llama.cpp b11103：为 HunyuanOCR 增加 DFlash 投机解码支持

AI 摘要

llama.cpp 发布 b11103 版本，为 HunyuanOCR 目标模型增加 DFlash 投机解码支持，在 Hunyuan 图构建中注册层输入张量，修复了此前因张量为空导致的断言中止问题。同时修复了针对 HunYuan 目标转换 DFlash 草稿模型时的词表处理错误，包括缺失的 _fix_special_tokens 方法和配置读取错误。测试显示图像请求可正常运行，草稿接受率约 0.5，OCR 输出与非投机运行逐字节一致。

正文节选

<details open> spec : support DFlash for HunyuanOCR (#28890) * model : add DFlash layer-input taps for HunyuanVL DFlash speculative decoding needs the target graph to expose the residual stream entering each layer (res->t_layer_inp[il]) - the draft model reads those tensors to build its cross-context. Qwen3 and the other DFlash-capable targets register them, but the Hunyuan graphs do not, so serving a DFlash draft against a HunyuanOCR target aborts during the first graph build: GGML_ASSERT

发布时间：2026-09-22 22:52

抓取时间：2026-09-22 23:13

来源机构：ggml-org