Search Results for "activation"

Ring is a reasoning Mixture-of-Experts (MoE) large language model (LLM) developed by inclusionAI. It is built from or derived from Ling. Its design emphasizes reasoning, efficiency, and modular expert activation. In its “flash” variant (Ring-flash-2.0), it optimizes inference by activating only a subset of experts. It applies reinforcement learning/reasoning optimization techniques. Its architectures and training approaches are tuned to enable efficient and capable reasoning performance. Reasoning-optimized model with reinforcement learning enhancements. ...

...Introduced in the ICML 2024 paper “MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases”, it focuses on delivering strong reasoning and generalization capabilities in models under one billion parameters. The framework integrates several architectural innovations—SwiGLU activation, deep and thin network design, embedding sharing, and grouped-query attention (GQA)—to achieve a superior trade-off between model size, inference speed, and accuracy. MobileLLM demonstrates remarkable performance, with the 125M and 350M variants outperforming previous state-of-the-art models of the same scale by up to 4.3% on zero-shot commonsense reasoning tasks.

Ling is a Mixture-of-Experts (MoE) large language model (LLM) provided and open-sourced by inclusionAI. The project offers different sizes (Ling-lite, Ling-plus) and emphasizes flexibility and efficiency: being able to scale, adapt expert activation, and perform across a range of natural language/reasoning tasks. Example scripts, inference pipelines, and documentation. The codebase includes inference, examples, models, documentation, and model download infrastructure. As more developers and researchers engage with the platform, we can expect rapid advancements and improvements, leading to even more sophisticated applications. ...

...It includes a “neuron explainer” component that, given a target neuron or latent feature, proposes natural language explanations or heuristics (e.g. “this neuron activates when the input has property X”) and then simulates activation behavior across example inputs to test whether the explanation holds. The project also contains a “neuron viewer” web component for browsing neurons, explanations, and activation patterns, making it more interactive and exploratory.

...Automatic batch size selection, distributed training (DDP, DeepSpeed), parameter efficient fine-tuning (PEFT), 4-bit quantization (QLoRA), and larger-than-memory datasets. Retain full control of your models down to the activation functions. Support for hyperparameter optimization, explainability, and rich metric visualizations. Experiment with different model architectures, tasks, features, and modalities with just a few parameter changes in the config. Think building blocks for deep learning.

...The project implements techniques that allow model components to be dynamically moved between CPU memory and GPU memory during inference, significantly reducing the amount of GPU VRAM required to run the model. This approach takes advantage of the sparse activation properties of mixture-of-experts architectures, where only a subset of expert networks are used for each token during generation. By selectively loading and caching the required experts, the system avoids keeping the entire model in GPU memory at once. The repository includes notebooks and code examples that demonstrate how to run large language models on consumer hardware such as personal GPUs or cloud notebook environments.

Open-source pre-training implementation of Google's LaMDA research paper in PyTorch. The totally not sentient AI. This repository will cover the 2B parameter implementation of the pre-training architecture as that is likely what most can afford to train. You can review Google's latest blog post from 2022 which details LaMDA here. You can also view their previous blog post from 2021 on the model.