pyramis-l3d

Pyramis-L3D

中文文档 · English

Pyramis-L3D is a hierarchical / sparse attention research prototype. Inspired by CPU multi-level caches (L1/L2/L3) and the TLB (translation lookaside buffer), it models attention as a three-tier pyramidal storage system — targeting sub-linear KV-cache storage and O(k) query-time sparse routing.

⚠️ Disclaimer: This is a research prototype, not production-ready. It explores hierarchical / sparse attention mechanisms and is intended for further development by researchers.

Highlights

Introduction

• 🔴 Pain point: Conventional Transformers grow the KV cache linearly with sequence length, making long contexts and high-dimensional data (e.g. 3D point clouds) increasingly expensive.
• 🔵 Solution: A hierarchical pyramid cache (L1/L2/L3) plus Latent-TLB sparse routing organizes KV into three tiers, attending only to the top_k hit rows per query for O(k) complexity.

Core Mechanism

Tiny Dev Config

d_model    = 256            n_layers   = 4
n_heads    = 8              head_dim   = 32
vocab_size = 121            intermediate_size = 1024
L1_window  = 128            L2_window  = 512
L2_stride  = 2              L2_conv_kernel    = 5
L3_codebook_size (M) = 1024        L3_latent_dim = 64
W_distinct = 64             top_k      = 16
max_position_embeddings = 1024
Parameters ≈ 4.47M

Quick Start

Model weights are hosted on Hugging Face; this repository contains only the code.

pip install transformers torch safetensors

from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "FemtoRhythm/pyramis-l3d", trust_remote_code=True
)

Local development from source:

from transformers import AutoConfig, AutoModelForCausalLM

config = AutoConfig.from_pretrained(".")          # reads config.json via auto_map
model = AutoModelForCausalLM.from_config(config)

Training & Evaluation

# Train (character-level tokenizer, reads cached IMDB parquet)
python train.py --nsamples 2000 --steps 200 --ctx 256 --batch 16 --lr 1e-3 --out ./checkpoint

# Evaluate (PPL / sub-linear row-pool scaling / routing health)
python eval.py --checkpoint ./checkpoint --ctx 256

Briefly trained on character-level IMDB, held-out PPL = 1.027. The row pool stays constant across context length (sub-linear storage):

Benchmark (vs MLA)

Standalone uv environment comparing Pyramis-L3D against DeepSeek-V2-style MLA (only the attention module differs):

cd benchmark
uv run python bench_compare.py --nsamples 500 --steps 300 --ctx 256 --batch 16

Project Status

• L1 dense GQA
• L2 CNN-pooled KV
• L3 addressable dictionary (codebook)
• Latent-TLB sparse routing
• distinctness protection
• CNN local gating
• Training script + safetensors checkpoint
• Benchmark evaluation (vs MLA)
• Paging mechanism
• Daemon thread
• PD dual-path separation

Limitations

• Research prototype: no long-sequence / production validation; system-level claims (paging, daemon thread, PD dual-path) are not implemented.
• Low codebook utilization: the VQ codebook collapses under short training (~0.2% utilization), a known phenomenon mitigated by longer training or better initialization.
• Character-level tokenizer: tiny vocabulary, unsuitable for general text modeling.
• Weights are briefly trained on 2000 samples, with no generalization guarantee.

Citation

@misc{pyramis-l3d,
  author       = {FemtoRhythm},
  title        = {Pyramis-L3D: A Hierarchical Sparse Attention Research Prototype},
  year         = {2026},
  howpublished = {\url{https://github.com/FemtoRhythm/pyramis-l3d}},
  note         = {Research prototype, not formally published}
}

License

Apache 2.0