/starwhale

an MLOps/LLMOps platform

Primary LanguageJavaApache License 2.0Apache-2.0

An MLOps/LLMOps Platform

πŸš€ ️☁️ Starwhale Cloud is now open to the public, try it! πŸŽ‰πŸ»

Artifact Hub PyPI - Python Version Client/SDK UT Server UT Starwhale E2E Test Codecov Codecov

English | δΈ­ζ–‡

What is Starwhale

Starwhale is an MLOps/LLMOps platform that brings efficiency and standardization to machine learning operations. It streamlines the model development liftcycle, enabling teams to optimize their workflows around key areas like model building, evaluation, release and fine-tuning.

products

Starwhale meets diverse deployment needs with three flexible configurations:

  • πŸ₯ Standalone - Deployed in a local development environment, managed by the swcli command-line tool, meeting development and debugging needs.
  • πŸ¦… Server - Deployed in a private data center, relying on a Kubernetes cluster, providing centralized, web-based, and secure services.
  • πŸ¦‰ Cloud - Hosted on a public cloud, with the access address https://cloud.starwhale.cn. The Starwhale team is responsible for maintenance, and no installation is required. You can start using it after registering an account.

As its core, Starwhale abstracts Model, Runtime and Dataset as first-class citizens - providing the fundamentals for streamlined operations. Starwhale further delivers tailored capabilities for common workflow scenarios including:

  • πŸ”₯ Models Evaluation - Implement robust, production-scale evaluations with minimal coding through the Python SDK.
  • 🌟 Live Demo - Interactively assess model performance through user-friendly web interfaces.
  • 🌊 LLM Fine-tuning - End-to-end toolchain from efficient fine-tuning to comparative benchmarking and publishing.

Starwhale is also an open source platform, using the Apache-2.0 license. The Starwhale framework is designed for clarity and ease of use, empowering developers to build customized MLOps features tailored to their needs.

framework

Key Concepts

🐘 Starwhale Dataset

Starwhale Dataset offers efficient data storage, loading, and visualization capabilities, making it a dedicated data management tool tailored for the field of machine learning and deep learning

dataset overview

import torch
from starwhale import dataset, Image

# build dataset for starwhale cloud instance
with dataset("https://cloud.starwhale.cn/project/starwhale:public/dataset/test-image", create="empty") as ds:
    for i in range(100):
        ds.append({"image": Image(f"{i}.png"), "label": i})
    ds.commit()

# load dataset
ds = dataset("https://cloud.starwhale.cn/project/starwhale:public/dataset/test-image")
print(len(ds))
print(ds[0].features.image.to_pil())
print(ds[0].features.label)

torch_ds = ds.to_pytorch()
torch_loader = torch.utils.data.DataLoader(torch_ds, batch_size=5)
print(next(iter(torch_loader)))

πŸ‡ Starwhale Model

Starwhale Model is a standard format for packaging machine learning models that can be used for various purposes, like model fine-tuning, model evaluation, and online serving. A Starwhale Model contains the model file, inference codes, configuration files, and any other files required to run the model.

overview

# model build
swcli model build . --module mnist.evaluate --runtime pytorch/version/v1 --name mnist

# model copy from standalone to cloud
swcli model cp mnist https://cloud.starwhale.cn/project/starwhale:public

# model run
swcli model run --uri mnist --runtime pytorch --dataset mnist
swcli model run --workdir . --module mnist.evaluator --handler mnist.evaluator:MNISTInference.cmp

🐌 Starwhale Runtime

Starwhale Runtime aims to provide a reproducible and sharable running environment for python programs. You can easily share your working environment with your teammates or outsiders, and vice versa. Furthermore, you can run your programs on Starwhale Server or Starwhale Cloud without bothering with the dependencies.

overview

# build from runtime.yaml, conda env, docker image or shell
swcli runtime build --yaml runtime.yaml
swcli runtime build --conda pytorch --name pytorch-runtime --cuda 11.4
swcli runtime build --docker pytorch/pytorch:1.9.0-cuda11.1-cudnn8-runtime
swcli runtime build --shell --name pytorch-runtime

# runtime activate
swcli runtime activate pytorch

# integrated with model and dataset
swcli model run --uri test --runtime pytorch
swcli model build . --runtime pytorch
swcli dataset build --runtime pytorch

πŸ„ Starwhale Evaluation

Starwhale Evaluation enables users to evaluate sophisticated, production-ready distributed models by writing just a few lines of code with Starwhale Python SDK.

import typing as t
import gradio
from starwhale import evaluation
from starwhale.api.service import api

def model_generate(image):
    ...
    return predict_value, probability_matrix

@evaluation.predict(
    resources={"nvidia.com/gpu": 1},
    replicas=4,
)
def predict_image(data: dict, external: dict) -> None:
    return model_generate(data["image"])

@evaluation.evaluate(use_predict_auto_log=True, needs=[predict_image])
def evaluate_results(predict_result_iter: t.Iterator):
    for _data in predict_result_iter:
        ...
    evaluation.log_summary({"accuracy": 0.95, "benchmark": "test"})

@api(gradio.File(), gradio.Label())
def predict_view(file: t.Any) -> t.Any:
    with open(file.name, "rb") as f:
        data = Image(f.read(), shape=(28, 28, 1))
    _, prob = predict_image({"image": data})
    return {i: p for i, p in enumerate(prob)}

🦍 Starwhale Fine-tuning

Starwhale Fine-tuning provides a full workflow for Large Language Model(LLM) tuning, including batch model evaluation, live demo and model release capabilities. Starwhale Fine-tuning Python SDK is very simple.

import typing as t
from starwhale import finetune, Dataset
from transformers import Trainer

@finetune(
    resources={"nvidia.com/gpu":4, "memory": "32G"},
    require_train_datasets=True,
    require_validation_datasets=True,
    model_modules=["evaluation", "finetune"],
)
def lora_finetune(train_datasets: t.List[Dataset], val_datasets: t.List[Dataset]) -> None:
    # init model and tokenizer
    trainer = Trainer(
        model=model, tokenizer=tokenizer,
        train_dataset=train_datasets[0].to_pytorch(), # convert Starwhale Dataset into Pytorch Dataset
        eval_dataset=val_datasets[0].to_pytorch())
    trainer.train()
    trainer.save_state()
    trainer.save_model()
    # save weights, then Starwhale SDK will package them into Starwhale Model

Installation

πŸ‰ Starwhale Standalone

Requirements: Python 3.7~3.11 in the Linux or macOS os.

python3 -m pip install starwhale

πŸ₯­ Starwhale Server

Starwhale Server is delivered as a Docker image, which can be run with Docker directly or deployed to a Kubernetes cluster. For the laptop environment, using swcli server start command is a appropriate choice that depends on Docker and Docker-Compose.

swcli server start

Quick Tour

We use MNIST as the hello world example to show the basic Starwhale Model workflow.

πŸͺ… MNIST Evaluation in Starwhale Standalone

πŸͺ† MNIST Evaluation in Starwhale Server

Examples

Documentation, Community, and Support

Contributing

πŸŒΌπŸ‘PRs are always welcomed πŸ‘πŸΊ. See Contribution to Starwhale for more details.

License

Starwhale is licensed under the Apache License 2.0.