Skip to content

SpinQuant rotate bias #2913

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

Merged
merged 1 commit into from
Sep 2, 2025
Merged

SpinQuant rotate bias #2913

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
15 changes: 15 additions & 0 deletions torchao/prototype/spinquant/hadamard_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -237,16 +237,24 @@ def apply_exact_had_to_linear(module, had_dim=-1, output=False, R2=None):
assert is_pow2(had_dim), "Hadamard dimension must be a power of 2!"

W = module.weight.data
if module.bias is not None:
B = module.bias.data
bias_dtype_orig = B.dtype
B = B.float()
dtype_orig = W.dtype
W = W.float()

if had_dim == -1:
if output:
had_K, K = get_hadK(out_features)
W = matmul_hadU(W.t(), had_K.to(W.device), K).t()
if module.bias is not None:
B = matmul_hadU(B, had_K.to(B.device), K)
else:
had_K, K = get_hadK(in_features)
W = matmul_hadU(W, had_K.to(W.device), K)
if module.bias is not None:
B = matmul_hadU(B, had_K.to(B.device), K)
else:
if R2 is not None:
hadK = R2.to(torch.float64)
Expand All @@ -260,8 +268,15 @@ def apply_exact_had_to_linear(module, had_dim=-1, output=False, R2=None):
temp = W.reshape(-1, shape[-1] // had_dim, had_dim)
temp = temp.to(torch.float64) @ hadK
W = temp.reshape(shape)
if module.bias is not None:
shape = B.shape
temp = B.reshape(-1, had_dim)
temp = temp.to(torch.float64) @ hadK
B = temp.reshape(shape)

if output:
W = W.t()

module.weight.data = W.to(dtype=dtype_orig)
if module.bias is not None:
module.bias.data = B.to(dtype=bias_dtype_orig)
Loading