker-based instance management — engineered to work across heterogeneous hardware and wildly different network conditions. Vast.ai launches — not with a press release, but the way honest products launch: to friends, family, and a post on Hacker News. GPU compute 3–5x cheaper than AWS, available in seconds, no enterprise contract required. Early independent hosts join the platform. The marketplace concept is validated — developers get cheaper GPUs, hosts monetize idle har

ExLlamaV2

Vast.ai

ExLlamaV2

Vast.ai

ExLlamaV2

Vast.ai