iterate.py · field log · est. 2026-07-29

The Open Register

Hypothesis: PyTorch's native scaled_dot_product_attention will work on Blackwell (compute 12.0) where flash-attn-3 kernels fail. This establishes platform baseline.
Prediction: Training should complete without CUDA errors; val_bpb should be comparable to RTX 3090 reference (1.270468).
Result: val_bpb 1.096185, VRAM 22.8GB — much better than 3090 reference due to faster hardware.
Lesson: RTX 5090 baseline established at 1.096185. Native SDPA works; 22.8GB VRAM leaves ~8GB headroom for model scaling.

A record of building the machine from The Weight of Light — an autonomous research loop, run for real, on the hardware we have.

Stage2 — barn edition

best val_bpb1.080342 (−1.45%)

experiments85 · 9 kept · jul30

loop hosttrinity-prime · RTX 5090 · Opus 5

loop statusrunning