Gemma 4 12B IT — INT4 (MLP-only, auto-round)

4-bit (W4A16, group size 128) quantization of google/gemma-4-12B-it, produced with intel/auto-round in RTN mode and exported to the GPTQ-compatible format.

Following NVIDIA's dense-Gemma-4 recipe (nvidia/Gemma-4-31B-IT-NVFP4), only the MLP / feed-forward linear layers are quantized to 4-bit, while the attention projections (q/k/v/o) are kept in BF16. Gemma 4's attention activations carry large per-channel outliers that 4-bit quantization cannot represent well, so quantizing attention degrades quality and breaks several inference kernels. Keeping attention in BF16 avoids this.

The vision/audio embedders, embeddings and lm_head are also kept at full precision.

Serving with vLLM (verified)

Tested on RTX 5090 (Blackwell, sm120), CUDA 13.

Gemma 4 12B "unified" support landed in vllm-project/vllm#44429 and is not yet in a stable release — you need a vLLM nightly build. On Blackwell, the FlashInfer sampler fails to JIT-compile, so disable it.

Install nightly (CUDA 13; use cu129 URLs on CUDA 12.9 hosts):

uv pip install -U vllm --pre \
  --extra-index-url https://wheels.vllm.ai/nightly/cu130 \
  --extra-index-url https://download.pytorch.org/whl/cu130 \
  --index-strategy unsafe-best-match

Serve:

export VLLM_USE_FLASHINFER_SAMPLER=0
vllm serve <path-to-this-model> \
  --served-model-name gemma4-12b \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.90 \
  --host 0.0.0.0 --port 8000

The model loads in ~11 GB, leaving plenty of room on a 32 GB card for KV cache (raise --max-model-len accordingly). Recommended sampling for Gemma 4: temperature=1.0, top_p=0.95, top_k=64.

Quick test:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma4-12b",
    "messages": [{"role": "user", "content": "Explain quantization in one paragraph."}],
    "max_tokens": 200, "temperature": 1.0, "top_p": 0.95, "top_k": 64
  }'

Usage (transformers)

Also loads under transformers (requires gptqmodel):

pip install transformers torch gptqmodel optimum

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "aleksandard/gemma-4-12B-it-int4-MLPonly-AutoRound"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype="auto", device_map="cuda"
)

messages = [{"role": "user", "content": "Explain quantization in one paragraph."}]
ids = tok.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=False
).to("cuda")
out = model.generate(ids, max_new_tokens=256)
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))

Notes

• vLLM stable (<= 0.22.0) does not serve Gemma 4 dense 12B — it hits a shape mismatch in the attention path caused by Gemma 4's heterogeneous head dimensions (head_dim 256 for sliding-window layers vs 512 for global layers). Use a nightly build as described above.
• On Blackwell, VLLM_USE_FLASHINFER_SAMPLER=0 is required to avoid a FlashInfer JIT-compile failure during sampling.

Reproduce

auto-round \
  --model google/gemma-4-12B-it \
  --scheme W4A16 \
  --iters 0 \
  --disable_opt_rtn \
  --layer_config '{"model.language_model.layers.\d+.self_attn.q_proj":{"bits":16},"model.language_model.layers.\d+.self_attn.k_proj":{"bits":16},"model.language_model.layers.\d+.self_attn.v_proj":{"bits":16},"model.language_model.layers.\d+.self_attn.o_proj":{"bits":16}}' \
  --format auto_gptq \
  --output_dir ./gemma-4-12B-it-int4-MLPonly

Limitations

This is a quantized derivative; it inherits all limitations and biases of the base model and may show additional deviation due to 4-bit quantization. See the base model card for full details. Quantization was calibration-free (RTN); a calibrated build may recover some quality.

License

Apache 2.0, inherited from the base model. This repository changes only the numeric precision of the weights.