Search Results for "mla"

FlashMLA is a high-performance decoding kernel library designed especially for Multi-Head Latent Attention (MLA) workloads, targeting NVIDIA Hopper GPU architectures. It provides optimized kernels for MLA decoding, including support for variable-length sequences, helping reduce latency and increase throughput in model inference systems using that attention style. The library supports both BF16 and FP16 data types, and includes a paged KV cache implementation with a block size of 64 to efficiently manage memory during decoding. ...

DeepSeek-V3 is a robust Mixture-of-Experts (MoE) language model developed by DeepSeek, featuring a total of 671 billion parameters, with 37 billion activated per token. It employs Multi-head Latent Attention (MLA) and the DeepSeekMoE architecture to enhance computational efficiency. The model introduces an auxiliary-loss-free load balancing strategy and a multi-token prediction training objective to boost performance. Trained on 14.8 trillion diverse, high-quality tokens, DeepSeek-V3 underwent supervised fine-tuning and reinforcement learning to fully realize its capabilities. ...

OpenMythos is an experimental, open-source implementation that attempts to reconstruct a hypothesized architecture behind advanced language models using a design called a Recurrent-Depth Transformer. The project explores the idea that instead of stacking hundreds of unique transformer layers, a smaller set of layers can be reused iteratively during inference to achieve deeper reasoning without increasing parameter count. It divides computation into three main stages, including a...

Kimi Linear is a hybrid linear attention architecture developed for efficient language modeling across short, long-context, and reinforcement learning workloads. Its core mechanism, Kimi Delta Attention, refines the gated delta rule with fine-grained controls for managing finite-state recurrent memory. The architecture combines KDA and global Multi-head Latent Attention layers at a 3:1 ratio to preserve model quality while reducing memory demands. Released Base and Instruct checkpoints...

Ling-3.0-tiny is inclusionAI’s lightweight hybrid reasoning Mixture-of-Experts model, designed to provide capable reasoning and agentic performance at low inference cost. It contains 7.9B total parameters while activating only 1.3B per token, using a hybrid architecture that alternates Kimi Delta Attention and Multi-Head Latent Attention with a sparse 128-expert MoE. The model supports both fast responses and configurable multi-step thinking, covering general agents, coding, mathematics,...

GigaChat 3 Ultra is a flagship instruct-model built on a custom Mixture-of-Experts architecture with 702B total and 36B active parameters. It leverages Multi-head Latent Attention to compress the KV cache into latent vectors, dramatically reducing memory demand and improving inference speed at scale. The model also employs Multi-Token Prediction, enabling multi-step token generation in a single pass for up to 40% faster output through speculative and parallel decoding techniques. Its...