engram-peft 1.0.3

Engram-PEFT: Efficient Parameter-Efficient Fine-Tuning with Engram

Engram-PEFT

[English] | 中文

[!IMPORTANT] This is an unofficial implementation of the DeepSeek Engram paper (arXiv:2601.07372). DeepSeek-AI official demo is here.

Engram-PEFT is a high-performance, 100% paper-aligned implementation of the DeepSeek Engram architecture. It provides a Parameter-Efficient Fine-Tuning (PEFT) interface to inject conditional memory into any Transformer-based LLM.

Engram decouples static knowledge storage from dynamic reasoning using a sparse retrieval mechanism, allowing models to scale their factual memory without increasing inference FLOPs or interfering with core logic.

---

🚀 Quick Start

Installation

pip install engram-peft

To run examples or contribute to development, install the project with development dependencies:

# Using uv (recommended)
uv sync --all-groups

# Using pip
pip install -e ".[dev]"

5-Minute Example

from transformers import AutoModelForCausalLM, AutoTokenizer
from engram_peft import EngramConfig, get_engram_model

# 1. Load base model
base_model = AutoModelForCausalLM.from_pretrained("TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T")
tokenizer = AutoTokenizer.from_pretrained("TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T")

# 2. Inject Engram layers (aligned with arXiv:2601.07372)
config = EngramConfig(target_layers=[2, 11, 20])
model = get_engram_model(base_model, config, tokenizer)

# 3. Model is ready for training! 
# Only Engram parameters (approx 1% of total) are trainable.

---

📊 Performance Comparison

Method	Params Added	Grad. Update Size	VRAM (1.1B)
FFT (Full Fine-Tune)	0	1,100M	~24GB (est.)
LoRA (r=16)	1.8M	1.8M	~5.1GB
Engram-PEFT	11.2M	~1.2M*	~6.8GB

* Engram employs sparse lookup; only a tiny fraction of parameters (approx. 1%) are active and receive gradient updates per step. For a detailed breakdown of VRAM usage and scaling, see our Memory Analysis.

---

🛠 Features

• 100% Paper Alignment: Implements Appendix A Table 5 parameters and the official DeepSeek gating/hashing logic.
• CPU-Side Precomputation: EngramDataCollator precomputes multi-head hashes on CPU, ensuring 100% GPU utilization.
• Tokenizer Compression: Built-in NFKC and lowercase normalization to achieve 23% vocabulary reduction (consistent with paper).
• Zero-Invasive: Injects via forward hooks; no modification to your base model architecture required.
• Dynamic Switching: Load and swap "knowledge packs" at runtime without reloading the base model.

---

📖 Documentation

For full details, see our documentation:

• Tutorials: Quickstart and domain knowledge injection.
• API Reference: Detailed class and function documentation.
• Paper Alignment: How we match the DeepSeek research.

---

🎯 Citation

If you use this implementation in your research, please cite the original DeepSeek paper:

@article{deepseek2026engram,
  title={Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models},
  author={DeepSeek-AI},
  journal={arXiv preprint arXiv:2601.07372},
  year={2026}
}

---

License

Apache License 2.0. See LICENSE for details.