Search Results for "expert"

...It keeps the dense portion of the quantized model in memory while streaming routed experts from a large disk-based store as they are needed. The runtime is implemented in pure C, requires no Python or BLAS during inference, and can operate without a GPU. Compressed attention caches, expert caching, optional hot tiers, and speculative decoding reduce memory pressure and improve repeated use. A planning tool calculates safe disk, RAM, and VRAM placement before loading the model, while a diagnostic command checks system readiness. Colibri includes terminal chat, an OpenAI-compatible text API, and a browser client, but disk-bound generation can be slow on cold caches.

...The runtime streams model trunk layers and routed experts from disk instead of keeping all weights resident in memory. Memory presets balance pinned layers and an expert LRU cache for laptops, desktops, workstations, and servers. Incremental generation preserves KV cache and recurrent state between tokens. The repository also includes tokenization, safetensors loading, diagnostics, benchmarks, trace replay, and comparisons against a PyTorch reference.

WASTE is an embeddable C inference engine for running extremely large mixture-of-experts models when the weights exceed available RAM. It keeps the shared model trunk in memory and streams only the experts selected for each token from fast NVMe storage. A bounded cache reuses recently needed experts, while lookahead routing begins disk reads before the next layer requires them. Its main target is the full 2.78-trillion-parameter Kimi K3 model, including multimodal image input. The engine has...