# turboderp/exllama

A more memory-efficient rewrite of the HF transformers implementation of Llama for use with quantized weights.

Repository: https://github.com/turboderp/exllama
Canonical: https://ross.abutalabs.com/products/exllama
Language: Python
License: MIT
License Family: permissive
Last push: 2023-09-30T19:06:04+00:00

## Health v2 (maintenance only)
Score: 29/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 0, release rhythm 35, longevity 86
- inputs: {"age_days": 1217, "days_push": 1068, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2936, forks 220 (observed 2026-08-28T04:07:30.848930+00:00)

## What it is
ExLlama is a standalone Python/C++/CUDA implementation of the Llama architecture optimized for running 4-bit GPTQ quantized weights on modern NVIDIA GPUs. It serves as a memory-efficient, fast alternative to the Hugging Face transformers implementation for local LLM inference.

## Use cases
- run quantized llama models on a consumer gpu
- load 4-bit gptq models with low vram usage
- fast local inference of llama on rtx 3090 or 4090
- build a local chatbot with a quantized language model
- benchmark inference speed of quantized llama models
- integrate llama inference into a python application

## When to choose
- you have a modern NVIDIA GPU (RTX 30-series or newer) and want memory-efficient 4-bit GPTQ inference
- you need faster generation than HF transformers with quantized Llama weights
- you want a lightweight standalone library rather than a full inference framework

## When to avoid
- you have an older GPU with poor FP16 support (Pascal or earlier) - AutoGPTQ or GPTQ-for-LLaMa are better fits
- you need AMD/ROCm support that is well tested and optimized
- you need a production-hardened, stable solution - the project is still a work in progress

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning, deep-learning
- domain: large-language-models, machine-learning, gpu-computing
- platform: python, windows, cross-platform
- tags: gptq, quantization, llama, 4-bit, cuda, inference-optimization, gpu, linux

## Member repositories
- turboderp/exllama (main) score 29

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:30.848930+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T07:33:09.896887+00:00, confidence not recorded.
  - readme: https://github.com/turboderp/exllama (fetched 2026-08-28T04:07:30.848930+00:00, sha 54a077b845b8)
- Data as of 2026-08-30T08:39:29.467469+00:00.