MTP/README.md

Gemma 4 E4B MTP drafter

Multi-Token Prediction (MTP) drafter for unsloth/gemma-4-E4B-it-GGUF. It runs as a speculative draft model that shares the target's KV cache and speeds up text generation.

Verified on a single B200 with the gemma-4-E4B-it-Q4_K_M.gguf target: 150 tok/s without MTP, 214 tok/s with MTP, 0.66 draft acceptance.

MTP for E4B was added to llama.cpp on 2026-06-08 (PR ggml-org/llama.cpp#24282). You need a llama.cpp build from after that date. Older builds cannot load these (arch gemma4-assistant).

Run E4B with -fa off. With flash attention on, the draft model currently aborts in the CUDA flash attention kernel.

Files

For -hf auto-discovery a Q8_0 drafter sits at the repo root as mtp-gemma-4-E4B-it.gguf. The same three precisions live in MTP/:

• mtp-gemma-4-E4B-it-Q8_0.gguf (smallest, recommended; mirrored at the repo root as mtp-gemma-4-E4B-it.gguf)
• mtp-gemma-4-E4B-it-BF16.gguf
• mtp-gemma-4-E4B-it-F16.gguf

Build llama.cpp

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp

# CUDA build. Set the arch for your GPU: 89 (RTX 4090), 90 (H100), 100 (B200).
cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=90
cmake --build build --config Release -j --target llama-server

Run, the easy way

A recent llama.cpp finds the drafter automatically from the root mtp- file, so -hf is all you need. No --model-draft.

./build/bin/llama-server \
  -hf unsloth/gemma-4-E4B-it-GGUF:Q4_K_M \
  --spec-type draft-mtp --spec-draft-n-max 4 \
  -ngl 999 -fa off

If your build is too old to auto-discover the sibling, use the explicit form below.

Run with an explicit drafter

Use this to choose a precision or point at a local file.

hf download unsloth/gemma-4-E4B-it-GGUF gemma-4-E4B-it-Q4_K_M.gguf --local-dir .
hf download unsloth/gemma-4-E4B-it-GGUF MTP/mtp-gemma-4-E4B-it-Q8_0.gguf --local-dir .

./build/bin/llama-server \
  -m gemma-4-E4B-it-Q4_K_M.gguf \
  --model-draft MTP/mtp-gemma-4-E4B-it-Q8_0.gguf \
  --spec-type draft-mtp --spec-draft-n-max 4 \
  -ngl 999 -fa off

Multi GPU: add --spec-draft-device CUDA0 -sm layer. The drafter pairs with any quant of the E4B. Quantized KV cache works.