bahadirakdemir/gemma-4-12B-it-text-fp8

Gemma 4 12B-it — Text FP8 (ModelOpt)

FP8-quantized text tower of google/gemma-4-12B-it, the unified (encoder-free multimodal) Gemma 4 12B model. The linear layers of the language model are quantized to FP8 (E4M3) with per-tensor static scales calibrated offline via NVIDIA ModelOpt; lm_head and the (tied) embeddings stay in BF16.

This checkpoint is text-only: the vision/audio encoder weights are not included. The config still advertises Gemma4UnifiedForConditionalGeneration and the weights live under model.language_model.*, which keeps the layer namespace compatible with the paired MTP drafter for speculative decoding (see below). Serve it with --limit-mm-per-prompt '{"image": 0, "audio": 0}' so the absent multimodal encoder is never invoked.

Produced by the same pipeline as bahadirakdemir/gemma-4-31B-it-text-fp8.

Requirements

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

• transformers ≥ 5.10.0 (when gemma4_unified was added)
• 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

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}'

For speculative decoding, pair it with the matching FP8 MTP drafter bahadirakdemir/gemma-4-12B-it-assistant-fp8:

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.

Quantization details

Method	ModelOpt FP8 PTQ (E4M3, per-tensor static scales)
Quantized	language-model linears (attention + MLP projections)
Kept in BF16	lm_head, tied embeddings, all norms
Calibration	32 instruct-style prompts, max length 1024

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