Qwen2.5-VL-3B-Instruct-MXFP4-W4A4-LM-Only

This is an MXFP4 W4A4 quantized version of Qwen/Qwen2.5-VL-3B-Instruct, created using llm-compressor.

Only the LLM decoder is quantized. The Vision Transformer (ViT) encoder remains in BF16 precision.

Model Summary

Property	Value
Base Model	Qwen/Qwen2.5-VL-3B-Instruct
Quantization	MXFP4 W4A4 (4-bit float weights, 4-bit float activations)
Quantization Scope	LLM decoder only (ViT encoder in BF16)
Strategy	Per-tensor-group (group_size=32), symmetric (minmax observer)
Format	compressed-tensors (mxfp4-pack-quantized)
Model Size	~3.3 GB (1 shard)
Ignored Layers	lm_head, all model.visual.* layers
Tool	llm-compressor
Supported Runtime	vLLM (with compressed-tensors)

Quantization Details

• Weights: FP4 E2M1 (4-bit float), per-tensor-group (group_size=32) with E8M0 (uint8 exponent) block scales, static quantization
• Activations: FP4 E2M1 (4-bit float), per-tensor-group (group_size=32), dynamic quantization
• Ignored: lm_head (kept in BF16) and all ViT encoder layers (model.visual.*)
• Calibration: 512 samples from CNN/DailyMail, max_seq_length=2048
• Standard: OCP MX Specification

Quantization Recipe

quant_stage:
  quant_modifiers:
    QuantizationModifier:
      ignore: ["lm_head", "re:model.visual.*"]
      scheme: "MXFP4"
      targets: ["Linear"]

Hardware Requirements

MXFP4 inference requires NVIDIA Blackwell (SM120+) GPUs with CUDA 12.8+ for native CUTLASS MXFP4 GEMM support.

Usage

With vLLM

export VLLM_ATTENTION_BACKEND=TORCH_SDPA

vllm serve JongYeop/Qwen2.5-VL-3B-Instruct-MXFP4-W4A4-LM-Only \
    --trust-remote-code \
    --max-model-len 4096 \
    --enforce-eager

With Transformers

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "JongYeop/Qwen2.5-VL-3B-Instruct-MXFP4-W4A4-LM-Only",
    torch_dtype="auto",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-3B-Instruct")

messages = [{"role": "user", "content": [
    {"type": "image", "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"},
    {"type": "text", "text": "Describe this image in detail."},
]}]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(text=[text], images=image_inputs, videos=video_inputs, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=256)
result = processor.batch_decode(output[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(result[0])

Model Architecture

• Architecture: Qwen2_5_VLForConditionalGeneration
• LLM Hidden Size: 2048
• LLM Intermediate Size: 11008
• LLM Layers: 36
• LLM Attention Heads: 16
• LLM KV Heads: 2 (GQA)
• Vision Encoder Depth: 32
• Vision Hidden Size: 1280
• Vocabulary Size: 151936

Differences from NVFP4

Feature	MXFP4	NVFP4
Scale Format	E8M0 (uint8 exponent)	E4M3 + FP32 global scale
Group Size	32	16
Standard	OCP MX Specification	NVIDIA proprietary
Hardware	SM120+ (Blackwell)	SM89+ (Ada/Hopper/Blackwell)

Acknowledgments

• Base model by Qwen Team
• Quantization powered by llm-compressor and compressed-tensors