WaveletLM

WaveletLM is a fully causal, attention-free language model that mixes tokens through learned lifting wavelet decomposition, a Fast Walsh-Hadamard Transform, per-scale gated spectral mixing with SwiGLU activation, an inverse FWHT, and wavelet reconstruction. Combined with expanded MLPs and sparse product-key memory, this yields an architecture with no attention and O(n log n) scaling in sequence length.

Full code, training details, ablations, and documentation: github.com/ramongougis/WaveletLM

Results

Dataset	Params	Perplexity	BPB
WikiText-103	883M	23.8	1.0140
PG-19 (1 epoch)	808M	27.4	1.0853

How to Use

import torch
from huggingface_hub import hf_hub_download

# Download the checkpoint
ckpt_path = hf_hub_download(repo_id="ragou19/WaveletLM", filename="best_model.pt")

Then follow the instructions in the GitHub repo to load and run: https://github.com/ramongougis/WaveletLM

Architecture

Training

• Trained on a single RTX 5090 for 5 epochs
• WikiText-103: best PPL of 23.749 with mean PPL of 23.818 across 3 seeds.
• PG-19: PPL of 27.40 (single seed).
• VRAM required: 18.3 GB.
• Time to train: 16 hours 15 minutes.

Generation

• VRAM: 5.0 GB by default, 4.5 GB with --ptq8 enabled.
• Can set compile:false to save 0.5-1 GB, but it's slower.
• 28.8 tokens/s. on a 5090 by default.
• Future enhancements expected to increase speed by up to 120%.

Logs

See runs.md for the full training history.

License

Apache 2.0. See LICENSE.