Model Architecture: Qwen3VLForConditionalGeneration
- Input: Text, Image
- Output: Text
Model Optimizations:
- Weight quantization: FP4
- Activation quantization: FP4
Release Date:
Version: 1.0
Model Developers:: Red Hat

Quantized version of Qwen/Qwen3-VL-32B-Instruct.

Model Optimizations

This model was obtained by quantizing the weights and activations of Qwen/Qwen3-VL-32B-Instruct to FP8 data type. This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 75%. Only the weights and activations of the linear operators within transformers blocks of the language model are quantized.

Deployment

Use with vLLM

Initialize vLLM server:

vllm serve RedHatAI/Qwen3-VL-32B-Instruct-NVFP4 --tensor_parallel_size 2

Send requests to the server:

python

from openai import OpenAI

# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://<your-server-host>:8000/v1"

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

model = "RedHatAI/Qwen3-VL-32B-Instruct-NVFP4"

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"},
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

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

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

Creation

This model was quantized using the llm-compressor library as shown below.

Creation details

python

import torch
from datasets import load_dataset
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.utils import dispatch_for_generation

MODEL_ID = "Qwen/Qwen3-VL-32B-Instruct"

# Load model.
model = Qwen3VLForConditionalGeneration.from_pretrained(MODEL_ID, torch_dtype="auto")
processor = AutoProcessor.from_pretrained(MODEL_ID)

DATASET_ID = "neuralmagic/calibration"
NUM_CALIBRATION_SAMPLES = 20
MAX_SEQUENCE_LENGTH = 8192

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


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

    return processor.apply_chat_template(
        messgages,
        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()
    }


# Configure the quantization algorithm and scheme.
# In this case, we:
#   * quantize the weights to fp4 with group-wise quantization
#   * quantize the activations to fp4 with dynamic group activations
recipe = QuantizationModifier(
    targets="Linear",
    scheme="NVFP4",
    ignore=[
        "re:.*lm_head",
        "re:visual.*",
        "re:model.visual.*",
        "re:.*mlp.gate$",
    ],
)

# Apply quantization.
oneshot(
    model=model,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    dataset=ds,
    data_collator=data_collator,
)

print("========== SAMPLE GENERATION ==============")
dispatch_for_generation(model)
input_ids = processor(text="Hello my name is", return_tensors="pt").input_ids.to("cuda")
output = model.generate(input_ids, max_new_tokens=20)
print(processor.decode(output[0]))
print("==========================================")


# Save to disk in compressed-tensors format.
SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-NVFP4"
model.save_pretrained(SAVE_DIR)
processor.save_pretrained(SAVE_DIR)

Evaluation

The model was evaluated on the OpenLLMv1 leaderboard task, using lm-evaluation-harness. vLLM was used for all evaluations.

Evaluation details

ChartQA

lm_eval \
  --model vllm-vlm \
  --model_args pretrained="RedHatAI/Qwen3-VL-32B-Instruct-NVFP4",dtype=auto,add_bos_token=False,max_model_len=262144,tensor_parallel_size=2,gpu_memory_utilization=0.9,enable_chunked_prefill=True,trust_remote_code=True,max_images=10 \
  --tasks chartqa \
  --apply_chat_template \
  --batch_size auto

MMLU

lm_eval \
  --model vllm-vlm \
  --model_args pretrained="RedHatAI/Qwen3-VL-32B-Instruct-NVFP4",dtype=auto,add_bos_token=False,max_model_len=262144,tensor_parallel_size=2,gpu_memory_utilization=0.9,enable_chunked_prefill=True,trust_remote_code=True,max_images=10 \
  --tasks mmlu \
  --apply_chat_template \
  --batch_size auto

Accuracy Comparison

ChartQA Results

Model	Accuracy	Recovery (%)
Qwen/Qwen3-VL-32B-Instruct	61.52	100.00
Qwen/Qwen3-VL-32B-Instruct-FP8	86.92	141.32
RedHatAI/Qwen3-VL-32B-Instruct-FP8-block	86.60	140.82
RedHatAI/Qwen3-VL-32B-Instruct-FP8-dynamic	86.68	140.95
RedHatAI/Qwen3-VL-32B-Instruct-NVFP4	86.56	140.7

MMLU Results

Model	Accuracy	Recovery (%)
Qwen/Qwen3-VL-32B-Instruct	78.03	100.00
Qwen/Qwen3-VL-32B-Instruct-FP8	77.80	99.71
RedHatAI/Qwen3-VL-32B-Instruct-FP8-block	77.72	99.60
RedHatAI/Qwen3-VL-32B-Instruct-FP8-dynamic	77.89	99.82
RedHatAI/Qwen3-VL-32B-Instruct-NVFP4	76.27	97.74