Day 71 of building in public!
I've been testing out EXL3 as per
@0xSero's suggestion yesterday using
Quick RTX 3090 test: same Qwen2.5-1.5B checkpoint, BF16 vs EXL3.
BF16: 27 tok/s, 2,945 MiB
EXL3 4.0 BPW: 157 tok/s, 929 MiB
EXL3 3.0 BPW: 141 tok/s, 764 MiB
4.0 BPW decoded 5.8x faster.
Deep dive coming soon