engram-peft 1.0.4

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. Quick check on trainable parameters
model.print_trainable_parameters()
# trainable params: 86,938,368 || all params: 1,186,986,752 || trainable%: 7.3243

---

📊 Performance Comparison

Method	Params Added	Speed (s/step)	Training Loss	Eval Loss	VRAM (nvtop)
LoRA (r=16)	~2.25 M	0.2738 s	1.231	0.989	9.35 GiB
Engram-PEFT	545.4 M	0.2961 s	1.263	1.017	10.82 GiB

[!TIP] Performance Insight: In our latest 3000-step benchmark on RTX 4090D, LoRA achieved slightly better loss and speed. However, Engram-PEFT provides 240x more parameter capacity (545M) for knowledge storage with only a ~8% latency penalty, making it ideal for tasks requiring massive factual recall.

Loss Curve Comparison

* 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 performance, computation, and memory, see our Performance Analysis.

---

🛠 Features

• 100% Paper Alignment: Implements Appendix A Table 5 parameters and the official DeepSeek gating/hashing logic.
• CPU Prefetching & Precomputation: EngramDataCollator pre-calculates multi-head hash indices on the CPU. By using num_workers > 0, these indices are prefetched in parallel with training, ensuring zero hashing overhead on the GPU.
• Tokenizer Compression: Built-in NFKC and lowercase normalization for 23% vocabulary reduction.
• Cross-Model Weight Migration: A unique feature (see weight_transfer.py) that allows migrating Engram weights between different models (e.g., Llama to Qwen) using character-level alignment on a corpus—effectively "recycling" learned knowledge.
• Zero-Invasive: Injects via forward hooks; no modification to your base model architecture required.
• Peft-like API: Familiar methods like print_trainable_parameters() and save_pretrained().
• Named Adapters: Industry-standard named adapter management (add/set/unload) for modular knowledge packs.
• Automated Training: Native EngramTrainer with built-in sparse Adam support and automatic sync of optimizer hyperparameters.

---

📖 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.