Search Results for "compress"

...Developed by researchers at Facebook (Meta AI), ToMe introduces an efficient technique that merges similar tokens within transformer layers, reducing redundant computation while preserving model accuracy. This approach differs from token pruning, which removes background tokens entirely; instead, ToMe merges tokens based on feature similarity, allowing it to compress both foreground and background information efficiently. ToMe integrates seamlessly into existing transformer models such as DeiT, MAE, SWAG, and timm ViTs, offering 2–3x speedups during inference and substantial efficiency gains during training. The method can be applied dynamically at inference time or incorporated into training for improved performance.

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 training corpus incorporates ten languages, enriched with books, academic sources, code datasets, mathematical tasks, and more than 5.5 trillion tokens of high-quality synthetic data. ...