User guides

These guides show you how to complete common tasks with Ray Data. If you’re new to Ray Data, start with the Ray Data quickstart.

• Loading data
  • Read files
  • Load data from other libraries
  • Read databases
  • Read from Kafka
  • Create synthetic data
  • Load other datasources
  • Community-maintained connectors
  • Performance considerations
• Inspecting data
  • Describe datasets
  • Inspect rows
  • Inspect batches
  • Inspect execution statistics
• Transforming Data
  • Transforming rows
  • Transforming batches
  • Stateful/Class-based Transforms
  • Specifying CPUs, GPUs, and Memory
  • Specifying Concurrency
  • Ordering of rows
  • Group-by and transforming groups
  • Advanced: Distributed UDFs with Placement Groups
  • Advanced: Asynchronous Transforms
  • Expressions (Alpha)
• Aggregating data
  • Use built-in aggregations
  • Create custom aggregations
• Iterating over data
  • Iterate over rows
  • Iterate over batches
  • Iterate over batches with shuffling
  • Split datasets for distributed parallel training
• Joining data
  • Configure joins
  • Configure the number of partitions
  • Configure the number of aggregators
• Shuffling data
  • Choose a shuffling method
  • Advanced: Optimize shuffles
• Weighted dataset mixing
  • Quickstart
  • Choose a mixing strategy
  • Choose a stopping condition
  • Limitations
  • See also
• Saving data
  • Write data to files
  • Convert datasets to other Python libraries
• Working with images
  • Read images
  • Transform images
  • Perform inference on images
  • Save images
• Working with text
  • Read text files
  • Transform text
  • Perform inference on text
  • Save text
• Working with tensors and NumPy
  • How does Ray Data represent tensors?
  • Transform tensor data
  • Save tensor data
• Working with PyTorch
  • Iterate over Torch tensors for training
  • Return Torch tensors from transformations
  • Run batch inference with PyTorch
  • Save datasets that contain Torch tensors
  • Migrate from PyTorch Datasets and DataLoaders
• Working with LLMs
  • Quickstart: Run batch inference with vLLM
  • How does the processor pipeline work?
  • Scale to multiple GPUs
  • Generate text
  • Run batch inference on multimodal data
  • Generate embeddings
  • Run classification models
  • Query OpenAI-compatible endpoints
  • Configure tokenization disaggregation
  • Use custom tokenizers
  • How does Ray Data LLM handle failures?
  • Advanced configuration
  • Troubleshooting
  • Usage data collection
• How to avoid out-of-memory errors (OOMs)
  • What do OOMs look like?
  • Apply best practices to avoid OOMs
  • What should you expect after tuning?
  • Further reading
• Monitoring your workload
  • Ray Data progress bars
  • Ray Data dashboard
  • Ray Data logs
  • Ray Data stats
• Execution configurations
  • Configure ExecutionOptions
  • Configure DataContext
• Run multiple Datasets in one cluster
  • How do subclusters work?
  • Assign each Dataset to a subcluster
  • Example: Two Datasets, two subclusters
  • Use subclusters with Ray Train
  • API reference
• End-to-end: Offline Batch Inference
  • Quickstart
  • Configuration and troubleshooting
• Advanced: Performance tips and tuning
  • Optimize transforms
  • Optimize reads
  • Reduce memory usage
  • Configure execution
  • Make execution reproducible
• Advanced: Scaling out expensive collate functions
  • Move the collate function to Ray Data
  • Create a custom collate function that runs in Ray Data
  • Ensure batch size alignment
  • Put it all together
  • Advanced: Handle custom data types
• Advanced: Read and write custom file types
  • Read data from files
  • Write data to files