llama.cpp b11103 Adds DFlash Speculative Decoding Support for HunyuanOCR
llama.cpp release b11103 adds layer-input tap support for HunyuanOCR target graphs, enabling DFlash speculative decoding for Hunyuan vision-language models. The update also fixes GGUF conversion issues related to special token ID lookups and model configuration handling during DFlash draft conversions. Speculative decoding with DFlash significantly boosts inference speed for vision-language models like HunyuanOCR without reducing output accuracy. Adding native support in llama.cpp allows developers to run high-throughput local OCR workflows more efficiently on edge devices and personal hardware. Testing on Tencent HunyuanOCR 1.5 achieved a draft acceptance rate of approximately 0.5 while generating byte-identical OCR output compared to non-speculative runs. The modification registers residual layer-input tensors required by DFlash draft models to construct cross-context without changing behavior when speculative draft models are disabled.
## BACKGROUND
llama.cpp is an open-source C/C++ framework for running LLMs efficiently on local hardware. Speculative decoding is an optimization technique where a smaller, faster 'draft' model predicts multiple candidate tokens that the larger target model verifies in parallel to reduce overall latency. DFlash is a block-diffusion draft framework designed to accelerate token generation in complex vision-language models such as HunyuanOCR.