Adaptive Harmony. Our internal LLM stack.
We have built a unified inference, pretraining, and reinforcement learning codebase in Rust+Python. It powers all of our products and internal research and development workloads.
Let's enable researchers to focus on research again.
Design philosophy.
Preference tuning methods are often an after-thought, either tacked-on distributed training codebase or built with method-specific codebase that struggle to generalize.
Adaptive Harmony is fully dedicated to preference tuning, accelerating researchers workflows.
Flexibility.
Typical LLM codebases are deeply entangled with their distributed training strategy, slowing down researchers. In Harmony, environments, logic, and model distributions are decoupled: researchers can focus on creative experiments and novel recipes.
Performance.
Models in Harmony are implemented with custom kernels, careful memory management, and extensive profiling to validate throughput in both compute and memory-bound regimes. Even in resource-constrained scenarios, for enterprises without clusters.
Robustness.
The core of Harmony is built in Rust, with the machine learning logic exposed in Python for easy iteration by researchers. It is also extensively tested, up to regular reproduction of full-scale PPO recipes to control for performance regressions.
🦀 + 🐍 = ❤️.
Most of Harmony is written in Rust, but machine learning logic is exposed in Python to enable users to easily iterate. Combining Rust with extensive test coverage and systematic end-to-end tests on typical use cases guarantees a robust, reliable codebase.
Decoupled environments, logic, and model distribution.
Advanced LLM use cases require repeated interactions between instances of models and complex environments. Often, researchers end-up spending more time dealing with distributed training idiosyncrasies than on experiments and iterations of their ideas.
In Harmony, users simply write recipes, focusing on the logic of the interactions. Recipes get lowered just-in-time for distribution, and it's easy to blend inference and training with custom objectives across countless models and environments.
In Harmony, users simply write recipes, focusing on the logic of the interactions. Recipes get lowered just-in-time for distribution, and it's easy to blend inference and training with custom objectives across countless models and environments.
High-performance across training and inference regimes.
Harmony integrates custom kernels and careful memory management to deliver optimal performance across both memory-bound and compute-bound regimes that may occur as researchers experiment with new recipes.
Read our latest research.
.jpg)
Looking for a packaged solution?
Check out Adaptive Engine, our platform for testing, serving, monitoring, and iterating on large language models.
Copyright © 2024 Adaptive ML, Inc. All rights reserved.