Skip to content

Fix ECC correction range: ±0.5 ULP → ±1 ULP #88

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
josejg wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix ECC correction range: ±0.5 ULP → ±1 ULP #88

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions heavyball/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1129,7 +1129,7 @@ def __init__(self, correction, smax):
self.smax = smax

def decode(self, x):
ls = (_log_ulp(x) - 1).float()
ls = _log_ulp(x).float()
return x.float() + _scale_by_exp2(self.correction.float() / self.smax, ls)

@staticmethod
Expand All @@ -1141,7 +1141,7 @@ def _bf16_to_f32(x):
def compute_correction(self, fp32, narrow):
narrow_f32 = self._bf16_to_f32(narrow) if narrow.dtype == torch.bfloat16 else narrow.float()
e = fp32 - narrow_f32
ls = (_log_ulp(narrow) - 1).float()
ls = _log_ulp(narrow).float()
e_norm = _scale_by_exp2(e, -ls)
scaled = e_norm.clamp(-1.0, 1.0) * self.smax
self.correction.copy_(scaled.abs().add(0.5).floor().copysign(scaled).to(self.correction.dtype))
Expand Down