FlashMHF: Flash Multi-Head Feed-Forward Network

Official open-source implementation for the paper: Flash Multi-Head Feed-Forward Network (arXiv:2512.06989)

Overview

FlashMHF introduces a novel Flash Multi-Head Feed-Forward Network architecture that addresses the fundamental limitations of naive Multi-Head FFNs through two key innovations:

1. Scale-Balanced Parallel Sub-Networks: Maintains optimal dimension ratios across model scales through dynamically weighted parallel sub-networks
2. Memory-Efficient Flash Kernel: I/O-aware fused kernel that computes outputs online in SRAM, eliminating the memory explosion problem

Key Features

• 🚀 Superior Performance: Consistently outperforms SwiGLU FFN baselines on perplexity and downstream tasks
• ⚡ Memory Efficient: 3-5x reduction in peak GPU memory usage compared to SwiGLU FFN
• 🔥 Fast Inference: Up to 1.08x speedup on Hopper architecture
• 📈 Scalable: Validated on models from 128M to 1.3B parameters
• 🛠️ Production Ready: Optimized kernels for both Triton and ThunderKittens (CUDA)

Architecture

FlashMHF draws inspiration from the structural symmetry between self-attention and FFNs:

Attention: softmax(QK^T/√d_k)V
FFN:       σ(XW₁)W₂

Our multi-head design partitions computation across parallel sub-networks with learned gating:

O = Σₑ Rₑ * ((SiLU(Q @ Kₑ^T) ⊙ (Q @ Uₑ^T)) @ Vₑ)

Key Innovations

1. Parallel FFN Sub-Networks: Each head uses E sub-networks with balanced dimension dₑ ≈ (8/3)dₕ
2. Flash Algorithm: Blockwise computation avoids materializing large intermediate tensors
3. Gating Mechanism: Sigmoid-normalized routing weights for dynamic expert selection

Model Configurations

We provide pre-configured models at different scales:

Model Size	Hidden Size	Layers	FFN Heads	Experts/Head	Config File
128M	768	10	8	8	configure-1-tinymhffnmoe.json
370M	1024	20	8	11	configure-2-mhffnmoe-TriDao-350M.json
1.3B	2048	20	16	22	configure-3-mhffnmoe-TriDao-1.3B.json

Performance Results

Language Modeling (370M Scale)

Variant	Eval Loss ↓	Eval PPL ↓
Baseline (SwiGLU)	3.030	20.69
PKV (d_h=128)	3.334	28.07
MH-FFN (d_h=128)	3.031	20.71
FlashMHF (d_h=128)	3.014	20.36

Memory & Speed Improvements

• Memory: 3-5x reduction in peak GPU memory vs SwiGLU FFN
• Speed: Up to 1.08x inference speedup on H100 GPUs
• Scaling: Consistent improvements from 128M to 1.3B parameters

Kernels

FlashMHF includes optimized kernel implementations:

Triton Kernels

• Location: kernels/triton/
• Features: Auto-tuned block sizes, memory-coalesced access patterns
• Usage: Default backend, works on most CUDA GPUs

ThunderKittens CUDA Kernels

• Location: kernels/thunderkittens/
• Features: Highly optimized for H100, warp-level programming
• Requirements: CUDA 12.0+, H100 GPU recommended

Repository Structure

FlashMHF/
├── models/flashmhf/              # Model implementations
│   ├── modeling_mhffnmoe.py      # Main model class
│   ├── configuration_mhffnmoe.py # Model configuration
│   └── convert_mhffnmoe_weights_to_hf.py
├── kernels/                      # Optimized kernel implementations  
│   ├── triton/                   # Triton-based kernels
│   │   ├── flash_mlp.py          # Single-head flash MLP
│   │   └── flash_mlp_moe.py      # Multi-expert flash MLP
│   └── thunderkittens/           # CUDA-based kernels
│       ├── flash_ffn_moe.cu      # CUDA kernel implementation
│       ├── flash_ffn_moe_torch.py # PyTorch wrapper
│       └── run_from_python.py    # Test script
├── configs/                      # Model configurations
│   ├── baseline/                 # SwiGLU baseline configs
│   ├── mhffn/                    # Naive multi-head FFN configs  
└── └── mhffnmoe/                 # FlashMHF configs

Paper

• Title: Flash Multi-Head Feed-Forward Network
• Authors: Minshen Zhang*, Xiang Hu*, Jianguo Li, Wei Wu, Kewei Tu
• arXiv: 2512.06989

If you find this repository useful, please cite:

@article{zhang2025flashmhf,
  title        = {Flash Multi-Head Feed-Forward Network},
  author       = {Zhang, Minshen and Hu, Xiang and Li, Jianguo and Wu, Wei and Tu, Kewei},
  journal      = {arXiv preprint arXiv:2512.06989},
  year         = {2025},
  url          = {https://www.arxiv.org/abs/2512.06989}
}