Skip to content

[Fix] Fixed a bug related to sparse logic. #4

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

[Fix] Fixed a bug related to sparse logic. #4

Show file tree
Hide file tree
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
9 changes: 8 additions & 1 deletion marlin/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -295,7 +295,7 @@ def __init__(self, infeatures, outfeatures, groupsize=-1):
self.n = outfeatures
self.groupsize = groupsize
self.register_buffer(
"B", torch.empty((self.k // 16, self.n * 16 // 8), dtype=torch.int)
"B", torch.empty((self.k // 16 // 2, self.n * 16 // 8), dtype=torch.int)
)
self.register_buffer(
"meta", torch.empty((self.n, self.k // 16), dtype=torch.int16)
Expand Down Expand Up @@ -365,8 +365,15 @@ def pack(self, linear, scales, trans=False):
s = s.reshape((-1, len(scale_perm_single)))[:, scale_perm_single]

mask = mask_creator(w.T).cuda().bool()
# Avoid confusing the pruned elements and the zero elements (-8 + (maxq + 1) // 2)
w += 1

w = mask * w.T
w, meta = sparse_semi_structured_from_dense_cutlass(w)

# Reover the original values
w -= 1

w = w.t()
self.k = self.k // 2
self.groupsize = self.groupsize // 2
Expand Down
5 changes: 5 additions & 0 deletions marlin/_semi_structured_conversions.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -304,6 +304,11 @@ def mask_creator(tensor):

num_groups = tensor.numel() // M

# Subtract the offset value for pruning
maxq = 2**4 - 1
ZERO_VALUE = (maxq + 1) // 2
tensor = tensor - ZERO_VALUE

# N:M sparsity for linear layers
tensor_temp = tensor.detach().abs().reshape(num_groups, M)
index = torch.argsort(tensor_temp, dim=1)[:, : int(M - N)]
Expand Down