hdlm-group/hdlm-base-gamma-0.01

HDLM-Gamma: Hybrid Diffusion Language Model

This is the model card for dlm-group/hdlm-base-gamma-0.01.

Model Description

HDLM-Gamma is a hybrid diffusion language model that unifies autoregressive and diffusion-based sequence generation through gamma-hybrid noising. This model interpolates transition operators between absorbing and uniform processes, making it conceptually closer to SEDD (Lou et al. 2024) while maintaining the benefits of both paradigms.

The gamma parameter (γ) controls the blend between absorbing and uniform transition matrices: Q_gamma = (1-γ) * Q_absorb + γ * Q_uniform, where smaller values emphasize the absorbing process and larger values incorporate more uniform transitions.

Model Architecture

• Base Model: Transformer architecture with staggered score conditioning
• Vocabulary Size: 50,258 tokens (GPT-2 vocabulary + absorbing token)
• Context Length: Variable (supports up to 2048 tokens)
• Training: Continuous-time diffusion with gamma-hybrid graph structure
• Inference: Analytic predictor with staggered score computation

Usage

Quick Start

from hdlm.hf_utils import smart_model_loader
from hdlm.gamma_hybrid.sampling import get_sa_sampling_fn
from transformers import GPT2TokenizerFast
import torch

# Load model using smart loader (automatically detects model type)
model, cfg, device, accelerator, metaschedule = smart_model_loader(
    model_path="hdlm-group/hdlm-base-gamma-0.01",
    model_type="auto",  # automatically detects gamma_hybrid
    device="cuda"
)

# Load tokenizer
tokenizer = GPT2TokenizerFast.from_pretrained('gpt2')

# Generate text
prompt = "The future of artificial intelligence"
prompt_ids = tokenizer.encode(prompt, return_tensors='pt').to(device)

# Configure sampling function (automatically set up from config)
sampling_fn = get_sa_sampling_fn(
    config=cfg,
    graph=None,  # Will be created from config
    noise=None,  # Will be created from config
    meta_schedule=metaschedule,
    batch_dims=(1,),
    eps=1e-4,
    device=device
)

# Generate samples
generated = sampling_fn(
    model=model,
    prompt=prompt_ids,
    context_length=1024
)

# Decode generated text  
generated_text = tokenizer.decode(generated[0], skip_special_tokens=True)
print(generated_text)

Evaluation

# Text generation evaluation
python hdlm/eval_generation.py \
    --checkpoint_path hdlm-group/hdlm-base-gamma-0.01 \
    --sampling_method SAR \
    --save_samples

# Perplexity evaluation
python hdlm/eval_modeling.py \
    --checkpoint_path hdlm-group/hdlm-base-gamma-0.01 \
    --work_dir "./logs/eval_modeling_gamma" \
    --dataset ptb

Training Details

• Dataset: OpenWebText
• Batch Size: 256
• Learning Rate: 3e-4 with lambda scheduling
• Gamma (γ): 0.01 (controls hybrid transition blend)
• Graph Type: QGamma with expanded sigma conditioning
• Noise Schedule: Log-linear (σ_min=1e-4, σ_max=10.0)
• Training Steps: 1M iterations
• Warmup: 50K steps

Key Components

Graph Structure

The QGamma graph combines absorbing and uniform transition matrices:

• Absorbing component: Transitions to absorbing state (mask token)
• Uniform component: Uniform transitions between all tokens
• Hybrid blend: Controlled by gamma parameter

Staggered Score

The model uses staggered score computation that applies different transformations to absorbing and uniform branches before combining them, enabling more flexible generation patterns.

Sampling Strategy

• Predictor: Analytic predictor with exact transition computation
• Strategy: Direct sampling with configurable strategy parameter
• Noise Removal: Optional final denoising step

Model Variants

Available gamma values and their characteristics:

• γ = 0.01: Minimal uniform transitions, closest to pure absorbing process
• γ = 0.1: Moderate hybrid behavior with increased uniform mixing
• γ = 0.5: Balanced absorbing-uniform transition blend

Citation

@article{fathi2025unifying,
  title={Unifying autoregressive and diffusion-based sequence generation},
  author={Fathi, Nima and Scholak, Torsten and No{\"e}l, Pierre-Andr{\'e}},
  journal={arXiv preprint arXiv:2504.06416},
  year={2025}
}

License

This model is released under the same license as the original HDLM codebase. Please refer to the GitHub repository for license details.