Skip to content

Fix sparse attention for small block-sizes #1545

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
jeffra merged 9 commits into from
Nov 12, 2021
Merged

Fix sparse attention for small block-sizes #1545

Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
f7ef4b5
fixing the softmax masking when using triangular masking
Oct 27, 2021
dfb603f
Merge branch 'master' of github.com:microsoft/DeepSpeed
Oct 29, 2021
c5ecf32
Merge branch 'master' of github.com:microsoft/DeepSpeed
Nov 2, 2021
426ecf7
Merge branch 'master' of github.com:microsoft/DeepSpeed
Nov 8, 2021
fde6310
Merge branch 'master' of github.com:microsoft/DeepSpeed
Nov 9, 2021
542b25c
fixing the sparse attention for low block-size
Nov 9, 2021
d926bf7
remove attn
Nov 9, 2021
544c0bc
Merge branch 'master' into fix-sparse-attn
RezaYazdaniAminabadi Nov 9, 2021
abfe900
Merge branch 'master' into fix-sparse-attn
jeffra Nov 12, 2021
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
2 changes: 1 addition & 1 deletion deepspeed/ops/sparse_attention/matmul.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -289,7 +289,7 @@ def make_sdd_lut(layout, block, dtype, device):
#_sparse_matmul._load_utils()
#start_width = 64 // block
#segmented = _sparse_matmul.sdd_segment(layout.type(torch.int32), start_width)
start_width = 128 // block
start_width = (128 if block > 16 else 32) // block
layout = layout.type(torch.int32)
segmented = libtriton.superblock(layout.data_ptr(),
layout.shape[0],
Expand Down