gemma-3-4b-it-quantized.w8a8

Model Overview

• Model Architecture: google/gemma-3-4b-it
  • Input: Vision-Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: INT8
  • Activation quantization: INT8
• Release Date: 6/4/2025
• Version: 1.0
• Model Developers: RedHatAI

Quantized version of google/gemma-3-4b-it.

Model Optimizations

This model was obtained by quantizing the weights of google/gemma-3-4b-it to INT8 data type, ready for inference with vLLM >= 0.8.0.

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

from vllm import LLM, SamplingParams
from vllm.assets.image import ImageAsset
from transformers import AutoProcessor

# Define model name once
model_name = "RedHatAI/gemma-3-4b-it-quantized.w8a8"

# Load image and processor
image = ImageAsset("cherry_blossom").pil_image.convert("RGB")
processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)

# Build multimodal prompt
chat = [
    {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "What is the content of this image?"}]},
    {"role": "assistant", "content": []}
]
prompt = processor.apply_chat_template(chat, add_generation_prompt=True)

# Initialize model
llm = LLM(model=model_name, trust_remote_code=True)

# Run inference
inputs = {"prompt": prompt, "multi_modal_data": {"image": [image]}}
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))

# Display result
print("RESPONSE:", outputs[0].outputs[0].text)

vLLM also supports OpenAI-compatible serving. See the documentation for more details.

Creation

This model was created with llm-compressor by running the code snippet below:

import base64
from io import BytesIO
import torch
from datasets import load_dataset
from transformers import AutoProcessor, Gemma3ForConditionalGeneration
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.transformers import oneshot


# Load model.
model_id = "google/gemma-3-4b-it"
model = Gemma3ForConditionalGeneration.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto",
)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "neuralmagic/calibration"
DATASET_SPLIT = {"LLM": "train[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048

# Load dataset and preprocess.
ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
ds = ds.shuffle(seed=42)

dampening_frac=0.05

def data_collator(batch):
    assert len(batch) == 1, "Only batch size of 1 is supported for calibration"
    item = batch[0]
    collated = {}
    import torch


    for key, value in item.items():
        if isinstance(value, torch.Tensor):
            collated[key] = value.unsqueeze(0)
        elif isinstance(value, list) and isinstance(value[0][0], int):
            # Handle tokenized inputs like input_ids, attention_mask
            collated[key] = torch.tensor(value)
        elif isinstance(value, list) and isinstance(value[0][0], float):
            # Handle possible float sequences
            collated[key] = torch.tensor(value)
        elif isinstance(value, list) and isinstance(value[0][0], torch.Tensor):
            # Handle batched image data (e.g., pixel_values as [C, H, W])
            collated[key] = torch.stack(value)  # -> [1, C, H, W]
        elif isinstance(value, torch.Tensor):
            collated[key] = value
        else:
            print(f"[WARN] Unrecognized type in collator for key={key}, type={type(value)}")
    
    return collated
   


# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        ignore=["re:.*lm_head.*", "re:.*embed_tokens.*", "re:vision_tower.*", "re:multi_modal_projector.*"],
        sequential_update=True,
        sequential_targets=["Gemma3DecoderLayer"],
        dampening_frac=dampening_frac,
    )
]

SAVE_DIR=f"{model_id.split('/')[1]}-quantized.w8a8"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=ds,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    output_dir=SAVE_DIR
)

Evaluation

The model was evaluated using lm_evaluation_harness for OpenLLM v1 text benchmark. The evaluations were conducted using the following commands:

OpenLLM v1

lm_eval \
  --model vllm \
  --model_args pretrained="<model_name>",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=<n>,gpu_memory_utilization=0.8,enable_chunked_prefill=True,trust_remote_code=True,enforce_eager=True \
  --tasks openllm \
  --batch_size auto

Accuracy