File size: 6,848 Bytes
2d0a056
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
---
title: CDD - Constrained Discrete Diffusion
emoji: 🧬
colorFrom: green
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
---

# Constrained Discrete Diffusion (CDD)

Implementation of **"Constrained Language Generation with Discrete Diffusion Models"** (Cardei et al., 2025 — [arXiv:2503.09790v3](https://arxiv.org/abs/2503.09790v3)).

CDD integrates discrete diffusion models with differentiable optimization to enforce constraints on text generation **without retraining**. It achieves **zero constraint violations** across toxicity mitigation, molecular generation, and instruction following.

## Key Idea

Discrete diffusion models (MDLM, UDLM) generate sequences by iteratively denoising from a noise distribution. At each reverse step, CDD inserts an **Augmented Lagrangian Method (ALM) projection** that modifies the denoiser's predicted token distributions to satisfy user-defined constraints:

```
z_T → x_θ(z_T) → [CDD Projection] → z_{T-1} → ... → z_0
```

The projection solves:
```
min_y  D_KL(x_θ || y)   s.t. g(argmax(y)) ∈ C
```

using Gumbel-Softmax relaxation for differentiability and ALM for constraint enforcement.

## Architecture

```
cdd/
├── samplers/
│   └── cdd_sampler.py      # Core ALM projection + sampling loop (§4)
├── constraints/
│   ├── toxicity.py          # Toxicity mitigation constraint (§5.1)
│   ├── molecular.py         # SA + Novelty constraints (§5.2)
│   └── instruction.py       # Counting + Lexical constraints (§5.3)
├── models/
│   └── load_models.py       # MDLM/UDLM model loading utilities
├── experiments/
│   ├── toxicity_mitigation.py     # Full toxicity experiment
│   ├── molecular_generation.py    # Full molecular experiment
│   ├── instruction_following.py   # Full instruction experiment
│   └── train_surrogates.py        # Surrogate model training
└── utils/
    ├── noise_schedule.py    # Diffusion math (schedules, posteriors)
    └── evaluation.py        # Metrics (PPL, entropy, violation rate)
```

## Installation

```bash
pip install -e .

# For GPU models (required for MDLM/UDLM inference):
pip install flash-attn --no-build-isolation

# For molecular evaluation:
pip install rdkit-pypi
```

## Quick Start

### 1. Unconstrained Generation with MDLM

```python
import torch
from cdd.models import load_mdlm
from cdd.samplers import CDDSampler, ALMConfig

# Load pretrained MDLM
model, tokenizer, info = load_mdlm(device="cuda")

# Create sampler (no constraints)
sampler = CDDSampler(
    model=model,
    tokenizer=tokenizer,
    constraint_fn=lambda x: torch.tensor(0.0),  # No constraint
    diffusion_type="mdlm",
    num_timesteps=1000,
    seq_length=128,
    device="cuda",
)

result = sampler.sample(batch_size=1)
print(result["text"][0])
```

### 2. Toxicity-Constrained Generation

```python
from cdd.samplers import CDDSampler, ALMConfig, TOXICITY_ALM_CONFIG

# Setup toxicity constraint (§5.1)
from cdd.constraints import create_toxicity_constraint
constraint = create_toxicity_constraint(threshold=0.5)

sampler = CDDSampler(
    model=model,
    tokenizer=tokenizer,
    constraint_fn=constraint,
    alm_config=TOXICITY_ALM_CONFIG,  # λ=0, μ=1, K=1000, M=10, η=0.2
    diffusion_type="mdlm",
    device="cuda",
)

# Conditional generation from toxic prompt
prefix = tokenizer.encode("The politician was accused of", return_tensors="pt").to("cuda")
result = sampler.sample(batch_size=1, prefix_ids=prefix)
print(result["text"][0])  # Non-toxic completion
```

### 3. Molecular Generation with SA Constraint

```python
from cdd.models import load_udlm
from cdd.samplers import CDDSampler, MOLECULAR_SA_ALM_CONFIG

model, tokenizer, info = load_udlm("kuleshov-group/udlm-qm9", device="cuda")

sampler = CDDSampler(
    model=model,
    tokenizer=tokenizer,
    constraint_fn=sa_constraint,  # SA ≤ 3.5
    alm_config=MOLECULAR_SA_ALM_CONFIG,  # λ=0, μ=1, μ_max=1000, K=1000, M=100, η=1.0
    diffusion_type="udlm",
    num_timesteps=1000,
    seq_length=32,
    device="cuda",
)

result = sampler.sample(batch_size=32)
```

## Reproducing Paper Experiments

### Toxicity Mitigation (§5.1, Table 3)

```bash
python -m cdd.experiments.toxicity_mitigation \
    --threshold 0.5 \
    --num_samples 1000 \
    --device cuda
```

Expected results (CDD τ=0.50): PPL ≈ 59.44, Violation = 0.0%

### Molecular Generation (§5.2, Table 4)

```bash
python -m cdd.experiments.molecular_generation \
    --sa_threshold 3.5 \
    --num_samples 1000 \
    --device cuda
```

### Instruction Following (§5.3, Table 5)

```bash
# Counting task
python -m cdd.experiments.instruction_following \
    --task counting --num_samples 100

# Lexical task
python -m cdd.experiments.instruction_following \
    --task lexical --num_samples 100
```

### Train Surrogate Models

```bash
# Toxicity surrogate (GPT-Neo 1.3B on Jigsaw)
python -m cdd.experiments.train_surrogates --task toxicity

# SA surrogate (GPT-2 on QM9)
python -m cdd.experiments.train_surrogates --task sa
```

## ALM Hyperparameters (Appendix B & C)

| Task | λ₀ | μ₀ | μ_max | K (outer) | M (inner) | η |
|------|-----|-----|-------|-----------|-----------|------|
| Toxicity (NLP) | 0.0 | 1.0 | — | 1000 | 10 | 0.20 |
| Molecular (QM9) | 0.0 | 1.0 | 1000 | 1000 | 100 | 1.0 |

## Base Models

| Model | Type | Params | Vocab | Length | Trained On |
|-------|------|--------|-------|--------|------------|
| [kuleshov-group/mdlm-owt](https://huggingface.co/kuleshov-group/mdlm-owt) | MDLM | 110M | 50,258 | 1024 | OpenWebText |
| [kuleshov-group/udlm-qm9](https://huggingface.co/kuleshov-group/udlm-qm9) | UDLM | 92M | 40 | 32 | QM9 |
| [kuleshov-group/udlm-lm1b](https://huggingface.co/kuleshov-group/udlm-lm1b) | UDLM | 139M | 30,522 | 128 | LM1B |

## Method Details

### Gumbel-Softmax Relaxation (Eq. 6)

Makes `argmax` differentiable:
```
φ̃(x)(v) = exp((log x(v) + ξ_v) / T) / Σ exp((log x(v') + ξ_v') / T)
```

### Augmented Lagrangian (Eq. 7-8)

Optimization at each diffusion step:
```
L(y, λ, μ) = D_KL(x_θ || y) + λ·g(ỹ) + (μ/2)·g(ỹ)²
```

Update rules:
```
λ ← λ + μ·g(ỹ)
μ ← min(2μ, μ_max)
```

### MDLM vs UDLM for CDD

| Property | MDLM | UDLM |
|----------|------|------|
| Noise type | Absorbing ([MASK]) | Uniform |
| Tokens re-sampled each step | Only masked | All |
| CDD paper usage | Text tasks | Molecular |
| Time conditioning | No | Yes |

## Citation

```bibtex
@article{cardei2025constrained,
  title={Constrained Language Generation with Discrete Diffusion Models},
  author={Cardei, Michael and Christopher, Jacob K and Hartvigsen, Thomas 
          and Bartoldson, Brian R. and Kailkhura, Bhavya and Fioretto, Ferdinando},
  journal={arXiv preprint arXiv:2503.09790},
  year={2025}
}
```