RedHatAI/gemma-4-12B-it-FP8-Dynamic

NVFP4 Quantized RedHatAI/gemma-4-12B-it-NVFP4

This is a preliminary version (and subject to change) of FP8_Dynamic quantized google/gemma-4-12B-it model. The model has both weights and activations quantized to FP8_Dynamic format with vllm-project/llm-compressor.

It is compatible and tested against vllm nightly.

Creation Script

Run this script with this LLM Compressor PR and latest transformers to quantize the model using iMatrix quantization

import torch
from compressed_tensors.offload import dispatch_model
from compressed_tensors.quantization import preset_name_to_scheme
from datasets import load_dataset
from transformers import AutoModelForImageTextToText, AutoProcessor

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.modifiers.transform.imatrix import IMatrixGatherer

MODEL_ID = "google/gemma-4-12B-it"
model = AutoModelForImageTextToText.from_pretrained(MODEL_ID, dtype="auto")
processor = AutoProcessor.from_pretrained(MODEL_ID)

DATASET_ID = "neuralmagic/calibration"
NUM_CALIBRATION_SAMPLES = 256
MAX_SEQUENCE_LENGTH = 2048

ds = load_dataset(DATASET_ID, name="LLM", split=f"train[:{NUM_CALIBRATION_SAMPLES}]")


def preprocess_function(example):
    messages = []
    for message in example["messages"]:
        messages.append(
            {
                "role": message["role"],
                "content": [{"type": "text", "text": message["content"]}],
            }
        )

    return processor.apply_chat_template(
        messages,
        return_tensors="pt",
        padding=False,
        truncation=True,
        max_length=MAX_SEQUENCE_LENGTH,
        tokenize=True,
        add_special_tokens=False,
        return_dict=True,
        add_generation_prompt=False,
    )


ds = ds.map(preprocess_function, batched=False, remove_columns=ds.column_names)


def data_collator(batch):
    assert len(batch) == 1
    return {
        key: (
            torch.tensor(value)
            if key != "pixel_values"
            else torch.tensor(value, dtype=torch.bfloat16).squeeze(0)
        )
        for key, value in batch[0].items()
    }


scheme = preset_name_to_scheme("FP8_DYNAMIC", ["Linear"])
scheme.weights.observer = "imatrix_mse"

recipe = [
    IMatrixGatherer(
        ignore=[
            "lm_head",
            "re:.*embed_vision.*",
            "re:.*embed_audio.*",
            "re:.*vision_embedder.*",
        ],
    ),
    QuantizationModifier(
        config_groups={"group_0": scheme},
        ignore=[
            "lm_head",
            "re:.*embed_vision.*",
            "re:.*embed_audio.*",
            "re:.*vision_embedder.*",
        ],
    ),
]

oneshot(
    model=model,
    recipe=recipe,
    dataset=ds,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    data_collator=data_collator,
)

print("\n\n")
print("========== SAMPLE GENERATION ==============")
dispatch_model(model)
input_ids = torch.tensor(
    [[
        2, 105, 2364, 107, 818, 3282, 506, 7217, 563, 3730, 563,
        1547, 106, 107, 105, 4368, 107
    ]]
).to(model.device)
output = model.generate(
    input_ids,
    max_new_tokens=100,
)
print(processor.tokenizer.decode(output[0]))
print("==========================================\n\n")

SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8_Dynamic-iMatrix"
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)

# Patch config: transformers renames checkpoint keys on load (vision_embedder ->
# embed_vision), but save_pretrained reverts them. The ignore list in config.json
# uses HF names (embed_vision) while safetensors keys use checkpoint names
# (vision_embedder), so vllm can't match them. Add the checkpoint name explicitly.
import json as _json
_cfg_path = SAVE_DIR + "/config.json"
with open(_cfg_path) as _f:
    _cfg = _json.load(_f)
_qcfg = _cfg.get("quantization_config")
if _qcfg:
    _ign = _qcfg.setdefault("ignore", [])
    if "model.vision_embedder.patch_dense" not in _ign:
        _ign.append("model.vision_embedder.patch_dense")
        with open(_cfg_path, "w") as _f:
            _json.dump(_cfg, _f, indent=2)
        print("Patched config.json: added vision_embedder.patch_dense to ignore list")

Preliminary Evaluations

1. GSM8K Platinum
2. Wikitext PPL

lm_eval --model vllm \
  --model_args "pretrained=RedHatAI/gemma-4-12B-it-FP8_Dynamic,dtype=auto,max_model_len=$MAX_MODEL_LEN,add_bos_token=True,gpu_memory_utilization=0.85" \
  --tasks gsm8k_platinum --num_fewshot 5 --apply_chat_template --batch_size auto

lm_eval --model vllm \
  --model_args "pretrained=RedHatAI/gemma-4-12B-it-FP8_Dynamic,dtype=auto,max_model_len=$MAX_MODEL_LEN,add_bos_token=True,gpu_memory_utilization=0.85" \
  --tasks wikitext --num_fewshot 0 --apply_chat_template --batch_size auto

Evals:

+---------------+------------------+--------------+----------------+----------+
| model_name    | flexible-extract | strict-match |  bits_per_byte | byte_ppl |
+---------------+------------------+--------------+----------------+----------+
| baseline-bf16 | 0.9082           | 0.8958       | 1.9125         | 3.7645   |
| FP8-RTN       | 0.9115           | 0.8999       | 1.9368         | 3.8285   |
| *FP8-iMatrix* | 0.9198           | 0.9032       | 1.9056         | 3.7465   |
| FP8-GPTQ      | 0.9098           | 0.8950       | 1.9357         | 3.8257   |
+---------------+------------------+--------------+----------------+----------+

Recovery
+---------------+------------------+--------------+---------------+----------+
|*NVFP4-iMatrix*| 100.17%          | 100.83%      | 100.36%       | 100.48%  |
+---------------+------------------+--------------+---------------+----------+