2.95 MB

Ctrl+K

2 contributors

History: 86 commits

Quazim0t0

Claude Opus 5.5

Re-check FTZ/DAZ before every optimizer step (2026-10-20)

a28437c about 14 hours ago

config
Old-hardware training through emulated GPU logic 3 months ago
daisychain
Re-check FTZ/DAZ before every optimizer step (2026-10-20) about 14 hours ago
docker
Old-hardware training through emulated GPU logic 3 months ago
docs
Old-hardware training through emulated GPU logic 3 months ago
examples
Old-hardware training through emulated GPU logic 3 months ago
scripts
Old-hardware training through emulated GPU logic 3 months ago
web
One arithmetic on every device: mixed CPU+GPU optimizer, replica hash check, FTZ pin, subnormal gate, live f32 audit, probe/fragment fixes (2026-10-14) 7 days ago
.gitattributes

1.52 kB
initial commit 3 months ago
.gitignore

119 Bytes
Old-hardware training through emulated GPU logic 3 months ago
LICENSE

1.09 kB
Old-hardware training through emulated GPU logic 3 months ago
README.md

33.9 kB
One arithmetic on every device: mixed CPU+GPU optimizer, replica hash check, FTZ pin, subnormal gate, live f32 audit, probe/fragment fixes (2026-10-14) 7 days ago
export_luts_web.py

3.06 kB
Refuse NaN/Inf at every int8 quantize; requant rounds half-up (retrained); bounded units; README for September 24 days ago
pyproject.toml

1.06 kB
Update URLs for DaisyChain-Train 3 months ago
requirements.txt

35 Bytes
Old-hardware training through emulated GPU logic 3 months ago
test_cluster_mixed.py

3.83 kB
One arithmetic on every device: mixed CPU+GPU optimizer, replica hash check, FTZ pin, subnormal gate, live f32 audit, probe/fragment fixes (2026-10-14) 7 days ago
test_cluster_reductions.py

5.69 kB
Re-check FTZ/DAZ before every optimizer step (2026-10-20) about 14 hours ago
test_verified_units.py

8.35 kB
Refuse NaN/Inf at every int8 quantize; requant rounds half-up (retrained); bounded units; README for September 24 days ago