github llama.cpp · 13h ago · inference

llama.cpp b10707 optimizes KV cache cell scanning for faster generation

from English

The llama.cpp project released version b10707, which includes a performance optimization to the key-value (KV) cache handling. The update modifies the `for_each_token_in` function to stop scanning once all sequences within a specific cell have been seen, rather than iterating through all possible maximum sequences.

This change targets the n-gram path via the `get_prev_tokens` caller.
On an RTX PRO 6000 with Qwen3.8-Flash-Next UD-Q4_K_XL, generation speed increased from 56.3 to 74.3 tokens per second at a 55k context.
At a 132k context, generation speed improved from 33.6 to 50.9 tokens per second.
The optimization scales with the number of used cells, making gains more visible at longer contexts while leaving prompt processing unchanged.

The release provides binaries for macOS (Apple Silicon and Intel), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Android, Windows (CPU, CUDA, Vulkan, OpenVINO, SYCL, ROCm), and openEuler.

Importance 1/3 Trust 1/3 llama.cpp Inference efficiency Local inference

Read original

github llama.cpp · just now · 1 view Live

llama.cpp b10721 fixes WebGPU crash and adds Windows CUDA 13.4 support

The llama.cpp project has released build b10721, which includes a critical fix for the WebGPU backend and expands hardware support on Windows.

github llama.cpp · 2h ago

llama.cpp b10718 extends MOE fusion to specdec and multi-token

The llama.cpp project released version b10718, which includes a significant update to its CUDA backend. The most notable change is the extension of Mixture of Experts (MOE) fusion to support speculative decoding (specdec), addressing previous limitations where MOE glu fusion and topk-router fusion were restricted to processing only one token at a time.

github llama.cpp · 5h ago · 1 view

llama.cpp b10721 fixes WebGPU crash and adds Windows CUDA 13.4 support

llama.cpp b10718 extends MOE fusion to specdec and multi-token

llama.cpp b10715 fuses DFlash encoder into KV cache injection

llama.cpp b10714 tunes Vulkan mat-vec rows for Strix Halo

llama.cpp b10706 adds SWIGLU_CLAMP and Vulkan shader support