SPB Git
spboucher.ai

Repository Public MIT

spbforge

History

Latest commit Add wave 3: DeepSeek-V3-style MoE routing, all config-selected 2e570b0

NameLast commitUpdated
..
adamw.metal Forge: LLM training from scratch in C++20 + Metal on Apple Silicon
attention.metal Add architecture-variant waves 1+2: train Mistral/Qwen/Gemma/OLMo-class models by config
cross_entropy.metal Forge: LLM training from scratch in C++20 + Metal on Apple Silicon
elementwise.metal Add wave 3: DeepSeek-V3-style MoE routing, all config-selected
embedding.metal Forge: LLM training from scratch in C++20 + Metal on Apple Silicon
flash_attention_mma.metal Add architecture-variant waves 1+2: train Mistral/Qwen/Gemma/OLMo-class models by config
flash_attention.metal Add architecture-variant waves 1+2: train Mistral/Qwen/Gemma/OLMo-class models by config
layernorm.metal Forge: LLM training from scratch in C++20 + Metal on Apple Silicon
matmul_mpp.metal Forge: LLM training from scratch in C++20 + Metal on Apple Silicon
matmul_simd.metal Forge: LLM training from scratch in C++20 + Metal on Apple Silicon
matmul.metal Forge: LLM training from scratch in C++20 + Metal on Apple Silicon
moe.metal Add wave 3: DeepSeek-V3-style MoE routing, all config-selected
precision_probe.metal Forge: LLM training from scratch in C++20 + Metal on Apple Silicon
quant.metal Add QAT, MoE, and architecture-variant knobs — all config-selected
softmax.metal Forge: LLM training from scratch in C++20 + Metal on Apple Silicon

Clone forge

Anonymous clone and fetch. Pushing requires the owner's personal access token.

Snapshot of 147560ec423d00ad5f58daaf762e57546c819b00

ZIP tar.gz
Files 147560ec423d00ad5f58daaf762e57546c819b00
Menu

Theme
↑↓navigate ↵open escclose /⌘Kanywhere