llama : MTP support for DeepSeek V3.2

[fairydreaming]

Overview

This PR adds MTP support to DeepSeek V3.2 model. Mostly useful for code generation (with --spec-draft-n-max 2 draft acceptance is around 80%).

Additional information

It's a shameless copy of #25980 as these two models share common architecture.

TODO for next PR: check if adding lightning indexer in MTP layer improves acceptance rate during long-context generations.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO, I copied the code myself.

[ggerganov]

Ack on the libllama changes.

[fairydreaming]

gpu-webgpu-nvidia

FLASH_ATTN_EXT(hsk=256,hsv=256,nh=8,nr23=[4,1],kv=1024,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]): OK
/home/ggml/actions-runner/_work/llama.cpp/llama.cpp/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp:2822: GGML_ASSERT(max_kv_tile > 0) failed

gpu-cuda

2026-08-02T16:43:40.8864070Z sentinel mismatch: a   ROPE(type=f16,ne_a=[36,16,2457,1],n_dims=36,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=2,inplace=0): �[1;31mFAIL�[0m

Uhh recently it feels like playing Russian roulette with CI.

[ggerganov]

The CUDA failure seems to be some issue with @taronaeo's runner - I've noticed it failing randomly on different CUDA tests:

https://github.com/ggml-org/llama.cpp/actions/runs/30755768689/job/91517537249?pr=26457#step:1:2

It's OK to ignore it for now.

[am17an]

At this point only you are running this model with lineage bench

[fairydreaming]

At this point only you are running this model with lineage bench

@am17an Tell me about it. Recently I feel the urge to get back to my old MiniMax-Text-01 implementation that I never finished.