Pixtral-Large-Instruct-2411-hf-quantized.w4a16

Pixtral-Large-Instruct-2411-hf-quantized.w4a16

Model Overview

• Model Architecture: neuralmagic/Pixtral-Large-Instruct-2411-hf
  • 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 neuralmagic/Pixtral-Large-Instruct-2411-hf.

Model Optimizations

This model was obtained by quantizing the weights of neuralmagic/Pixtral-Large-Instruct-2411-hf to INT4 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/Pixtral-Large-Instruct-2411-hf-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>

import requests
import torch
from PIL import Image
from transformers import AutoProcessor
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.transformers import oneshot
from llmcompressor.transformers.tracing import TraceableLlavaForConditionalGeneration
from compressed_tensors.quantization import QuantizationArgs, QuantizationType, QuantizationStrategy, ActivationOrdering, QuantizationScheme

    
# Load model.
model_id = "neuralmagic/Pixtral-Large-Instruct-2411-hf"
model = TraceableLlavaForConditionalGeneration.from_pretrained(
    model_id, device_map="auto", torch_dtype="auto"
)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048
dampening_frac=0.01

# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {
        "input_ids": torch.LongTensor(batch[0]["input_ids"]),
        "attention_mask": torch.tensor(batch[0]["attention_mask"]),
        "pixel_values": torch.tensor(batch[0]["pixel_values"]),
    }

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=["MistralDecoderLayer"],
    ignore=["re:.*lm_head", "re:vision_tower.*", "re:multi_modal_projector.*"],
    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=DATASET_ID,
    splits=DATASET_SPLIT,
    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
)

</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> </details>

Accuracy

<table> <thead> <tr> <th>Category</th> <th>Metric</th> <th>neuralmagic/Pixtral-Large-Instruct-2411-hf</th> <th>neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w4a16</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>63.56</td> <td>60.56</td> <td>95.28%</td> </tr> <tr> <td>VQAv2 (val)<br><i>vqa_match</i></td> <td>79.03</td> <td>79.04</td> <td>100.01%</td> </tr> <tr> <td>DocVQA (val)<br><i>anls</i></td> <td>89.55</td> <td>89.00</td> <td>99.39%</td> </tr> <tr> <td>ChartQA (test, CoT)<br><i>anywhere_in_answer_relaxed_correctness</i></td> <td>82.24</td> <td>81.52</td> <td>99.12%</td> </tr> <tr> <td>Mathvista (testmini, CoT)<br><i>explicit_prompt_relaxed_correctness</i></td> <td>67.3</td> <td>66.60</td> <td>98.96%</td> </tr> <tr> <td><b>Average Score</b></td> <td><b>76.34</b></td> <td><b>75.34</b></td> <td><b>98.69%</b></td> </tr> <tr> <td rowspan="2"><b>Text</b></td> <td>MGSM (CoT)</td> <td>76.05</td> <td>75.09</td> <td>98.74%</td> </tr> <tr> <td>MMLU (5-shot)</td> <td>82.8</td> <td>82.25</td> <td>99.33%</td> </tr> </tbody> </table>

Inference Performance

This model achieves up to 2.80x speedup in single-stream deployment and up to 1.75x 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/Pixtral-Large-Instruct-2411-hf-quantized.w4a16 --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></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>Queries Per Dollar</th> <th>Latency (s)</th> <th>Queries Per Dollar</th> <th>Latency (s)</th> <th>Queries Per Dollar</th> </tr> </thead> <tbody style="text-align: center"> <tr> <th rowspan="3" valign="top">A100</th> <td>4</td> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf</td> <td></td> <td>7.5</td> <td>67</td> <td>6.5</td> <td>77</td> <td>6.4</td> <td>79</td> </tr> <tr> <td>2</td> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w8a8</td> <td>1.86</td> <td>8.1</td> <td>124</td> <td>7.1</td> <td>142</td> <td>6.8</td> <td>148</td> </tr> <tr> <td>2</td> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w4a16</td> <td>2.52</td> <td>6.9</td> <td>147</td> <td>5.1</td> <td>199</td> <td>4.5</td> <td>221</td> </tr> <tr> <th rowspan="3" valign="top">H100</th> <td>4</td> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf</td> <td></td> <td>4.4</td> <td>67</td> <td>3.9</td> <td>74</td> <td>3.7</td> <td>79</td> </tr> <tr> <td>2</td> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf-FP8-Dynamic</td> <td>1.82</td> <td>4.7</td> <td>120</td> <td>4.1</td> <td>137</td> <td>3.9</td> <td>145</td> </tr> <tr> <td>2</td> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w4a16</td> <td>1.87</td> <td>4.7</td> <td>120</td> <td>3.9</td> <td>144</td> <td>3.8</td> <td>149</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>Queries Per Dollar</th> <th>Maximum throughput (QPS)</th> <th>Queries Per Dollar</th> <th>Maximum throughput (QPS)</th> <th>Queries Per Dollar</th> </tr> </thead> <tbody style="text-align: center"> <tr> <th rowspan="3" valign="top">A100x4</th> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf</td> <td></td> <td>0.4</td> <td>222</td> <td>0.7</td> <td>341</td> <td>0.8</td> <td>399</td> </tr> <tr> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w8a8</td> <td>1.70</td> <td>0.8</td> <td>383</td> <td>1.1</td> <td>571</td> <td>1.3</td> <td>674</td> </tr> <tr> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w4a16</td> <td>1.48</td> <td>0.5</td> <td>276</td> <td>1.0</td> <td>505</td> <td>1.4</td> <td>680</td> </tr> <tr> <<th rowspan="3" valign="top">H100x4</th> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf</td> <td></td> <td>1.0</td> <td>284</td> <td>1.6</td> <td>465</td> <td>1.8</td> <td>511</td> </tr> <tr> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf-FP8-Dynamic</td> <td>1.61</td> <td>1.7</td> <td>467</td> <td>2.6</td> <td>726</td> <td>3.2</td> <td>908</td> </tr> <tr> <td>neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w4a16</td> <td>1.33</td> <td>1.4</td> <td>393</td> <td>2.2</td> <td>726</td> <td>2.7</td> <td>764</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).