Lgr54HFi
/

chomera

Model card Files Files and versions

311 kB

Ctrl+K

Ctrl+K

1 contributor

History: 6 commits

Lgr54HFi's picture

perf: re-enable torch.compile now that STE uses detach() trick (zero graph breaks)"

dd57d33 verified 15 days ago

chimera
perf: replace _RoundTernarySTE autograd.Function with detach() trick — zero graph breaks for torch.compile\n\nThe detach() identity pattern (w + (round(clamp(w)) - w).detach()) is\nmathematically equivalent to the old STE but uses only standard aten ops\nthat torch.compile/Inductor can trace through. This eliminates 84+\ngraph breaks, enabling full kernel fusion of quantize+linear.\n\nPattern from official BitNet b1.58 implementation (1bitLLM/bitnet_b1_58-large).\nRef: arxiv 2402.17764" 15 days ago
tests
Upload folder using huggingface_hub 15 days ago
.gitattributes

1.52 kB
initial commit 15 days ago
.gitignore

230 Bytes
Upload folder using huggingface_hub 15 days ago
README.md

9.11 kB
Upload folder using huggingface_hub 15 days ago
chimera_turbo.py

17.1 kB
perf: re-enable torch.compile now that STE uses detach() trick (zero graph breaks)" 15 days ago
config.json

25.5 kB
Upload folder using huggingface_hub 15 days ago
gguf_import.py

29.1 kB
Upload folder using huggingface_hub 15 days ago
inference.py

12.3 kB
Upload folder using huggingface_hub 15 days ago
launch_turbo.sh

1.9 kB
Upload folder using huggingface_hub 15 days ago
pyproject.toml

789 Bytes
Upload folder using huggingface_hub 15 days ago
train.py

9.22 kB
Upload folder using huggingface_hub 15 days ago
train_fast.py

4.96 kB
Upload folder using huggingface_hub 15 days ago
train_hyper.py

8.73 kB
Upload folder using huggingface_hub 15 days ago