nm-testing/Qwen2.5-VL-72B-Instruct-quantized.w4a16

Qwen2.5-VL-72B-Instruct-quantized-w4a16

Model Overview

• Model Architecture: Qwen/Qwen2.5-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/Qwen2.5-VL-72B-Instruct.

Model Optimizations

This model was obtained by quantizing the weights of Qwen/Qwen2.5-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.

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

# prepare model
llm = LLM(
    model="neuralmagic/Qwen2.5-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

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 (
    TraceableQwen2_5_VLForConditionalGeneration,
)
from compressed_tensors.quantization import QuantizationArgs, QuantizationType, QuantizationStrategy, ActivationOrdering, QuantizationScheme

# Load model.
model_id = "Qwen/Qwen2.5-VL-72B-Instruct"

model = TraceableQwen2_5_VLForConditionalGeneration.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)

# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

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=["Qwen2_5_VLDecoderLayer"],
    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=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 \
  --num_fewshot 0 \
  --batch_size auto \
  --output_path output_dir

Accuracy

Category	Metric	Qwen/Qwen2.5-VL-72B-Instruct	neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w4a16	Recovery (%)
Vision	MMMU (val, CoT) explicit_prompt_relaxed_correctness	64.33	62.89	97.76%
	VQAv2 (val) vqa_match	81.94	81.87	99.91%
	DocVQA (val) anls	94.71	94.72	100.01%
	ChartQA (test, CoT) anywhere_in_answer_relaxed_correctness	88.96	88.96	100.00%
	Mathvista (testmini, CoT) explicit_prompt_relaxed_correctness	78.18	77.68	99.36%
	Average Score	81.62	81.22	99.51
Text	MGSM (CoT)	75.45	75.13	99.58%
	MMLU (5-shot)	86.16	85.36	99.07%

Inference Performance

This model achieves up to xxx speedup in single-stream deployment and up to xxx 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/Qwen2.5-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)	Queries Per Dollar	Latency (s)th>	Queries Per Dollar	Latency (s)	Queries Per Dollar
A100	4	Qwen/Qwen2.5-VL-72B-Instruct		6.4	78	4.5	111	4.4	113
	2	neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w8a8	1.85	7.0	143	4.9	205	4.8	211
	1	neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w4a16	3.33	9.4	213	5.1	396	4.8	420
H100	4	Qwen/Qwen2.5-VL-72B-Instruct		4.3	68	3.0	97	2.9	100
	2	neuralmagic/Qwen2.5-VL-72B-Instruct-FP8-Dynamic	1.79	4.6	122	3.3	173	3.2	177
	1	neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w4a16	5.66	4.3	252	4.3	252	1.0	1065

**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)	Queries Per Dollar	Maximum throughput (QPS)	Queries Per Dollar	Maximum throughput (QPS)	Queries Per Dollar
A100x4	Qwen/Qwen2.5-VL-72B-Instruct		0.4	180	1.1	539	1.2	595
	neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w8a8	1.80	1.2	578	4.0	2040	4.6	2266
	neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w4a16	2.75	2.8	1364	12.8	6352	16.4	8148
H100x4	Qwen/Qwen2.5-VL-72B-Instruct		0.5	134	1.2	357	1.3	379
	neuralmagic/Qwen2.5-VL-72B-Instruct-FP8-Dynamic	1.73	1.8	479	4.4	1203	4.8	1296
	neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w4a16	8.27	13.2	3652	13.2	3652	99.2	27108

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

**QPS: Queries per second.

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