spb/forge Public MIT
Forge — LLM training from scratch in pure C++20 + Metal on Apple Silicon.
C++ 61.2%
C 23%
Python 7.6%
TeX 7.2%
CMake 1.1%
1{2 "model": {3 "name": "gpt-10m",4 "n_layers": 6,5 "d_model": 384,6 "n_heads": 6,7 "n_kv_heads": 6,8 "d_ff": 1024,9 "vocab_size": 4096,10 "context_length": 512,11 "tied_embeddings": true,12 "use_rope": true,13 "rope_theta": 10000.0,14 "norm": "rmsnorm",15 "norm_eps": 1e-6,16 "activation": "swiglu",17 "dropout": 0.018 },19 "train": {20 "lr": 6e-4,21 "min_lr_ratio": 0.1,22 "warmup_steps": 1000,23 "max_steps": 20000,24 "beta1": 0.9,25 "beta2": 0.95,26 "eps": 1e-8,27 "weight_decay": 0.1,28 "grad_clip": 1.0,29 "batch_size": 64,30 "grad_accum_steps": 1,31 "precision": "f32",32 "checkpoint_every": 1000,33 "eval_every": 500,34 "eval_batches": 20,35 "seed": 133736 }37}38