pixtral-12b-FP8-Dynamic

Model Overview

• Model Architecture: mgoin/pixtral-12b
  • 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 mgoin/pixtral-12b.

Model Optimizations

This model was obtained by quantizing the weights of mgoin/pixtral-12b 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/pixtral-12b-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 TraceableLlavaForConditionalGeneration
from llmcompressor.modifiers.quantization import QuantizationModifier
import os

# Load model.
model_id = mgoin/pixtral-12b
model = TraceableLlavaForConditionalGeneration.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.w4a16 \
        --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="neuralmagic/pixtral-12b-quantized.w4a16 ",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 \

HumanEval

Generation

python3 codegen/generate.py \
  --model neuralmagic/pixtral-12b-quantized.w4a16 \
  --bs 16 \
  --temperature 0.2 \
  --n_samples 50 \
  --root "." \
  --dataset humaneval

Sanitization

python3 evalplus/sanitize.py \
  humaneval/neuralmagic/pixtral-12b-quantized.w4a16_vllm_temp_0.2

Evaluation

evalplus.evaluate \
  --dataset humaneval \
  --samples humaneval/neuralmagic/pixtral-12b-quantized.w4a16_vllm_temp_0.2-sanitized

Accuracy

Category	Metric	mgoin/pixtral-12b	neuralmagic/pixtral-12b-FP8-Dynamic	Recovery (%)
Vision	MMMU (val, CoT) explicit_prompt_relaxed_correctness	48.00	50.11	104.40%
	VQAv2 (val) vqa_match	78.71	78.44	99.66%
	DocVQA (val) anls	89.47	89.20	99.70%
	ChartQA (test, CoT) anywhere_in_answer_relaxed_correctness	81.68	81.76	100.10%
	Mathvista (testmini, CoT) explicit_prompt_relaxed_correctness	56.50	58.70	103.89%
	Average Score	70.07	71.24	101.67%
Text	HumanEval pass@1	68.40	69.50	101.61%
	MMLU (5-shot)	71.40	69.50	97.34%

Inference Performance

This model achieves up to 1.80x speedup in single-stream deployment and up to 1.36x 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-12b-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	mgoin/pixtral-12b		5.7	796	4.8	929	4.7	964
	neuralmagic/pixtral-12b-quantized.w8a8	1.55	3.7	1220	3.1	1437	3.0	1511
	neuralmagic/pixtral-12b-quantized.w4a16	2.16	3.2	1417	2.1	2093	1.9	2371
A100x1	mgoin/pixtral-12b		3.0	676	2.4	825	2.3	859
	neuralmagic/pixtral-12b-quantized.w8a8	1.38	2.2	904	1.7	1159	1.7	1201
	neuralmagic/pixtral-12b-quantized.w4a16	1.83	1.8	1096	1.3	1557	1.2	1702
H100x1	mgoin/pixtral-12b		1.8	595	1.5	732	1.4	764
	neuralmagic/pixtral-12b-FP8-Dynamic	1.35	1.4	767	1.1	1008	1.0	1056
	neuralmagic/pixtral-12b-quantized.w4a16	1.37	1.4	787	1.1	1018	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
A6000x1	mgoin/pixtral-12b		0.6	2632	0.9	4108	1.1	4774
	neuralmagic/pixtral-12b-quantized.w8a8	1.50	0.9	3901	1.4	6160	1.6	7292
	neuralmagic/pixtral-12b-quantized.w4a16	1.41	0.6	2890	1.3	5758	1.8	8312
A100x1	mgoin/pixtral-12b		1.1	2291	1.8	3670	2.1	4284
	neuralmagic/pixtral-12b-quantized.w8a8	1.38	1.5	3096	2.5	5076	3.0	5965
	neuralmagic/pixtral-12b-quantized.w4a16	1.40	1.4	2728	2.6	5133	3.5	6943
H100x1	BF16		2.6	2877	4.0	4372	4.7	5095
	neuralmagic/pixtral-12b-FP8-Dynamic	1.33	3.4	3753	5.4	5862	6.3	6917
	neuralmagic/pixtral-12b-quantized.w4a16	1.22	2.8	3115	5.0	5511	6.2	6777

**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).