Skip to content

ARCHITECTURE_DECISION.md

Latest commit

 

History

History
46 lines (30 loc) · 1.34 KB

ARCHITECTURE_DECISION.md

File metadata and controls

46 lines (30 loc) · 1.34 KB
Raw

Architecture Decision: v1 as Base

Date: 2026-05-18

Decision

Use the v1 sparse-edge architecture as the default base for future work.

Why

The matched stress run below shows better speed, lower VRAM, and stronger learning signal for v1.

Matched Run (v1 vs v2)

Command used on both repositories:

python3 brain3d.py --grid 48 --steps 200 --strong-input --3d-inputs

Environment:

  • GPU: NVIDIA GeForce GTX 1050 (4 GB)
  • PyTorch: 2.6.0+cu126
  • Python env: /tmp/RetroCodeMess/pyenv-qtun-cu126
Model Steps/s Post-init VRAM Post-sim VRAM Final VRAM Corr (L/R) Spike (L/R) Weight Delta
v1 28.04 217 MiB 351 MiB 239 MiB 0.707 / 0.752 1.000 / 0.938 +11,014,204
v2 10.71 1075 MiB 1965 MiB 1965 MiB 0.002 / -0.004 0.000 / 0.000 +0.000000

Selective v2 Cherry-Picks Kept in v1

  • Added --device {auto,cpu,cuda} CLI option.
  • Added fail-fast CLI validation for --grid and --steps.
  • Added CUDA OOM graceful retry on CPU (single retry) for both --device auto and --device cuda.
  • Added elapsed-time print and cross-pattern correlation prints for easier benchmarking.

Rule for Future Cherry-Picks

Only keep changes that improve at least one of:

  1. learning signal,
  2. throughput,
  3. VRAM efficiency,

without regressing the other two materially.