Your Solution to Distributed Model Training
Increase iteration speed—without increasing cost. Anyscale makes it easy to scale from a single machine to a cloud cluster with just one line of code.
Why Anyscale?
Reduce Costs with Spot Instances
With elastic training on Anyscale, train with minimal interruption from spot instance preemption and node failure—while reducing costs by up to 90%.
Faster Iteration, Same Cost
Get the same results—faster. Train with parallelized compute to complete training in minutes, rather than hours. Increase iteration speed with the ability to scale across nodes during development.
Improve Model Quality By Training on All Your Data
Train higher quality models by training on all of your data—not just a subset.
One Unified API
Easily scale your training for any machine learning library—from XGBoost to Tensorflow to PyTorch and more.
How DoorDash Built Reliable Delivery Forecasting with Ray
Balancing supply and demand is crucial for marketplace app health—and DoorDash relies on Ray to get it done. See how Doordash achieved 46% reduction in training time and 42% reduction in cost with Ray.
Easily Get Started with Distributed Training at Scale
Elastic Training & Spot Instance Support
Job Retries & Fault Tolerance Support
Fast Node Launching and Autoscaling
Fractional Heterogeneous Resource Allocation
Detailed Training Dashboard
Last-Mile Data Preprocessing
Autoscaling Development Environment
Distributed Debugger
Data Integrations (Databricks, Snowflake, S3, GCS, etc)
Framework Support (Pytorch, Huggingface, Tensorflow, XGBoost, etc)
Experiment Tracking Integrations (Weights and Biases, MLflow, etc)
Orchestration Integrations (Prefect, Apache Airflow, etc)
Alerting
Resumable Jobs
Priority Scheduling
Job Queues
EFA Support
| | | |
|---|---|---|---|
Elastic Training & Spot Instance Support | - | - | |
Job Retries & Fault Tolerance Support | | - | |
Fast Node Launching and Autoscaling | - | - | 60 sec |
Fractional Heterogeneous Resource Allocation | - | | |
Detailed Training Dashboard | - | - | |
Last-Mile Data Preprocessing | - | | |
Autoscaling Development Environment | - | - | |
Distributed Debugger | - | - | |
Data Integrations (Databricks, Snowflake, S3, GCS, etc) | | | |
Framework Support (Pytorch, Huggingface, Tensorflow, XGBoost, etc) | | | |
Experiment Tracking Integrations (Weights and Biases, MLflow, etc) | | | |
Orchestration Integrations (Prefect, Apache Airflow, etc) | | | |
Alerting | | - | |
Resumable Jobs | | | |
Priority Scheduling | - | | |
Job Queues | | - | |
EFA Support | | Custom | |
Set-it-and-Forget-it Model Training
With built-in fault tolerance and automatic job retries, Anyscale will ensure your training job completes regardless of any errors. Easily recover from system failures and resume training from a recent checkpoint.
Training Utilization Dashboard
Gain insight into your distributed training job progress and track utilization to ensure you’re getting the most out of your compute resources.
Reduce Costs By Using GPUs and CPUs
Anyscale makes it easy to leverage heterogeneous compute. Use CPUs and GPUs in the same pipeline to increase utilization, fully saturate GPUs, and decrease costs.
Out-of-the-Box Templates & App Accelerators
Jumpstart your development process with custom-made templates, only available on Anyscale.
End-to-End LLM Workflows
Execute end-to-end LLM workflows to develop and productionize LLMs at scale
Fine-Tune Stable Diffusion
Fine-tune a personalized Stable Diffusion XL model with Ray Train
Pre-Train Stable Diffusion
Pre-train a Stable Diffusion V2 model with Ray Train and Ray Data
“We have no ceiling on scale, and an incredible opportunity to bring AI features and value to our 170 million users.”
Greg Roodt
ML Lead, Canva
FAQs
Anyscale's Ray Train integrates with your preferred training frameworks, including PyTorch, Hugging Face, Tensorflow, and more.
Distributed AI Model Training at Scale
Enable simple, fast, and affordable distributed model training with Anyscale. Learn more, or get started today.