A minimalist, dependency-free implementation of the EGGROLL (Evolution Guided General Optimization via Low-rank Learning) algorithm family.

Mission: To get the most capable pairs of models / platforms cross-implementations of EGGROLL family of algorithms as possible and squeeze every possible bit of performance from the equipment, implementing everything required from the scratch in hardware-optimized fashion.

This project demonstrates integer-only training of a language model, completely bypassing the need for standard floating-point arithmetic or heavy ML frameworks like PyTorch or JAX.

• Pure C / Bare Metal: Old-school fashioned and goal-oriented. Zero external dependencies on the CPU side, keeping it close to the metal.
• Apple Silicon Optimized: Vectorized operations using ARM NEON intrinsics and parallelized via Grand Central Dispatch (GCD).
• NVIDIA CUDA Optimized: Custom GPU kernels utilizing Warp-level primitives, Shared Memory, and CUB/Thrust for maximum throughput.
• Integer Only: Operates primarily on int8 weights/activations with int32 (CPU) or int64 (GPU) accumulation—sticking to integer math as long as it yields the best performance for the hardware.
• Gradient Free: Uses Evolution Strategies (ES) with low-rank perturbations instead of backpropagation. It's both wisdom and freedom!

Ensure you have a text dataset named input.txt in the current directory.

clang -O3 full_trained_egg.c -o egg
./egg

nvcc -O3 full_cuda_train_egg.cu -o egg_cuda
./egg_cuda

Training Output

An int8 model.

• Native int8 Architecture: Operates on raw bytes with a compact N-layer, H-dim topology.
• Quantized Sigmoid Self-Attention: An int32/int64 accumulation scheme and quantized weighting.
• Auto-Norm & Entropy Monitoring: Adaptive normalization layers.
• EGG DEBUG: debug-printing "tool" to monitor entropy flow through the network and weights distribution and saturation.
• Information-Regulated Optimizer: A hybrid ES-AdamW approach where the optimizer (float32) regulates the amount of updates applied to the integer weights, ensuring stable learning.
• Performance: Achieves ~300k tokens/second with a population of 40,000+ (×ばつ5) on a single 4090 GPU setup, reaching loss rates (~1.45 bits/byte).

The system employs a Synchronous Replicated Model with Sharded Evaluation:

• Sharded Evaluation: The population is split across GPUs, with each evaluating a subset of perturbations in parallel.
• Implicit Synchronization: Instead of exchanging gradients (All-Reduce), GPUs receive only fitness scores. Since noise is deterministic, each GPU independently reconstructs the update, keeping replicas synchronized with negligible bandwidth.

image-20251206052311833

nvcc -O3 -arch=native full_cuda_train_transformer_adam_mgpu.cu -o egg_transformer_mgpu
./egg_transformer_mgpu

A lightweight header-only tool for monitoring integer model stability and detecting saturation or mode collapse.

• Metrics: Tracks Mean, StdDev, bit-level Entropy (0.00-8.00), and Saturation percentages per layer.
• Usage: Define EGG_DEBUG during compilation to enable ANSI-colored logs for activations and attention scores.

A robust, distributed training system designed to scale EGGROLL across multiple nodes and GPUs.

For full documentation, architecture details, and usage instructions, please see d-eggs/README.md.

• Architecture: Coordinator-Worker model with fault tolerance.
• Key Features: Custom Binary Protocol, CUDA Graphs, Ternary Packing, Muon Optimizer.

Configuration

We are friendly and welcome all sorts of contributions!

• Testers: Open issues with a description of your available compute, join existing issues if you can platforms described there.
• Moderators: To keep all this under control.
• Creatives: Even if you have nice creative IDEA on README design - you're welcome.

• Original JAX Implementation: ESHyperscale/nano-egg
• Original Paper & Project: EGGROLL Website