---

🧭 Navigation

• 📌 News & Highlights
• 🔍 Core Insights & Key Findings
• ⚙️ Method Overview & Sparsity Probing
• 📊 Multimodal Benchmark Results
• 🤗 MoE Adaptation Checkpoints
• 📦 Installation
• 🧩 Supported Architectures
• 🚀 Quickstart & Inference
• 🔬 Probing & Evaluation Workflows
• 📂 Code Structure
• 📑 Citation & Contact

---

📌 News & Highlights

• [2025.12] 📄 Paper released on arXiv: arXiv:2512.02351!
• [2025.12] 🌐 Interactive Project Page live at shwai-he.github.io/SparseUnifiedModel with an interactive multimodal sparsity visualizer!
• [2025.12] 🤗 Released MoE adaptation checkpoints for BAGEL (BAGEL-MoE-7B-GEN-16to8 & BAGEL-MoE-7B-GEN-32to16) on Hugging Face.
• [2025.12] 🚀 Full codebase released supporting training-free depth/width probing and sparse MoE conversion for BAGEL, Ming-Omni, and Qwen-Image.

---

🔍 Core Insights & Key Findings

Unified multimodal models integrate understanding (e.g., visual question answering, reasoning, cross-modal retrieval) and generation (e.g., text-to-image synthesis, visual editing) into a single, cohesive architecture. However, unifying these distinct modalities introduces critical inference bottlenecks, compute imbalances, and parameter redundancies.

We conduct a systematic, training-free probing investigation across depth and width dimensions:

┌──────────────────────────────────────────────────────────────────────────────────────────┐
│                                SYSTEMIC COMPRESSION ASYMMETRY                            │
├─────────────────────────────────────────────┬────────────────────────────────────────────┤
│         🧠 Understanding Components         │          🎨 Generation Components          │
├─────────────────────────────────────────────┼────────────────────────────────────────────┤
│ • High compressibility in generation tasks  │ • High sensitivity to compression          │
│ • 50%+ depth/width pruned with ~0% drop     │ • Severe quality collapse if pruned static │
│ • Serves as coarse high-level semantic prior│ • Requires high-precision continuous token │
│ • Tolerates aggressive layer dropping       │ • Best handled via dynamic sparse MoE      │
└─────────────────────────────────────────────┴────────────────────────────────────────────┘

1. Understanding components are heavily compressible during generation:
   • In generation tasks (e.g., text-to-image synthesis), understanding layers primarily extract coarse conditioning representations.
   • Up to 50% of understanding layers/neurons can be dropped with negligible loss in image alignment and visual quality (GenEval drops < 1.2%).
2. Generation components are exceptionally fragile:
   • In contrast, generation layers model complex, fine-grained pixel distributions. Moderate pruning leads to rapid artifacts, semantic distortion, and image quality collapse.
3. Dynamic Sparsity Motivates Sparse MoE Adaptation:
   • Inspection of neuron activation distributions reveals sample-dependent, input-specific activation patterns across generation layers.
   • Rather than static pruning, we convert dense MLP layers into Mixture-of-Experts (MoE) and sparsely activate them (e.g., top-8 out of 16 experts, or top-16 out of 32 experts).
   • Result: Halves active generation FLOPs while fully preserving dense generation quality!

Figure 1: Two-stage optimization pipeline: (1) Training-free component probing across depth and width, followed by (2) Sparse MoE adaptation to recover generation fidelity with 50% active parameters.

---

⚙️ Method Overview & Sparsity Probing

Our framework provides a unified pipeline for probing and exploiting multimodal sparsity:

  Unified Multimodal Model
            │
            ├───► 1. Depth Probing (Layer Dropping via Cosine Similarity / Output Impact)
            │         └── Identify redundant transformer layers across tasks
            │
            ├───► 2. Width Probing (Neuron Activation Calibration & Partitioning)
            │         └── Measure activation frequency across calibration datasets
            │
            └───► 3. MoE Adaptation (Dense MLP ──► Sparse Expert Routing)
                      └── Partition weight matrices into N experts and route Top-K dynamically

Probing Dimensions

---

📊 Multimodal Benchmark Results

1. Visual Generation Benchmarks (GenEval, DPG-Bench, ImageReward)

Evaluated on BAGEL-7B and its compressed / MoE variants:

Key Takeaway: Static generation pruning drops GenEval from 0.652 → 0.518, while BAGEL-MoE recovers it to 0.651 with only 54% active compute!

---

2. Visual Understanding Benchmarks (MME, MMBench, POPE, TextVQA)

---

🤗 MoE Adaptation Checkpoints

Pretrained sparse MoE checkpoints are hosted on Hugging Face:

---

📦 Installation

1. Create and Activate Environment

# Create conda environment
conda create -n sparse_um python=3.10 -y
conda activate sparse_um

# Install PyTorch (CUDA 12.1+ recommended)
pip install torch==2.1.2 torchvision==0.16.2 --index-url https://download.pytorch.org/whl/cu121

# Install requirements
pip install -r requirements.txt

# Install FlashAttention-2 (optional but recommended for speed)
pip install flash-attn --no-build-isolation

---

🧩 Supported Architectures

We provide unified compression hooks via modeling/compression_mixin.py across:

---

🚀 Quickstart & Inference

1. Interleaved Generation & Understanding with BAGEL

import torch
from inferencer import InterleaveInferencer
from modeling.bagel import Bagel, BagelConfig
from modeling.autoencoder import load_ae
from modeling.qwen2 import Qwen2Tokenizer

# 1. Load model and tokenizer
model_path = "LLM-Drop/BAGEL-MoE-7B-GEN-16to8"
tokenizer = Qwen2Tokenizer.from_pretrained(model_path)
model = Bagel.from_pretrained(model_path, torch_dtype=torch.bfloat16, device_map="auto")
vae_model = load_ae("vae_path")

inferencer = InterleaveInferencer(
    model=model,
    vae_model=vae_model,
    tokenizer=tokenizer,
    vae_transform=None,
    vit_transform=None,
    new_token_ids={}
)

# 2. Perform text-to-image synthesis
context = inferencer.init_gen_context()
inferencer.update_context_text("A cinematic shot of a red vintage sports car driving through a misty neon-lit forest at dusk.", context)
images = inferencer.generate_image(context, cfg_text_scale=4.0)
images[0].save("output.png")

---

🔬 Probing & Evaluation Workflows

1. Depth Pruning Probing (Layer Dropping)

Run Layer Dropping evaluation on GenEval for BAGEL, Ming-Omni, or Qwen:

# Generation task (GenEval benchmark)
bash scripts/eval/bagel/run_geneval_ld.sh
bash scripts/eval/ming/run_geneval_ld.sh
bash scripts/eval/qwen/run_geneval_ld.sh

# Understanding task (MME benchmark)
bash scripts/eval/bagel/run_vlm_ld.sh
bash scripts/eval/ming/run_vlm_ld.sh

2. Width Reduction Probing (Neuron Pruning)

Profile and prune intermediate neuron activations based on calibration samples:

# 1. Calculate neuron importance scores and partition masks
python scripts/neuron_partition.py   --model_path /path/to/bagel   --calibration_task understanding   --calibration_samples 64

# 2. Run width reduction evaluation
bash scripts/eval/bagel/run_geneval_wr.sh
bash scripts/eval/bagel/run_vlm_wr.sh

3. Converting Dense Models to Sparse MoE

Follow our interactive step-by-step Jupyter Notebook:

jupyter notebook notebooks/dense2sparse.ipynb

Or convert programmatically by specifying expert partitions:

# Partition dense intermediate MLP weights into 16 or 32 sparse expert slices
from utils.moe_utils import convert_dense_to_sparse_moe

sparse_model = convert_dense_to_sparse_moe(
    dense_model=model,
    num_experts=16,
    active_experts=8,
    target_modules=["mlp.down_proj", "mlp.gate_up_proj"]
)

---

📂 Code Structure

SparseUnifiedModel/
├── modeling/                   # Core multimodal model definitions
│   ├── bagel/                  # BAGEL unified multimodal LLM
│   ├── ming/                   # Ming-Omni MoE multimodal architecture
│   ├── qwen/                   # Qwen-Image text encoder & vision stack
│   ├── qwen2/                  # Base Qwen2 language model components
│   ├── siglip/                 # SigLIP vision transformer encoder
│   ├── diffusers/              # Diffusion pipelines and schedulers
│   └── compression_mixin.py    # Layer dropping and pruning instrumentation
│
├── eval/                       # Comprehensive evaluation suites
│   ├── gen/                    # Generation benchmarks (GenEval, DPG-Bench)
│   │   ├── gen_images.py       # Width reduction image generation runner
│   │   ├── gen_images_ld.py    # Layer drop image generation runner
│   │   ├── compress_utils.py   # Shared pruning utilities and hooks
│   │   └── geneval/            # GenEval scoring and prompt datasets
│   └── vlm/                    # Multimodal understanding benchmarks
│       └── eval/               # MME, MMBench, POPE, MMMU, MathVista, etc.
│
├── scripts/                    # Execution and evaluation scripts
│   ├── eval/                   # Benchmark runner bash scripts
│   │   ├── bagel/              # BAGEL evaluation scripts (LD, WR, Baseline)
│   │   ├── ming/               # Ming-Omni evaluation scripts
│   │   └── qwen/               # Qwen-Image evaluation scripts
│   ├── inferencer.py           # Interleaved inference engine
│   └── neuron_partition.py     # Neuron importance profiling and partitioner
│
├── tools/                      # Utilities and deployment integrations
│   ├── ming_sdk/               # Ming SDK client
│   ├── vllm/                   # vLLM integration patches
│   └── gradio_demo.py          # Interactive web UI demo
│
├── notebooks/                  # Interactive exploration and walkthroughs
│   ├── dense2sparse.ipynb      # Step-by-step Dense -> Sparse MoE conversion
│   ├── inference_bagel.ipynb   # BAGEL interactive inference demo
│   ├── inference_qwen.ipynb    # Qwen-Image interactive inference demo
│   └── inference_ming.ipynb    # Ming-Omni interactive inference demo
│
├── docs/                       # Project website & visual assets
│   ├── index.html              # Modern academic project webpage
│   └── static/images/          # Figures, architecture charts, and plots
├── requirements.txt            # Python dependencies
└── LICENSE                     # Apache-2.0 License

---

📑 Citation & Contact

If you find this work or codebase helpful in your research, please cite our paper:

@misc{he2025understandingharnessingsparsityunified,
  title={Understanding and Harnessing Sparsity in Unified Multimodal Models},
  author={Shwai He and Chaorui Deng and Ang Li and Shen Yan},
  year={2025},
  eprint={2512.02351},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2512.02351},
}

Contact & Inquiries

• Shwai He: shwai.he@bytedance.com
• Shen Yan: sheny@bytedance.com

---

Released under the Apache 2.0 License.