FP8-quantized version of google/gemma-4-12B-it-assistant, the Multi-Token Prediction (MTP) drafter that pairs with the Gemma 4 12B-it target for speculative decoding. The drafter is a small 4-layer model; its linear layers are quantized to FP8 (E4M3) with per-tensor static scales via NVIDIA ModelOpt. The drafter↔target handshake projections (pre_projection, post_projection) and lm_head stay in BF16.

This drafter is not a standalone text model — it requires a target model to provide shared_kv_states at inference time. Use it as a spec model in vLLM, paired with the 12B-it target.

Compatible target

Pairs with bahadirakdemir/gemma-4-12B-it-text-fp8 — the FP8-quantized 12B-it text tower produced by the same pipeline. The FP8 scales were calibrated by running real speculative decoding against the 12B-it target over 32 instruct-style prompts.

Requirements

This drafter and its target use the unified Gemma 4 architecture (gemma4_unified), which is newer than the classic gemma4 (e.g. 31B). You need:

transformers ≥ 5.10.0
vLLM with gemma4_unified support — at the time of writing this is on the main branch / nightly (uv pip install -U vllm --pre), not yet in a tagged stable release (≤ 0.22.0). It will be in the next stable release.

Usage with vLLM

bash

vllm serve bahadirakdemir/gemma-4-12B-it-text-fp8 \
  --quantization modelopt \
  --max-model-len 8192 \
  --max-num-batched-tokens 8192 \
  --gpu-memory-utilization 0.5 \
  --limit-mm-per-prompt '{"image": 0, "audio": 0}' \
  --speculative-config '{"model": "bahadirakdemir/gemma-4-12B-it-assistant-fp8", "num_speculative_tokens": 4}'

Tested with vllm/vllm-openai:gemma4-0505-arm64-cu130 on NVIDIA GB10.

This is the 12B counterpart of bahadirakdemir/gemma-4-31B-it-assistant-fp8, produced by the same pipeline.

License: Apache 2.0, inherited from upstream Gemma 4 — see the Gemma 4 license.