Qwen2-VL-72B-Instruct-FP8-dynamic

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

Model Overview

• Model Architecture: Qwen/Qwen2-VL-72B-Instruct
  • Input: Vision-Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: FP8
  • Activation quantization: FP8
• 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 FP8 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-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.

<details> <summary>Model Creation Code</summary>

from transformers import AutoProcessor, Qwen2VLForConditionalGeneration

from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers import oneshot, wrap_hf_model_class

MODEL_ID = "Qwen/Qwen2-VL-72B-Instruct"

# Load model.
model_class = wrap_hf_model_class(Qwen2VLForConditionalGeneration)
model = model_class.from_pretrained(MODEL_ID, device_map="auto", torch_dtype="auto")
processor = AutoProcessor.from_pretrained(MODEL_ID)

# Configure the quantization algorithm and scheme.
# In this case, we:
#   * quantize the weights to fp8 with per channel via ptq
#   * quantize the activations to fp8 with dynamic per token
recipe = QuantizationModifier(
    targets="Linear",
    scheme="FP8_DYNAMIC",
    ignore=["re:.*lm_head", "re:visual.*"],
)

# Apply quantization and save to disk in compressed-tensors format.
SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-dynamic"
oneshot(model=model, recipe=recipe, output_dir=SAVE_DIR)
processor.save_pretrained(SAVE_DIR)

# Confirm generations of the quantized model look sane.
print("========== SAMPLE GENERATION ==============")
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("==========================================")

</details>

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:

<details> <summary>Evaluation Commands</summary>

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

</details>

Accuracy

<table> <thead> <tr> <th>Category</th> <th>Metric</th> <th>Qwen/Qwen2-VL-72B-Instruct</th> <th>neuralmagic/Qwen2-VL-72B-Instruct-FP8-Dynamic</th> <th>Recovery (%)</th> </tr> </thead> <tbody> <tr> <td rowspan="6"><b>Vision</b></td> <td>MMMU (val, CoT)<br><i>explicit_prompt_relaxed_correctness</i></td> <td>62.11</td> <td>60.67</td> <td>97.68%</td> </tr> <tr> <td>VQAv2 (val)<br><i>vqa_match</i></td> <td>82.51</td> <td>82.44</td> <td>99.91%</td> </tr> <tr> <td>DocVQA (val)<br><i>anls</i></td> <td>95.01</td> <td>95.10</td> <td>100.09%</td> </tr> <tr> <td>ChartQA (test, CoT)<br><i>anywhere_in_answer_relaxed_correctness</i></td> <td>83.40</td> <td>83.68</td> <td>100.34%</td> </tr> <tr> <td>Mathvista (testmini, CoT)<br><i>explicit_prompt_relaxed_correctness</i></td> <td>66.57</td> <td>67.07</td> <td>100.75%</td> </tr> <tr> <td><b>Average Score</b></td> <td><b>77.12</b></td> <td><b>77.39</b></td> <td><b>100.35%</b></td> </tr> <tr> <td rowspan="2"><b>Text</b></td> <td>MGSM (CoT)</td> <td>68.60</td> <td>67.78</td> <td>98.80%</td> </tr> <tr> <td>MMLU (5-shot)</td> <td>82.70</td> <td>82.60</td> <td>99.88%</td> </tr> </tbody> </table>

Inference Performance

This model achieves up to 1.84x speedup in single-stream deployment and up to 1.85x 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.

<details> <summary>Benchmarking Command</summary> ``` guidellm --model neuralmagic/Qwen2-VL-72B-Instruct-FP8-Dynamic --target "http://localhost:8000/v1" --data-type emulated --data prompt_tokens=<prompt_tokens>,generated_tokens=<generated_tokens>,images=<num_images>,width=<image_width>,height=<image_height> --max seconds 120 --backend aiohttp_server ``` </details>

Single-stream performance (measured with vLLM version 0.7.2)

<table border="1" class="dataframe"> <thead> <tr> <th></th> <th></th> <th></th> <th style="text-align: center;" colspan="2" >Document Visual Question Answering<br>1680W x 2240H<br>64/128</th> <th style="text-align: center;" colspan="2" >Visual Reasoning <br>640W x 480H<br>128/128</th> <th style="text-align: center;" colspan="2" >Image Captioning<br>480W x 360H<br>0/128</th> </tr> <tr> <th>Hardware</th> <th>Number of GPUs</th> <th>Model</th> <th>Average Cost Reduction</th> <th>Latency (s)</th> <th>QPD</th> <th>Latency (s)th> <th>QPD</th> <th>Latency (s)</th> <th>QPD</th> </tr> </thead> <tbody> <tr> <th rowspan="3" valign="top">A100</th> <td>4</td> <td>Qwen/Qwen2-VL-72B-Instruct</td> <td></td> <td>6.5</td> <td>77</td> <td>4.6</td> <td>110</td> <td>4.4</td> <td>113</td> </tr> <tr> <td>2</td> <td>neuralmagic/Qwen2-VL-72B-Instruct-quantized.w8a8</td> <td>1.85</td> <td>7.2</td> <td>139</td> <td>4.9</td> <td>206</td> <td>4.8</td> <td>211</td> </tr> <tr> <td>1</td> <td>neuralmagic/Qwen2-VL-72B-Instruct-quantized.w4a16</td> <td>3.32</td> <td>10.0</td> <td>202</td> <td>5.0</td> <td>398</td> <td>4.8</td> <td>419</td> </tr> <tr> <th rowspan="3" valign="top">H100</td> <td>4</td> <td>Qwen/Qwen2-VL-72B-Instruct</td> <td></td> <td>4.4</td> <td>66</td> <td>3.0</td> <td>97</td> <td>2.9</td> <td>99</td> </tr> <tr> <td>2</td> <td>neuralmagic/Qwen2-VL-72B-Instruct-FP8-Dynamic</td> <td>1.79</td> <td>4.7</td> <td>119</td> <td>3.3</td> <td>173</td> <td>3.2</td> <td>177</td> </tr> <tr> <td>1</td> <td>neuralmagic/Qwen2-VL-72B-Instruct-quantized.w4a16</td> <td>2.60</td> <td>6.4</td> <td>172</td> <td>4.3</td> <td>253</td> <td>4.2</td> <td>259</td> </tr> </tbody> </table>

**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)

<table border="1" class="dataframe"> <thead> <tr> <th></th> <th></th> <th></th> <th style="text-align: center;" colspan="2" >Document Visual Question Answering<br>1680W x 2240H<br>64/128</th> <th style="text-align: center;" colspan="2" >Visual Reasoning <br>640W x 480H<br>128/128</th> <th style="text-align: center;" colspan="2" >Image Captioning<br>480W x 360H<br>0/128</th> </tr> <tr> <th>Hardware</th> <th>Model</th> <th>Average Cost Reduction</th> <th>Maximum throughput (QPS)</th> <th>QPD</th> <th>Maximum throughput (QPS)</th> <th>QPD</th> <th>Maximum throughput (QPS)</th> <th>QPD</th> </tr> </thead> <tbody> <tr> <th rowspan="3" valign="top">A100x4</th> <td>Qwen/Qwen2-VL-72B-Instruct</td> <td></td> <td>0.3</td> <td>169</td> <td>1.1</td> <td>538</td> <td>1.2</td> <td>595</td> </tr> <tr> <td>neuralmagic/Qwen2-VL-72B-Instruct-quantized.w8a8</td> <td>1.84</td> <td>0.6</td> <td>293</td> <td>2.0</td> <td>1021</td> <td>2.3</td> <td>1135</td> </tr> <tr> <td>neuralmagic/Qwen2-VL-72B-Instruct-quantized.w4a16</td> <td>2.73</td> <td>0.6</td> <td>314</td> <td>3.2</td> <td>1591</td> <td>4.0</td> <td>2019</td> </tr> <tr> <th rowspan="3" valign="top">H100x4</td> <td>Qwen/Qwen2-VL-72B-Instruct</td> <td></td> <td>0.5</td> <td>137</td> <td>1.2</td> <td>356</td> <td>1.3</td> <td>377</td> </tr> <tr> <td>neuralmagic/Qwen2-VL-72B-Instruct-FP8-Dynamic</td> <td>1.70</td> <td>0.8</td> <td>236</td> <td>2.2</td> <td>623</td> <td>2.4</td> <td>669</td> </tr> <tr> <td>neuralmagic/Qwen2-VL-72B-Instruct-quantized.w4a16</td> <td>2.35</td> <td>1.3</td> <td>350</td> <td>3.3</td> <td>910</td> <td>3.6</td> <td>994</td> </tr> </tbody> </table>

**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).