Vibeleaderboard
← Back to Vibers
Amazon Science
Company

Amazon Science

Index Rank192
@AmazonScience

3 Intel

Is this you? Sign in with X to claim this profile.

Intel

AWS 'priming' builds efficient hybrid LLMs from Transformers(www.amazon.science)

AWS shows how to turn a pretrained Transformer into a hybrid state-space model with under 0.5% of pretraining tokens. The 32B result stays within about 1% of its source, roughly halves KV cache, and decodes up to 2.3x faster, with code, models and a vLLM plugin released.

AI Toolsbuilt by @AmazonScience
TorchNeuron opens up Trainium to PyTorch, validated by agents(www.amazon.science)

Amazon open-sourced TorchNeuron, a PyTorch backend for AWS Trainium with adaptive eager mode and shared compile caching. The team used agents to cull Hugging Face down to 841 representative models and built a tensor-shape registry for fast staged regression tests.

Developer Toolsbuilt by @AmazonScience
Amazon's TorchNeuron brings native PyTorch to Trainium chips(www.aboutamazon.com)

Amazon announced TorchNeuron, a native PyTorch backend for Trainium2 and Trainium3 that runs existing code with a configuration change. It plugs into PyTorch's device registration, works with TorchTitan and Hugging Face tooling, and is slated for Apache 2.0 release on Oct 20.

Developer Toolsbuilt by @AmazonScience