omgformer 2.0.5

Parallel Diffusion Language Model with 60 features — MoE, LoRA, attention variants, and more

OMGFormer

Parallel Diffusion Language Model — 60 features, production-ready

OMGFormer is a modular PyTorch library for building and training parallel masked diffusion language models. It ships with a comprehensive set of attention variants, MoE routing strategies, LoRA fine-tuning methods, and training utilities.

---

Installation

pip install omgformer
# with safetensors support (recommended for LoRA saving)
pip install omgformer[safetensors]

---

Quick Start

from omgformer import OMGConfig, OMGModel

config = OMGConfig(
    vocab_size=32000,
    hidden_size=768,
    num_layers=12,
    num_heads=12,
)
model = OMGModel(config)

---

Feature Highlights

Attention (Features #1–#16)

• Grouped Query Attention (GQA) — reduced KV heads for efficient inference
• Multi-Head Latent Attention (MLA) — DeepSeek-style latent compression
• Sliding Window Attention — Mistral-style local context windows
• Linear Attention — O(n) complexity via kernel feature maps
• Block Sparse Attention — memory-efficient sparse patterns
• RoPE variants: standard, YaRN, NTK-aware, LongRoPE
• ALiBi and T5 relative position biases

Layers (Features #17–#28)

• SwiGLU, GeGLU, ReGLU FFN variants
• RMSNorm and ScaleNorm
• AdaLN modulation for diffusion timestep conditioning
• Token Merging — dynamic sequence length reduction
• Stochastic Depth — drop-path regularization

Mixture of Experts (Features #37–#40)

• Standard top-K token-choice routing
• Expert Choice routing (Google Switch-style) — perfect load balance
• Soft MoE (Google Brain 2023) — fully differentiable routing
• Shared expert (DeepSeek MoE) — always-on dense + sparse experts

LoRA / PEFT (Features #41–#44)

• Standard LoRA with configurable rank and alpha
• DoRA — weight-decomposed LoRA (Liu et al., 2024)
• rsLoRA — rank-stabilized scaling for high-rank training
• LoRA+ — different learning rates for A and B matrices
• Save/load only adapter weights (~MB, not GB)

Training (Features #45–#52)

• Lion optimizer (Chen et al., 2023)
• EMA with warmup annealing
• Warmup + cosine LR schedule with optional restarts
• Gradient checkpointing helper
• FSDP wrapping helper

Advanced (Features #53–#60)

• KV Cache for autoregressive decoding
• Multi-Token Prediction Head (MTP)
• Model merging: SLERP, DARE, TIES
• Reward model head + PPO step
• INT8 / INT4 quantization stubs
• GGUF export stub
• RAG context injector
• Dynamic batching engine
• Chunked long-document attention

---

LoRA Fine-tuning Example

from omgformer import OMGConfig, OMGModel, add_lora, merge_lora, save_lora

model = OMGModel(OMGConfig())

# Add LoRA adapters (freezes base weights automatically)
model = add_lora(model, rank=16, alpha=32)

# ... fine-tune ...

# Merge and save
model = merge_lora(model)
save_lora(model, "./my_adapter")

---

MoE Example

from omgformer import OMGConfig, OMGModel

config = OMGConfig(
    use_moe=True,
    num_experts=8,
    num_experts_per_token=2,
    moe_expert_choice=False,   # or True for Expert Choice routing
)
model = OMGModel(config)
output, aux_loss = model(input_ids)
loss = ce_loss + aux_loss  # add load-balancing loss during training

---

License

Apache 2.0 — see LICENSE.

Links

• GitHub: https://github.com/fastloraoffical/OMGformers
• PyPI: https://pypi.org/project/omgformer/