llama.cpp b10237: MTP support for DeepSeek V3.2

llama.cpp build b10237 adds Multi-Token Prediction (MTP) support for DeepSeek V3.2, so the model’s native multi-token architecture works correctly when running locally instead of being stripped or misidentified.

What changed in b10237:

• MTP support for DeepSeek V3.2 — PR #26457, co-authored by Stanisław Szymczyk
• MTP layers excluded from model type discovery — the loader no longer treats MTP heads as part of the base model’s layer count, which previously could trigger architecture misidentification
• Pre-built binaries ship for macOS Apple Silicon, iOS, Ubuntu (CPU/Vulkan/ROCm 7.2/OpenVINO/SYCL FP32+FP16), Windows (CPU/CUDA 12.4/CUDA 13.3/Vulkan/HIP), Android arm64, and openEuler (310p/910b ACL Graph)

MTP is DeepSeek’s approach to speculative decoding baked into the architecture — the model predicts multiple future tokens in a single forward pass, then verifies them. On llama.cpp this means faster generation for DeepSeek V3.2 without a separate draft model. The type-discovery fix matters because older builds could miscount layers when MTP heads were present, causing load failures or silent degradation.

Available now across all listed platforms on the b10237 release page. If you’re running DeepSeek V3.2 on an earlier llama.cpp build, upgrade — older versions either ignored or misclassified the MTP layers.