Skip to content

Add bf16 support #5

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
wants to merge 3 commits into
base: gma/cpu_support
Choose a base branch
from
Open

Add bf16 support #5

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
2c89fca
rebase to guokai branch
baodii Mar 13, 2023
0321941
fix softmaxOp: alibi index bug
baodii Mar 23, 2023
d987f44
Merge branch 'gma/cpu_support' into add_bf16_support
baodii Mar 23, 2023
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
5 changes: 5 additions & 0 deletions .gitignore
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,3 +29,8 @@ tests/unit/saved_checkpoint/
# Dev/IDE data
.vscode
.theia

# ctags and cscope
cscope.*
cscope/
tags
16 changes: 11 additions & 5 deletions deepspeed/ops/transformer/inference/op_binding/softmax.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,18 +45,24 @@ def forward(self,
self.config.mp_size)
else:
# fallback
alibi = alibi[head_offset:head_offset +
self.num_attention_heads_per_partition]
if alibi is not None:
bs = attn_scores.shape[0]
total_heads = alibi.shape[0]
heads = int(total_heads / bs)
seq_length = alibi.shape[-1]
alibi = alibi.view(bs, heads, 1, seq_length)

alibi_split = alibi[:, head_offset:head_offset +
self.num_attention_heads_per_partition, :, :]
attn_scores = attn_scores * layer_scale + alibi_split
input_dtype = attn_scores.dtype
if (triangular):
tri = ~torch.tril(
torch.ones(attn_scores.size(),
device=attn_scores.device)).to(bool)
attn_scores = torch.masked_fill(attn_scores * layer_scale,
attn_scores = torch.masked_fill(attn_scores,
tri,
torch.finfo(input_dtype).min)
if alibi is not None:
attn_scores += alibi
if attn_mask is not None:
# expand atten_mask from two dim into 4 dim, insert two dims in the middle
attn_mask = attn_mask[:, None, None, :]
Expand Down