NeuroFlow C++ LLM

NeuroFlow is a memory-augmented causal language model implemented entirely in C++17, featuring a three-brain architecture (SN/ECN/DMN) inspired by cognitive neuroscience. Designed for efficient training and inference on consumer GPUs.

Model Architecture

Parameter	Value	Description
`d_model`	512	Model dimension
`hidden_dim`	2048	FFN hidden dimension
`memory_dim`	512	Memory dimension
`num_layers`	12	ECN layers
`memory_slots`	64	Memory slots
`num_associations`	8	DMN association heads
`vocab_size`	128,000	128K multilingual BPE tokenizer
`max_seq_len`	512	Maximum sequence length
`causal_window_size`	64	Causal attention window
`lm_num_attn_layers`	2	Causal LM attention layers
`params`	~120M	Total parameters

Three-Brain Architecture

SN (Sensory Network): Input encoding and feature extraction
ECN (Executive Control Network): Core reasoning and processing layers
DMN (Default Mode Network): Memory-augmented association and retrieval

Files

File	Size	Description
`output/checkpoint_step1000/model.nfv1`	431 MB	Checkpoint at 1000 steps
`output/checkpoint_step2000/model.nfv1`	431 MB	Checkpoint at 2000 steps
`output/lm_head_lmh1.nfv1`	255 MB	LM head weights (native format v1)
`configs/config.json`	—	Model architecture configuration
`configs/tokenizer_128k.json`	—	128K BPE tokenizer
`configs/huggingface/`	—	HuggingFace-compatible tokenizer files (vocab.json, merges.txt)

Source Code

The full C++ source is included under src/ and include/ directories:

Core: tensor.hpp/cpp, model.hpp/cpp, tokenizer.hpp/cpp
Architecture: causal_lm.hpp/cpp, generative_model.hpp/cpp, networks.hpp
Training: train_lm.hpp/cpp, train_v2.cpp, sft_train.cpp, dpo_train.cpp
CUDA: cuda_context.hpp/cpp, cuda_kernels.hpp, tensor_ops.cpp
Optimizers: adamw.hpp/cpp, scheduler.hpp/cpp, grad_scaler.hpp/cpp

Build & Train

Prerequisites

CMake ≥ 3.15
C++17 compiler (GCC ≥ 9, MSVC 2019+)
CUDA Toolkit ≥ 11.4 (optional, for GPU training)
BLAS (OpenBLAS recommended)

Build

# CPU only
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)

# With CUDA
mkdir build_cuda && cd build_cuda
cmake .. -DNEUROFLOW_USE_CUDA=ON -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)

Train

./build_cuda/neuroflow_train_v2 \
  --config configs/config_distill.json \
  --data data/distill_train.txt \
  --output output \
  --epochs 20 \
  --batch-size 64 \
  --lr 0.0003 \
  --use-cuda --adam

Training Scripts

Key Python scripts in scripts/:

train_distill.py — Knowledge distillation training pipeline
preprocess_distill.py — Data preprocessing for distillation
deploy_dsw.sh — One-click deployment for Alibaba Cloud DSW (A10 GPU)
train_optimized.sh — Optimized multi-stage training

License

Apache 2.0

cwenzi
/

neuroflow-cpp