You can deploy this model and start earning money today!

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

Name: gemma-4-12B-it-FP8-Dynamic
Availability: InStock

RedHatAI/gemma-4-12B-it

quantized

apache-2.0

Views

No views yet

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

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

python

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

GSM8K Platinum
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%  |
+---------------+------------------+--------------+---------------+----------+

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

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

Creation Script

Preliminary Evaluations

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