llama.cpp 发布: v0.5.0

来源摘要

## Overview This release focuses on backend performance and correctness, broader model coverage, and more robust server/router operation. It adds HRM-Text (DFM Mimir 1B) support, MiMo-V2.6 and HunyuanOCR conversion support, ggml 0.25.0 backend improvements, multi-address HTTP binding, image outputs from function calls, and several chat parser/UI fixes. ### Highlights - Accelerate CUDA `conv2d` with implicit GEMM ([#29135](https://github.com/ggml-org/llama.cpp/pull/29135)) - Add Metal MoE and SSM_CONV fusion optimizations ([#28948](https://github.com/ggml-org/llama.cpp/pull/28948)) - Allow the server to bind to multiple addresses ([#28690](https://github.com/ggml-org/llama.cpp/pull/28690)) ### API changes - Add `llama_adapter_lora_init_from_file_ptr()` for loading LoRA from an open FILE ([#28993](https://github.com/ggml-org/llama.cpp/pull/28993)) - Document `llama_model_load_from_file_ptr()` as reading from the current position and requiring aligned mmap ([#28993](https://github.com/ggml-org/llama.cpp/pull/28993)) - Add `LLAMA_VOCAB_TYPE_TEST` dummy tokenizer ([#29084](https://github.com/ggml-org/llama.cpp/pull/29084)) - Add `input_image` support to server function-call outputs ([#22575](https://github.com/ggml-org/llama.cpp/pull/22575)) - Allow `--host` to accept comma-separated TCP addresses and UNIX sockets ([#28690](https://github.com/ggml-org/llama.cpp/pull/28690)) ### New models - Add HRM-Text / DFM Mimir 1B support ([#27625](https://github.com/ggml-org/llama.cpp/pull/27625)) - Add MiMo-V2.6 conversion support ([#29257](https://github.com/ggml-org/llama.cpp/pull/29257)) - Add DFlash support for HunyuanOCR ([#28890](https://github.com/ggml-org/llama.cpp/pull/28890)) - Extend Nemotron MTP and Nemotron-H model handling ([#29018](https://github.com/ggml-org/llama.cpp/p