pixtral-12b-FP8-Dynamic

Model Overview

Model Architecture: Qwen2.5-VL-3B-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/Qwen2.5-VL-3B-Instruct.

Model Optimizations

This model was obtained by quantizing the weights of Qwen/Qwen2.5-VL-3B-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/Qwen2.5-VL-3B-Instruct-FP8-Dynamic",
    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 requests
import torch
from PIL import Image
from transformers import AutoProcessor
from llmcompressor.transformers import oneshot
from llmcompressor.transformers.tracing import (
    TraceableQwen2_5_VLForConditionalGeneration,
)
from llmcompressor.modifiers.quantization import QuantizationModifier

# Load model.
model_id = Qwen/Qwen2.5-VL-3B-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)

# Recipe
recipe = [
    QuantizationModifier(
        targets="Linear",
        scheme="FP8_DYNAMIC",
        sequential_targets=["MistralDecoderLayer"],
        ignore=["re:.*lm_head", "re:vision_tower.*", "re:multi_modal_projector.*"],
    ),
]

SAVE_DIR=f"{model_id.split('/')[1]}-FP8-Dynamic"

# Perform oneshot
oneshot(
    model=model,
    recipe=recipe,
    trust_remote_code_model=True,
    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-3B-Instruct	nm-testing/Qwen2.5-VL-3B-Instruct-FP8-Dynamic	Recovery (%)
Vision	MMMU (val, CoT) explicit_prompt_relaxed_correctness	44.56	45.78	102.74%
	VQAv2 (val) vqa_match	75.94	76.22	100.37%
	DocVQA (val) anls	92.53	92.40	99.86%
	ChartQA (test, CoT) anywhere_in_answer_relaxed_correctness	81.20	80.72	99.41%
	Mathvista (testmini, CoT) explicit_prompt_relaxed_correctness	54.15	53.25	98.34%
	Average Score	69.28	69.67	100.56%
Text	MGSM (CoT)	52.49	43.14	82.17%
Text	MMLU (5-shot)	65.32	65.03	99.56%

Inference Performance

This model achieves up to 1.10x speedup in single-stream deployment and up to 1.32x 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-3B-Instruct-FP8-Dynamic --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	Model	Average Cost Reduction	Latency (s)	Queries Per Dollar	Latency (s)		Queries Per Dollar	Latency (s)	Queries Per Dollar
A6000x1	Qwen/Qwen2.5-VL-3B-Instruct		3.1	1454	1.8	2546	1.7	2610
	neuralmagic/Qwen2.5-VL-3B-Instruct-quantized.w8a8	1.27	2.6	1708	1.3	3340	1.3	3459
	neuralmagic/Qwen2.5-VL-3B-Instruct-quantized.w4a16	1.57	2.4	1886	1.0	4409	1.0	4409
A100x1	Qwen/Qwen2.5-VL-3B-Instruct		2.2	920	1.3	1603	1.2	1636
	neuralmagic/Qwen2.5-VL-3B-Instruct-quantized.w8a8	1.09	2.1	975	1.2	1743	1.1	1814
	neuralmagic/Qwen2.5-VL-3B-Instruct-quantized.w4a16	1.20	2.0	1011	1.0	2015	1.0	2012
H100x1	Qwen/Qwen2.5-VL-3B-Instruct	1.5	740	0.9	1221	0.9	1276
	neuralmagic/Qwen2.5-VL-3B-Instruct-FP8-Dynamic	1.06	1.4	768	0.9	1276	0.8	1399
	neuralmagic/Qwen2.5-VL-3B-Instruct-quantized.w4a16	1.24	0.9	1219	0.9	1270	0.8	1304

**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 Dollarv	Maximum throughput (QPS)	Queries Per Dollar	Maximum throughput (QPS)	Queries Per Dollar
A6000x1	Qwen/Qwen2.5-VL-3B-Instruct		0.5	2405	2.6	11889	2.9	12909
	neuralmagic/Qwen2.5-VL-3B-Instruct-quantized.w8a8	1.26	0.6	2725	3.4	15162	3.9	17673
	neuralmagic/Qwen2.5-VL-3B-Instruct-quantized.w4a16	1.39	0.6	2548	3.9	17437	4.7	21223
A100x1	Qwen/Qwen2.5-VL-3B-Instruct		0.8	1663	3.9	7899	4.4	8924
	neuralmagic/Qwen2.5-VL-3B-Instruct-quantized.w8a8	1.06	0.9	1734	4.2	8488	4.7	9548
	neuralmagic/Qwen2.5-VL-3B-Instruct-quantized.w4a16	1.10	0.9	1775	4.2	8540	5.1	10318
H100x1	Qwen/Qwen2.5-VL-3B-Instruct		1.1	1188	4.3	4656	4.3	4676
	neuralmagic/Qwen2.5-VL-3B-Instruct-FP8-Dynamic	1.15	1.4	1570	4.3	4676	4.8	5220
	neuralmagic/Qwen2.5-VL-3B-Instruct-quantized.w4a16	1.96	4.2	4598	4.1	4505	4.4	4838