xdavxd

gemma-4-12B-it-heretic-NVFP4

README

License: apache-2.0

Model Overview

Model Architecture: Gemma 4
- Input: Text / Image / Audio
- Output: Text
Model Optimizations:
- Weight quantization: FP4
- Activation quantization: FP4
Release Date: 2026-07-03
Version: 1.0
Quantized by: xdavxd
Base Model: coder3101/gemma-4-12B-it-heretic
Original Model: google/gemma-4-12B-it

This model is a quantized version of coder3101/gemma-4-12B-it-heretic. It was evaluated on several tasks to assess its quality in comparison to the original model.

Model Optimizations

This model was obtained by quantizing the weights and activations of coder3101/gemma-4-12B-it-heretic to FP4 data type using the NVFP4 format, ready for inference with vLLM. This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 65%.

Weights are quantized with FP4 (group_size=16), and activations are quantized with FP4 using local per-group scaling. Only the weights and activations of the linear operators within transformer blocks are quantized using LLM Compressor. Vision, audio, embedding, and output head layers are kept in their original precision.

Deployment

Use with vLLM

This model can be deployed using vLLM. For detailed instructions including multi-GPU deployment, multimodal inference, thinking mode, function calling, and benchmarking, see the Gemma 4 vLLM usage guide.

Note: For 12B, ensure you have the latest package of transformers (5.13.0 or later) or it may fail to load.

Start the vLLM server:

markdown
vllm serve xdavxd/gemma-4-12B-it-heretic-NVFP4 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.90

To enable thinking/reasoning and tool calling:

markdown
vllm serve xdavxd/gemma-4-12B-it-heretic-NVFP4 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.90 \
  --enable-auto-tool-choice \
  --reasoning-parser gemma4 \
  --tool-call-parser gemma4 \
  --chat-template examples/tool_chat_template_gemma4.jinja \
  --limit-mm-per-prompt '{"image": 4, "audio": 1}' \
  --async-scheduling

Tip: For text-only workloads, pass --limit-mm-per-prompt '{"image": 0, "audio": 0}' to skip vision encoder memory allocation and free up GPU memory for a longer context window.

Send requests to the server:

python
from openai import OpenAI

openai_api_key = "EMPTY"
openai_api_base = "http://<your-server-host>:8001/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

model = "xdavxd/gemma-4-12B-it-heretic-NVFP4"

messages = [
    {"role": "user", "content": "Explain quantum mechanics clearly and concisely."},
]

outputs = client.chat.completions.create(
    model=model,
    messages=messages,
)

generated_text = outputs.choices[0].message.content
print(generated_text)

Creation

This model was created by applying NVFP4 quantization with LLM Compressor, as presented in the code snippet below.

python
from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier

MODEL_ID = "./gemma-4-12B-it-heretic"
SAVE_DIR = "gemma-4-12B-it-heretic-NVFP4"

model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

recipe = QuantizationModifier(
    targets="Linear",
    scheme="NVFP4",
    ignore=["re:.*vision.*", "re:.*audio.*", "lm_head", "re:.*embed.*"],
)

oneshot(
    model=model,
    tokenizer=tokenizer,
    recipe=recipe,
    dataset="ultrachat_200k",
    splits="train_sft",
    num_calibration_samples=512,
    max_seq_length=2048,
)

model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)

Evaluation

This model was evaluated on GSM8K Platinum, MMLU-Pro, IFEval, MATH-500, AIME 2025, GPQA Diamond, and LiveCodeBench v6 using lm-evaluation-harness and lighteval, served with vLLM (OpenAI-compatible API). All evaluations were performed with thinking enabled.

Performance

(as reported by coder3101)

Table with columns: Metric, This model, Original model (google/gemma-4-12B-it)
Metric	This model	Original model (google/gemma-4-12B-it)
KL divergence	0.1576	0 (by definition)
Refusals	7/100	99/100

Accuracy

Original and RedHatAI benchmarks taken from RedHatAI, comparison to see how the heretic model converted to NVFP4 stacks up against the non-heretic versions. Recovery vs. Original to see how the combination of heretic ARA (Arbitrary-Rank Ablation) and then quantization affects accuracy.

Reproduction

The results were obtained using the following commands:

Each benchmark was run 3 times with different judgmental seeds (1234, 2345, 3456) and the scores were averaged; AIME 2025 used 8 seeds.

Ran using a docker image, you will need to tweak these for your own environment. Kept parameters similar to the RedHatAI runs.

vLLM server (all benchmarks):

Note: The GB10 image I was using did not have a new enough transformers version, command below runs an in-container upgrade before launching vLLM to get around the error.

markdown
sudo docker run --rm -it \
  --gpus all \
  --ipc=host \
  --network host \
  -v ~/quant-workspace/gemma-4-12B-it-heretic-NVFP4:/models/heretic:ro \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -e VLLM_USE_V2_MODEL_RUNNER=1 \
  ghcr.io/timothystewart6/vllm-gb10:latest \
  bash -c "pip install -q --upgrade transformers && vllm serve /models/heretic \
  --host 0.0.0.0 --port 8001 \
  --served-model-name heretic-12b \
  --max-model-len 69632 \
  --gpu-memory-utilization 0.50 \
  --enable-auto-tool-choice \
  --reasoning-parser gemma4 \
  --tool-call-parser gemma4 \
  --limit-mm-per-prompt '{"image":0,"audio":0}' \
  --kv-cache-dtype fp8 \
  --max-num-seqs 32 \
  --async-scheduling"

GSM8K Platinum (lm-eval, 0-shot, 3 repetitions)

markdown
for SEED in 1234 2345 3456; do
  lm_eval --model local-chat-completions \
    --tasks gsm8k_platinum_cot_llama \
    --model_args "model=heretic,max_length=36096,base_url=http://0.0.0.0:8001/v1/chat/completions,num_concurrent=32,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=2400" \
    --num_fewshot 0 \
    --apply_chat_template \
    --output_path "results_gsm8k_seed${SEED}.json" \
    --seed $SEED \
    --gen_kwargs "do_sample=True,temperature=1.0,top_p=0.95,top_k=64,max_gen_toks=32000,seed=${SEED}"
  echo "Seed $SEED complete"
done

IFEval (lm-eval, 0-shot, 3 repetitions)

markdown
for SEED in 1234 2345 3456; do
  lm_eval --model local-chat-completions \
    --tasks ifeval \
    --model_args "model=heretic,max_length=36096,base_url=http://0.0.0.0:8001/v1/chat/completions,num_concurrent=32,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=2400" \
    --num_fewshot 0 \
    --apply_chat_template \
    --output_path "results_ifeval_seed${SEED}.json" \
    --seed $SEED \
    --gen_kwargs "do_sample=True,temperature=1.0,top_p=0.95,top_k=64,max_gen_toks=32000,seed=${SEED}"
  echo "IFEval seed $SEED complete"
done

MATH-500, (lighteval, 3 repetitions)

litellm_config.yaml:

yaml
model_parameters:
  provider: hosted_vllm
  model_name: hosted_vllm/heretic
  base_url: http://0.0.0.0:8001/v1
  api_key: ''
  timeout: 3600
  concurrent_requests: 32
  generation_parameters:
    temperature: 1.0
    max_new_tokens: 65536
    top_p: 0.95
    top_k: 64
    seed: 1234

markdown
# MATH-500 (3 seeds)
for SEED in 1234 2345 3456; do
  sed -i "s/seed: .*/seed: $SEED/" litellm_config.yaml
  lighteval endpoint litellm litellm_config.yaml 'math_500|0' \
    --output-dir "results_math500_seed${SEED}/" --save-details
  echo "MATH-500 seed $SEED complete"
done

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider