Views
No views yet
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.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.bahadirakdemir/gemma-4-31B-it-text-fp8.model_type: gemma4_unified), which is
newer than the classic gemma4 (e.g. 31B). You need:gemma4_unified was added)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.1vllm serve bahadirakdemir/gemma-4-12B-it-text-fp8 \
2 --quantization modelopt \
3 --max-model-len 8192 \
4 --max-num-batched-tokens 8192 \
5 --gpu-memory-utilization 0.5 \
6 --limit-mm-per-prompt '{"image": 0, "audio": 0}'bahadirakdemir/gemma-4-12B-it-assistant-fp8:1vllm serve bahadirakdemir/gemma-4-12B-it-text-fp8 \
2 --quantization modelopt \
3 --max-model-len 8192 \
4 --max-num-batched-tokens 8192 \
5 --gpu-memory-utilization 0.5 \
6 --limit-mm-per-prompt '{"image": 0, "audio": 0}' \
7 --speculative-config '{"model": "bahadirakdemir/gemma-4-12B-it-assistant-fp8", "num_speculative_tokens": 4}'vllm/vllm-openai:gemma4-0505-arm64-cu130 on NVIDIA GB10.| 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 |