Skip to content

Latest commit

 

History

History
32 lines (26 loc) · 2.07 KB

File metadata and controls

32 lines (26 loc) · 2.07 KB

FusionLLM-v1 Documentation

Conceptual reference for the architecture, training stack, and data pipeline of FusionLLM-v1. These notes capture the why and how behind the code; the code itself is kept clean and free of explanatory comments.

For the authoritative project overview, see the top-level README.md and AGENTS.md. This documentation/ folder supplements those with the detailed rationale that previously lived inline in the source files.

Index

Document Component Source file(s)
architecture.md 24-layer hybrid topology, μP init, logit softcap, tied embeddings models/fusionllm.py
mla.md Multi-Head Latent Attention (Q LoRA 192, KV rank 96, decoupled RoPE, absorption trick, FA2) models/mla.py
gdn.md Gated Delta Net linear attention (32 heads, chunk 64, FP32 recurrent state, snake/sigmoid gating) models/gdn.py
moe.md DeepSeekMoE (8 routed top-2 + 1 shared, aux-loss-free biased-sigmoid routing) models/moe.py
mtp.md Multi-Token Prediction heads (depth 2 λ=0.10, depth 3 λ=0.05, shared output head) models/mtp.py
fusionllm.md Top-level FusionLLMBlock wiring, forward / forward_with_hidden, generate models/fusionllm.py
training.md Dual optimizer (NorMuon + CautiousAdamW), WSD scheduler, BF16 + safetensors checkpoints, benchmark training/
data_pipeline.md 6-stage pipeline (download → preprocess → 64K BPE → tokenize → 4096×4096 shards → streaming mmap loader), source mix data/
utils.md Checkpoint, scheduler, validation, benchmark, async data loader notes training/

Conventions

  • Raw PyTorch only — no HuggingFace Trainer, no PyTorch Lightning, no DeepSpeed.
  • No magic numbers — every architectural constant is defined in code and its rationale is recorded here.
  • Separate documentation — decisions live in this folder, not in code comments.