Search Results for "cache memory simulator"

TurboFieldfare is a custom Swift and Metal runtime for running the instruction-tuned Gemma 4 26B-A4B model on Apple Silicon Macs with limited memory. Instead of loading the entire model, it keeps the shared core and KV cache in RAM while streaming only the routed experts required for each token from SSD. This approach reduces active memory use to roughly 2 GB while the installed model occupies about 14.3 GB of storage. The project includes a native Mac application, command-line tools, a streaming installer, a Swift library, and an experimental OpenAI-compatible local server. ...

...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. On very compute-bound settings, it can reach up to ~660 TFLOPS on H800 SXM5 hardware, while in memory-bound configurations it can push memory throughput to ~3000 GB/s. The team regularly updates it with performance improvements; for example, a 2025 update claims 5 % to 15 % gains on compute-bound workloads while maintaining API compatibility.

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 contain 48 billion total parameters, activate 3 billion parameters per token, and support contexts up to one million tokens. The models were trained on 5.7 trillion tokens and can reduce KV cache requirements by up to 75 percent. ...

...The model contains 33B total parameters with only 3B activated per token, allowing it to deliver strong coding performance while remaining efficient enough to run locally on modern consumer hardware. It uses a hybrid attention architecture that combines Sliding Window Attention and global attention layers, reducing memory requirements and improving inference speed. Laguna XS.2 supports native reasoning with interleaved thinking between tool calls, enabling more capable autonomous coding agents and multi-step workflows. The model features a 262K-token context window, preserved reasoning across interactions, FP8 KV-cache optimization, and compatibility with local deployment ecosystems such as Ollama and vLLM.

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. ...