Model Architecture: neuralmagic/Pixtral-Large-Instruct-2411-hf
- Input: Vision-Text
- Output: Text
Model Optimizations:
- Weight quantization: INT8
- Activation quantization: INT8
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 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/Pixtral-Large-Instruct-2411-hf-quantized.w8a8",
    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 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
        
# 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


# 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
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        sequential_targets=["MistralDecoderLayer"],
        ignore=["re:.*lm_head", "re:vision_tower.*", "re:multi_modal_projector.*"],
    ),
]

SAVE_DIR==f"{model_id.split('/')[1]}-quantized.w8a8"

# 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
)

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	neuralmagic/Pixtral-Large-Instruct-2411-hf	neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w8a8	Recovery (%)
Vision	MMMU (val, CoT) explicit_prompt_relaxed_correctness	63.56	63.89	100.52%
VQAv2 (val) vqa_match	79.03	79.12	100.11%
DocVQA (val) anls	89.55	89.80	100.28%
ChartQA (test, CoT) anywhere_in_answer_relaxed_correctness	82.24	80.44	97.81%
Mathvista (testmini, CoT) explicit_prompt_relaxed_correctness	67.3	66.50	98.81%
Average Score	76.34	75.95	99.49%
Text	MGSM (CoT)	76.05	74.76	98.30%
MMLU (5-shot)	82.8	82.9	100.12%

Inference Performance

This model achieves up to 1.87x speedup in single-stream deployment and up to 2.0x 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/Pixtral-Large-Instruct-2411-hf-quantized.w8a8 --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)	Queries Per Dollar	Latency (s)	Queries Per Dollar
A100	4	neuralmagic/Pixtral-Large-Instruct-2411-hf		7.5	67	6.5	77	6.4	79
2	neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w8a8	1.86	8.1	124	7.1	142	6.8	148
2	neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w4a16	2.52	6.9	147	5.1	199	4.5	221
H100	4	neuralmagic/Pixtral-Large-Instruct-2411-hf		4.4	67	3.9	74	3.7	79
2	neuralmagic/Pixtral-Large-Instruct-2411-hf-FP8-Dynamic	1.82	4.7	120	4.1	137	3.9	145
2	neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w4a16	1.87	4.7	120	3.9	144	3.8	149

**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	neuralmagic/Pixtral-Large-Instruct-2411-hf		0.4	222	0.7	341	0.8	399
neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w8a8	1.70	0.8	383	1.1	571	1.3	674
neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w4a16	1.48	0.5	276	1.0	505	1.4	680
H100x4	neuralmagic/Pixtral-Large-Instruct-2411-hf		1.0	284	1.6	465	1.8	511
neuralmagic/Pixtral-Large-Instruct-2411-hf-FP8-Dynamic	1.61	1.7	467	2.6	726	3.2	908
neuralmagic/Pixtral-Large-Instruct-2411-hf-quantized.w4a16	1.33	1.4	393	2.2	726	2.7	764