Infinity is a high-throughput, low-latency REST API for serving text-embeddings, reranking models, clip, clap and colpali. Infinity is developed under MIT License.
- Deploy any model from HuggingFace: deploy any embedding, reranking, clip and sentence-transformer model from HuggingFace
- Fast inference backends: The inference server is built on top of PyTorch, optimum (ONNX/TensorRT) and CTranslate2, using FlashAttention to get the most out of your NVIDIA CUDA, AMD ROCM, CPU, AWS INF2 or APPLE MPS accelerator. Infinity uses dynamic batching and tokenization dedicated in worker threads.
- Multi-modal and multi-model: Mix-and-match multiple models. Infinity orchestrates them.
- Tested implementation: Unit and end-to-end tested. Embeddings via infinity are correctly embedded. Lets API users create embeddings till infinity and beyond.
- Easy to use: Built on FastAPI. Infinity CLI v2 allows launching of all arguments via Environment variable or argument. OpenAPI aligned to OpenAI's API specs. View the docs at https://michaelfeil.github.io/infinity on how to get started.
- [2024/11] AMD, CPU, ONNX docker images
- [2024/10]
pip install infinity_client
- [2024/07] Inference deployment example via Modal and a free GPU deployment
- [2024/06] Support for multi-modal: clip, text-classification & launch all arguments from env variables
- [2024/05] launch multiple models using the
v2
cli, including--api-key
- [2024/03] infinity supports experimental int8 (cpu/cuda) and fp8 (H100/MI300) support
- [2024/03] Docs are online: https://michaelfeil.github.io/infinity/latest/
- [2024/02] Community meetup at the Run:AI Infra Club
- [2024/01] TensorRT / ONNX inference
- [2023/10] Initial release
pip install infinity-emb[all]
After your pip install, with your venv active, you can run the CLI directly.
infinity_emb v2 --model-id BAAI/bge-small-en-v1.5
Check the v2 --help
command to get a description for all parameters.
infinity_emb v2 --help
Instead of installing the CLI via pip, you may also use docker to run michaelf34/infinity
.
Make sure you mount your accelerator ( i.e. install nvidia-docker
and activate with --gpus all
).
port=7997
model1=michaelfeil/bge-small-en-v1.5
model2=mixedbread-ai/mxbai-rerank-xsmall-v1
volume=$PWD/data
docker run -it --gpus all \
-v $volume:/app/.cache \
-p $port:$port \
michaelf34/infinity:latest \
v2 \
--model-id $model1 \
--model-id $model2 \
--port $port
The cache path inside the docker container is set by the environment variable HF_HOME
.
Docker container for CPU
Use the `latest-cpu` image or `x.x.x-cpu` for slimer image. Run like any other cpu-only docker image. Optimum/Onnx is often the prefered engine.docker run -it \
-v $volume:/app/.cache \
-p $port:$port \
michaelf34/infinity:latest-cpu \
v2 \
--engine optimum \
--model-id $model1 \
--model-id $model2 \
--port $port
Docker Container for ROCm (MI200 Series and MI300 Series)
Use the `latest-rocm` image or `x.x.x-rocm` for rocm compatible inference. **This image is currently not build via CI/CD (to large), consider pinning to exact version.** Make sure you have ROCm is correctly installed and ready to use with Docker.Visit Docs for more info.
Docker Container for Onnx-GPU, Cuda Extensions, TensorRT
Use the `latest-trt-onnx` image or `x.x.x-trt-onnx` for nvidia compatible inference. **This image is currently not build via CI/CD (to large), consider pinning to exact version.**This image has support for:
- ONNX-Cuda "CudaExecutionProvider"
- ONNX-TensorRT "TensorRTExecutionProvider" (may not always work due to version mismatch with ORT)
- CudaExtensions and packages, e.g. Tri-Dao's
pip install flash-attn
package when using Pytorch. - nvcc compiler support
docker run -it \
-v $volume:/app/.cache \
-p $port:$port \
michaelf34/infinity:latest-trt-onnx \
v2 \
--engine optimum \
--device cuda \
--model-id $model1 \
--port $port
Launching multiple models at once
Since infinity_emb>=0.0.34
, you can use cli v2
method to launch multiple models at the same time.
Checkout infinity_emb v2 --help
for all args and validation.
Multiple Model CLI Playbook:
-
- cli options can be repeated e.g.
v2 --model-id model/id1 --model-id/id2 --batch-size 8 --batch-size 4
. This will create two modelsmodel/id1
andmodel/id2
- cli options can be repeated e.g.
-
- or adapt the defaults by setting ENV Variables separated by
;
:INFINITY_MODEL_ID="model/id1;model/id2;" && INFINITY_BATCH_SIZE="8;4;"
- or adapt the defaults by setting ENV Variables separated by
-
- single items are broadcasted to
--model-id
length,v2 --model-id model/id1 --model-id/id2 --batch-size 8
making both models have batch-size 8.
- single items are broadcasted to
-
- Everything is broadcasted to the number of
--model-id
+ API requests are routed to the--served-model-name/--model-id
- Everything is broadcasted to the number of
Using environment variables instead of the cli
All CLI arguments are also launchable via environment variables.Environment variables start with INFINITY_{UPPER_CASE_SNAKE_CASE}
and often match the --{lower-case-kebab-case}
cli arguments.
The following two are equivalent:
- CLI
infinity_emb v2 --model-id BAAI/bge-base-en-v1.5
- ENV-CLI:
export INFINITY_MODEL_ID="BAAI/bge-base-en-v1.5" && infinity_emb v2
Multiple arguments can be used via ;
syntax: INFINITY_MODEL_ID="model/id1;model/id2;"
API Key
Supply an `--api-key secret123` via CLI or ENV INFINITY_API_KEY="secret123".Chosing the fastest engine
With the command --engine torch
the model must be compatible with https://github.com/UKPLab/sentence-transformers/ and AutoModel
With the command --engine optimum
, there must be an onnx file. Models from https://huggingface.co/Xenova are recommended.
With the command --engine ctranslate2
- only BERT
models are supported.
Telemetry opt-out
See which telemetry is collected: https://michaelfeil.eu/infinity/main/telemetry/
# Disable
export INFINITY_ANONYMOUS_USAGE_STATS="0"
Infinity aims to be the inference server supporting most functionality for embeddings, reranking and related RAG tasks. The following Infinity tests 15+ architectures and all of the below cases in the Github CI. Click on the sections below to find tasks and validated example models.
Text Embeddings
Text embeddings measure the relatedness of text strings. Embeddings are used for search, clustering, recommendations. Think about a private deployed version of openai's text embeddings. https://platform.openai.com/docs/guides/embeddings
Tested embedding models:
- mixedbread-ai/mxbai-embed-large-v1
- WhereIsAI/UAE-Large-V1
- BAAI/bge-base-en-v1.5
- Alibaba-NLP/gte-large-en-v1.5
- jinaai/jina-embeddings-v2-base-code
- sentence-transformers/all-MiniLM-L6-v2
- intfloat/multilingual-e5-large-instruct
- intfloat/multilingual-e5-small
- jinaai/jina-embeddings-v3
- BAAI/bge-m3, no sparse
- decoder-based models. Keep in mind that they are ~20-100x larger (&slower) than bert-small models:
Other models:
- Most embedding model are likely supported: https://huggingface.co/models?pipeline_tag=feature-extraction&other=text-embeddings-inference&sort=trending
- Check MTEB leaderboard for models https://huggingface.co/spaces/mteb/leaderboard.
Reranking
Given a query and a list of documents, Reranking indexes the documents from most to least semantically relevant to the query. Think like a locally deployed version of https://docs.cohere.com/reference/rerankTested reranking models:
- mixedbread-ai/mxbai-rerank-xsmall-v1
- Alibaba-NLP/gte-multilingual-reranker-base
- BAAI/bge-reranker-base
- BAAI/bge-reranker-large
- BAAI/bge-reranker-v2-m3
- jinaai/jina-reranker-v1-turbo-en
Other reranking models:
- Reranking Models supported by infinity are bert-style classification Models with one category.
- Most reranking model are likely supported: https://huggingface.co/models?pipeline_tag=text-classification&other=text-embeddings-inference&sort=trending
- https://huggingface.co/models?pipeline_tag=text-classification&sort=trending&search=rerank
Multi-modal and cross-modal - image and audio embeddings
Specialized embedding models that allow for image<->text or image<->audio search. Typically, these models allow for text<->text, text<->other and other<->other search, with accuracy tradeoffs when going cross-modal.Image<->text models can be used for e.g. photo-gallery search, where users can type in keywords to find photos, or use a photo to find related images. Audio<->text models are less popular, and can be e.g. used to find music songs based on a text description or related music songs.
Tested image<->text models:
- wkcn/TinyCLIP-ViT-8M-16-Text-3M-YFCC15M
- jinaai/jina-clip-v1
- google/siglip-so400m-patch14-384
- Models of type: ClipModel / SiglipModel in
config.json
Tested audio<->text models:
- Clap Models from LAION
- limited number open source organizations training these models
-
- Note: The sampling rate of the audio data needs to match the model *
Not supported:
- Plain vision models e.g. nomic-ai/nomic-embed-vision-v1.5
ColBert-style late-interaction Embeddings
ColBert Embeddings don't perform any special Pooling methods, but return the raw **token embeddings**. The **token embeddings** are then to be scored with the MaxSim Metric in a VectorDB (Qdrant / Vespa)For usage via the RestAPI, late-interaction embeddings may best be transported via base64
encoding.
Example notebook: https://colab.research.google.com/drive/14FqLc0N_z92_VgL_zygWV5pJZkaskyk7?usp=sharing
Tested colbert models:
ColPali-style late-interaction Image<->Text Embeddings
Similar usage to ColBert, but scanning over an image<->text instead of only text.For usage via the RestAPI, late-interaction embeddings may best be transported via base64
encoding.
Example notebook: https://colab.research.google.com/drive/14FqLc0N_z92_VgL_zygWV5pJZkaskyk7?usp=sharing
Tested ColPali/ColQwen models:
- vidore/colpali-v1.2-merged
- michaelfeil/colqwen2-v0.1
- No lora adapters supported, only "merged" models.
Text classification
A bert-style multi-label text classification. Classifies it into distinct categories.Tested models:
- ProsusAI/finbert, financial news classification
- SamLowe/roberta-base-go_emotions, text to emotion categories.
- bert-style text-classifcation models with more than >1 label in
config.json
Instead of the cli & RestAPI use infinity's interface via the Python API.
This gives you most flexibility. The Python API builds on asyncio
with its await/async
features, to allow concurrent processing of requests. Arguments of the CLI are also available via Python.
import asyncio
from infinity_emb import AsyncEngineArray, EngineArgs, AsyncEmbeddingEngine
sentences = ["Embed this is sentence via Infinity.", "Paris is in France."]
array = AsyncEngineArray.from_args([
EngineArgs(model_name_or_path = "BAAI/bge-small-en-v1.5", engine="torch", embedding_dtype="float32", dtype="auto")
])
async def embed_text(engine: AsyncEmbeddingEngine):
async with engine:
embeddings, usage = await engine.embed(sentences=sentences)
# or handle the async start / stop yourself.
await engine.astart()
embeddings, usage = await engine.embed(sentences=sentences)
await engine.astop()
asyncio.run(embed_text(array[0]))
Reranking gives you a score for similarity between a query and multiple documents. Use it in conjunction with a VectorDB+Embeddings, or as standalone for small amount of documents. Please select a model from huggingface that is a AutoModelForSequenceClassification compatible model with one class classification.
import asyncio
from infinity_emb import AsyncEngineArray, EngineArgs, AsyncEmbeddingEngine
query = "What is the python package infinity_emb?"
docs = ["This is a document not related to the python package infinity_emb, hence...",
"Paris is in France!",
"infinity_emb is a package for sentence embeddings and rerankings using transformer models in Python!"]
array = AsyncEmbeddingEngine.from_args(
[EngineArgs(model_name_or_path = "mixedbread-ai/mxbai-rerank-xsmall-v1", engine="torch")]
)
async def rerank(engine: AsyncEmbeddingEngine):
async with engine:
ranking, usage = await engine.rerank(query=query, docs=docs)
print(list(zip(ranking, docs)))
# or handle the async start / stop yourself.
await engine.astart()
ranking, usage = await engine.rerank(query=query, docs=docs)
await engine.astop()
asyncio.run(rerank(array[0]))
When using the CLI, use this command to launch rerankers:
infinity_emb v2 --model-id mixedbread-ai/mxbai-rerank-xsmall-v1
CLIP models are able to encode images and text at the same time.
import asyncio
from infinity_emb import AsyncEngineArray, EngineArgs, AsyncEmbeddingEngine
sentences = ["This is awesome.", "I am bored."]
images = ["http://images.cocodataset.org/val2017/000000039769.jpg"]
engine_args = EngineArgs(
model_name_or_path = "wkcn/TinyCLIP-ViT-8M-16-Text-3M-YFCC15M",
engine="torch"
)
array = AsyncEngineArray.from_args([engine_args])
async def embed(engine: AsyncEmbeddingEngine):
await engine.astart()
embeddings, usage = await engine.embed(sentences=sentences)
embeddings_image, _ = await engine.image_embed(images=images)
await engine.astop()
asyncio.run(embed(array["wkcn/TinyCLIP-ViT-8M-16-Text-3M-YFCC15M"]))
CLAP models are able to encode audio and text at the same time.
import asyncio
from infinity_emb import AsyncEngineArray, EngineArgs, AsyncEmbeddingEngine
import requests
import soundfile as sf
import io
sentences = ["This is awesome.", "I am bored."]
url = "https://bigsoundbank.com/UPLOAD/wav/2380.wav"
raw_bytes = requests.get(url, stream=True).content
audios = [raw_bytes]
engine_args = EngineArgs(
model_name_or_path = "laion/clap-htsat-unfused",
dtype="float32",
engine="torch"
)
array = AsyncEngineArray.from_args([engine_args])
async def embed(engine: AsyncEmbeddingEngine):
await engine.astart()
embeddings, usage = await engine.embed(sentences=sentences)
embedding_audios = await engine.audio_embed(audios=audios)
await engine.astop()
asyncio.run(embed(array["laion/clap-htsat-unfused"]))
Use text classification with Infinity's classify
feature, which allows for sentiment analysis, emotion detection, and more classification tasks.
import asyncio
from infinity_emb import AsyncEngineArray, EngineArgs, AsyncEmbeddingEngine
sentences = ["This is awesome.", "I am bored."]
engine_args = EngineArgs(
model_name_or_path = "SamLowe/roberta-base-go_emotions",
engine="torch", model_warmup=True)
array = AsyncEngineArray.from_args([engine_args])
async def classifier(engine: AsyncEmbeddingEngine):
async with engine:
predictions, usage = await engine.classify(sentences=sentences)
# or handle the async start / stop yourself.
await engine.astart()
predictions, usage = await engine.classify(sentences=sentences)
await engine.astop()
asyncio.run(classifier(array["SamLowe/roberta-base-go_emotions"]))
Infinity has a generated client code for RestAPI client side usage.
If you want to call a remote infinity instance via RestAPI, install the following package locally:
pip install infinity_client
For more information, check out the Client Readme https://github.com/michaelfeil/infinity/tree/main/libs/client_infinity/infinity_client
- Serverless deployments at Runpod
- Truefoundry Cognita
- Langchain example
- imitater - A unified language model server built upon vllm and infinity.
- Dwarves Foundation: Deployment examples using Modal.com
- infiniflow/Ragflow
- SAP Core AI
- gpt_server - gpt_server is an open-source framework designed for production-level deployment of LLMs (Large Language Models) or Embeddings.
- KubeAI: Kubernetes AI Operator for inferencing
- LangChain
- Batched, modification of the Batching algoritm in Infinity
View the docs at https:///michaelfeil.github.io/infinity on how to get started.
After startup, the Swagger Ui will be available under {url}:{port}/docs
, in this case http://localhost:7997/docs
. You can also find a interactive preview here: https://infinity.modal.michaelfeil.eu/docs (and https://michaelfeil-infinity.hf.space/docs)
Install via Poetry 1.8.1, Python3.11 on Ubuntu 22.04
cd libs/infinity_emb
poetry install --extras all --with lint,test
To pass the CI:
cd libs/infinity_emb
make precommit
All contributions must be made in a way to be compatible with the MIT License of this repo.
@software{feil_2023_11630143,
author = {Feil, Michael},
title = {Infinity - To Embeddings and Beyond},
month = oct,
year = 2023,
publisher = {Zenodo},
doi = {10.5281/zenodo.11630143},
url = {https://doi.org/10.5281/zenodo.11630143}
}