# predibase/lorax

Multi-LoRA inference server that scales to 1000s of fine-tuned LLMs

Repository: https://github.com/predibase/lorax
Canonical: https://ross.abutalabs.com/products/lorax
Homepage: https://loraexchange.ai
Language: Python
License: Apache-2.0
License Family: permissive
Topics: fine-tuning, gpt, llama, llm, llm-inference, llm-serving, llmops, lora, model-serving, pytorch, transformers
Last push: 2026-05-28T18:12:20+00:00

## Health v2 (maintenance only)
Score: 62/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 84, release rhythm 28, longevity 74
- inputs: {"age_days": 1048, "days_push": 97, "days_rel": 597, "gap_med": 48, "n_releases_24m": 4}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 3826, forks 326 (observed 2026-08-28T04:08:22.647255+00:00)

## What it is
LoRAX is a multi-LoRA inference server that serves thousands of fine-tuned LLM adapters on a single GPU by sharing a common base model. It provides an OpenAI-compatible REST API with dynamic adapter loading, heterogeneous continuous batching, and production features like Prometheus metrics and Kubernetes Helm charts.

## Use cases
- serve thousands of fine-tuned lora models on one gpu
- host a multi-tenant llm inference server with per-request adapters
- reduce cost of serving many fine-tuned llms
- run an openai-compatible api for custom fine-tuned models
- dynamically load huggingface lora adapters per request
- deploy llm serving on kubernetes with helm
- merge multiple lora adapters into an ensemble at inference time
- serve structured  output from fine-tuned llms

## When to choose
- you need to serve many fine-tuned LoRA adapters of the same base model cost-effectively
- you want an OpenAI-compatible API with per-request adapter selection
- you need production features like metrics, tracing, and Kubernetes deployment
- you want to merge adapters per request to build ensembles

## When to avoid
- you only serve one or a few full fine-tuned models rather than LoRA adapters
- you need non-LoRA fine-tuning methods like full-weight tuning served together
- you need a simple local playground rather than a production serving server
- your base models differ per request, defeating shared-base-model efficiency

## Facets
- artifact type: service
- maturity: active
- function: llm-inference, http-server, api-framework, monitoring, tracing
- domain: large-language-models, machine-learning, apis, self-hosted, gpu-computing
- platform: python, self-hosted
- tags: lora, model-serving, openai-compatible-api, fine-tuned-models, adapter-hot-swapping, continuous-batching, quantization, helm-charts, docker, kubernetes, gpu, linux

## Member repositories
- predibase/lorax (main) score 62

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:08:22.647255+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 2026-08-29T18:26:21.836370+00:00, confidence not recorded.
  - readme: https://github.com/predibase/lorax (fetched 2026-08-28T04:08:22.647255+00:00, sha 9ce55f7eebce)
  - homepage: https://loraexchange.ai (fetched 2026-08-29T09:21:11.777160+00:00, sha 60e544500dc0)
- Data as of 2026-08-30T08:39:29.467469+00:00.