本機推論
KoboldCpp 1.120 Adds Direct I/O Loading Mode, Allows mmap and mlock to Be Enabled Together
The new release adds `--usedirectio` for large GGUF weights and allows the relevant memory pages to be locked when a model is mapped with mmap. It also supports Qwen3.8-Flash-Next, Ling-3.0-flash, and custom JavaScript tools, but the project has not published loading-speed or memory benchmarks.
KoboldCpp 1.120 exposes more model-file loading strategies to local inference users. The [official release notes](https://github.com/LostRuins/koboldcpp/releases/tag/v1.120) add `--usedirectio`, enabling supported platforms to read GGUF files using direct I/O. Another change allows mmap and mlock to be used together. mmap lets the operating system map model pages on demand, making it suitable when a model exceeds available physical memory or when multiple processes share the same file. mlock requests that the system not swap mapped pages out. Combining the two can reduce latency variability caused by paging during inference, but it increases resident-memory pressure and is subject to operating-system limits on locked memory.
The main value of direct I/O is during loading: it provides a read option that bypasses the conventional cache path, potentially preventing a one-time load of tens of gigabytes of weights from evicting other files from the cache. Its actual impact, however, depends heavily on the SSD, file system, read alignment, and GPU offload configuration. The release notes provide no cold-start timing, throughput figures, or supported-platform matrix, so it should not be assumed to be faster than mmap. Engineering teams should benchmark the same GGUF under both cold- and warm-load conditions, along with time to first token after loading and peak system-memory usage.
Version 1.120 also adds support for Qwen3.8-Flash-Next and Ling-3.0-flash, while warning that incorrectly quantized files are already circulating in the community. Kobold Lite can now load user-configured JavaScript tools and expose them to models through the standard tool-calling interface. This also means that the file, network, and process permissions available to tool code must be isolated separately. The release additionally fixes unintended assistant prefill triggering and incorrect failsafe-mode selection. A [LocalLLaMA discussion](https://www.reddit.com/r/LocalLLaMA/comments/1w2c4el/koboldcpp_v1120_released/) indicates that DirectIO and combined page locking are the changes attracting the most community attention, but current reports remain insufficient to support conclusions across different hardware configurations.