Anyscale's LLMForge
Anyscale Exclusive ML Library for LLM Fine-Tuning
What is LLMForge?
LLMForge is a Ray library for LLM fine-tuning, only available on Anyscale.
Combining a collection of design patterns across RayTurbo (Anyscale’s leading compute engine, built on Ray), Ray Train, and Ray Data—alongside other open-source libraries like Deepspeed, HuggingFace accelerate—LLMForge is the easiest library for LLM fine-tuning at any scale.
Any Type of Fine-Tuning
Full-Parameter
Full-parameter fine-tuning takes the LLM "as is" and trains it on the given dataset. In principle, this is similar to the pre-training stage of the LLM optimizing all the parameters of the neural network.
LoRA
LoRA (Low-Rank Adaptation) is a fine-tuning technique that freezes most of your LLM's weights, and instead adds and optimizes select parameters. This technique is typically more resource-efficient It also helps to regularize the model, so it can effectively retain learned information.
Use Cases
Casual Language Modeling
Run any type of LLM fine-tuning, including casual language modeling, where each token is predicted based on all past tokens.
Benefits
Advanced Customization
The LLMForge “custom” mode offers more flexibility and control over the fine-tuning parameters including cluster shape and type, allowing for advanced optimizations and customization.
Streamlined Deployment
LLMForge integrates directly with Anyscale, so you can build an LLM fine-tuning job in Anyscale Jobs and easily easily integrate it with production flows through CI/CD pipelines.
Improved Observability
Take advantage of standard logging frameworks such as W&B and MLFlow, plus use Ray dashboard and Anyscale loggers for debugging and progress monitoring.
Any Model, Any Prompt
Get out-of-the-box support for popular models like common LLaMA LLMs, or configure any HuggingFace model and prompt format in custom mode.
Feature Highlights
Multi-Stage Fine-Tuning
Combine fine-tuning across multiple datasets by using a previously-created checkpoint as initialization for another round of fine-tuning.
Context Length Extension
Extend the context length of the model using methods like RoPE scaling.
Configure Any Hyperparameters
LLMForge gives you full control over learning hyperparameters such as learning rate, number of epochs, batch size, and more.
“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
Nope! In addition to LLMForge for fine-tuning, the Ray and Anyscale suite for distributed computing also includes the following libraries:
Open source libraries:
Anyscale-only ML libraries:
- Ray LLM for LLM Inference
Try Anyscale Today
Build, deploy, and manage scalable AI and Python applications on the leading AI platform. Unlock your AI potential with Anyscale.