Model Architecture: Qwen/Qwen2-VL-72B-Instruct
- Input: Vision-Text
- Output: Text
Model Optimizations:
- Weight quantization: INT4
- Activation quantization: FP16
Release Date: 2/24/2025
Version: 1.0
Model Developers: Neural Magic

Quantized version of Qwen/Qwen2-VL-72B-Instruct.

Model Optimizations

This model was obtained by quantizing the weights of Qwen/Qwen2-VL-72B-Instruct to INT8 data type, ready for inference with vLLM >= 0.5.2.

Deployment

Use with vLLM

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

python

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

# prepare model
llm = LLM(
    model="neuralmagic/Qwen2-VL-72B-Instruct-quantized.w4a16",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")

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 as part a multimodal announcement blog.

Model Creation Code

python

import base64
from io import BytesIO
import torch
from datasets import load_dataset
from qwen_vl_utils import process_vision_info
from transformers import AutoProcessor
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.transformers import oneshot
from llmcompressor.transformers.tracing import TraceableQwen2VLForConditionalGeneration
from llmcompressor.transformers.utils.data_collator import qwen2_vl_data_collator
from compressed_tensors.quantization import QuantizationArgs, QuantizationType, QuantizationStrategy, ActivationOrdering, QuantizationScheme

# Load model.
model_id = "Qwen/Qwen2-VL-72B-Instruct"

model = TraceableQwen2VLForConditionalGeneration.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto",
)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "lmms-lab/flickr30k"
DATASET_SPLIT = {"calibration": "test[: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=0.01

# Apply chat template and tokenize inputs.
def preprocess_and_tokenize(example):
    # preprocess
    buffered = BytesIO()
    example["image"].save(buffered, format="PNG")
    encoded_image = base64.b64encode(buffered.getvalue())
    encoded_image_text = encoded_image.decode("utf-8")
    base64_qwen = f"data:image;base64,{encoded_image_text}"
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "image": base64_qwen},
                {"type": "text", "text": "What does the image show?"},
            ],
        }
    ]
    text = processor.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    image_inputs, video_inputs = process_vision_info(messages)

    # tokenize
    return processor(
        text=[text],
        images=image_inputs,
        videos=video_inputs,
        padding=False,
        max_length=MAX_SEQUENCE_LENGTH,
        truncation=True,
    )

ds = ds.map(preprocess_and_tokenize, remove_columns=ds["calibration"].column_names)

# Recipe
recipe = GPTQModifier(
    targets="Linear",
    config_groups={
        "config_group": QuantizationScheme(
            targets=["Linear"],
            weights=QuantizationArgs(
                num_bits=4,
                type=QuantizationType.INT,
                strategy=QuantizationStrategy.GROUP,
                group_size=128,
                symmetric=True,
                dynamic=False,
                actorder=ActivationOrdering.WEIGHT,
            ),
        ),
    },
    sequential_targets=["Qwen2VLDecoderLayer"],
    ignore=["lm_head", "re:visual.*"],
    update_size=NUM_CALIBRATION_SAMPLES,
    dampening_frac=dampening_frac
)

SAVE_DIR=f"{model_id.split('/')[1]}-quantized.w4a16

# 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=qwen2_vl_data_collator,
    output_dir=SAVE_DIR
)

Evaluation

The model was evaluated using mistral-evals for vision-related tasks and using lm_evaluation_harness for select text-based benchmarks. The evaluations were conducted using the following commands:

Evaluation Commands

Vision Tasks

vqav2
docvqa
mathvista
mmmu
chartqa

vllm serve neuralmagic/pixtral-12b-quantized.w8a8 --tensor_parallel_size 1 --max_model_len 25000 --trust_remote_code --max_num_seqs 8 --gpu_memory_utilization 0.9 --dtype float16 --limit_mm_per_prompt image=7

python -m eval.run eval_vllm \
        --model_name neuralmagic/pixtral-12b-quantized.w8a8 \
        --url http://0.0.0.0:8000 \
        --output_dir ~/tmp \
        --eval_name <vision_task_name>

Text-based Tasks

MMLU

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 \
  --tasks mmlu \
  --num_fewshot 5 \
  --batch_size auto \
  --output_path output_dir

MGSM

lm_eval \
  --model vllm \
  --model_args pretrained="<model_name>",dtype=auto,max_model_len=4096,max_gen_toks=2048,max_num_seqs=128,tensor_parallel_size=<n>,gpu_memory_utilization=0.9 \
  --tasks mgsm_cot_native \
  --apply_chat_template \
  --num_fewshot 0 \
  --batch_size auto \
  --output_path output_dir

Accuracy

Category	Metric	Qwen/Qwen2-VL-72B-Instruct	nm-testing/Qwen2-VL-72B-Instruct-quantized.W4A16	Recovery (%)
Vision	MMMU (val, CoT) explicit_prompt_relaxed_correctness	62.11	60.11	96.78%
VQAv2 (val) vqa_match	82.51	82.38	99.84%
DocVQA (val) anls	95.01	94.94	99.93%
ChartQA (test, CoT) anywhere_in_answer_relaxed_correctness	83.40	80.72	96.78%
Mathvista (testmini, CoT) explicit_prompt_relaxed_correctness	66.57	64.66	97.13%
Average Score	77.92	76.56	98.26
Text	MGSM (CoT)	68.60	66.45	96.87%
MMLU (5-shot)	82.70	82.35	99.58%

Inference Performance

This model achieves up to 3.7x speedup in single-stream deployment and up to 3.3x speedup in multi-stream asynchronous deployment, depending on hardware and use-case scenario. The following performance benchmarks were conducted with vLLM version 0.7.2, and GuideLLM.

Benchmarking Command

``` guidellm --model neuralmagic/Qwen2-VL-72B-Instruct-quantized.w4a16 --target "http://localhost:8000/v1" --data-type emulated --data prompt_tokens=,generated_tokens=,images=,width=,height= --max seconds 120 --backend aiohttp_server ```

Single-stream performance (measured with vLLM version 0.7.2)

			Document Visual Question Answering 1680W x 2240H 64/128	Visual Reasoning 640W x 480H 128/128	Image Captioning 480W x 360H 0/128
Hardware	Number of GPUs	Model	Average Cost Reduction	Latency (s)	QPD	Latency (s)th>	QPD	Latency (s)	QPD
A100	4	Qwen/Qwen2-VL-72B-Instruct		6.5	77	4.6	110	4.4	113
2	neuralmagic/Qwen2-VL-72B-Instruct-quantized.w8a8	1.85	7.2	139	4.9	206	4.8	211
1	neuralmagic/Qwen2-VL-72B-Instruct-quantized.w4a16	3.32	10.0	202	5.0	398	4.8	419
H100	4	Qwen/Qwen2-VL-72B-Instruct		4.4	66	3.0	97	2.9	99
2	neuralmagic/Qwen2-VL-72B-Instruct-FP8-Dynamic	1.79	4.7	119	3.3	173	3.2	177
1	neuralmagic/Qwen2-VL-72B-Instruct-quantized.w4a16	2.60	6.4	172	4.3	253	4.2	259

**Use case profiles: Image Size (WxH) / prompt tokens / generation tokens

**QPD: Queries per dollar, based on on-demand cost at Lambda Labs (observed on 2/18/2025).

Multi-stream asynchronous performance (measured with vLLM version 0.7.2)

			Document Visual Question Answering 1680W x 2240H 64/128	Visual Reasoning 640W x 480H 128/128	Image Captioning 480W x 360H 0/128
Hardware	Model	Average Cost Reduction	Maximum throughput (QPS)	QPD	Maximum throughput (QPS)	QPD	Maximum throughput (QPS)	QPD
A100x4	Qwen/Qwen2-VL-72B-Instruct		0.3	169	1.1	538	1.2	595
neuralmagic/Qwen2-VL-72B-Instruct-quantized.w8a8	1.84	0.6	293	2.0	1021	2.3	1135
neuralmagic/Qwen2-VL-72B-Instruct-quantized.w4a16	2.73	0.6	314	3.2	1591	4.0	2019
H100x4	Qwen/Qwen2-VL-72B-Instruct		0.5	137	1.2	356	1.3	377
neuralmagic/Qwen2-VL-72B-Instruct-FP8-Dynamic	1.70	0.8	236	2.2	623	2.4	669
neuralmagic/Qwen2-VL-72B-Instruct-quantized.w4a16	2.35	1.3	350	3.3	910	3.6	994