Andrej Karpathy 2ff7d51252 integrate Flash Attention 3. +9% tok_per_sec for d12 with ctx even as low as 2048 out of the box nice. also, ready to tune windows huge

2026-01-11 20:33:19 +00:00

..

test_engine.py

integrate Flash Attention 3. +9% tok_per_sec for d12 with ctx even as low as 2048 out of the box nice. also, ready to tune windows huge

2026-01-11 20:33:19 +00:00