Skip to content
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

Addition of booleans is currently wrong in iree-compile #341

Merged
merged 3 commits into from
Oct 28, 2024
Merged

Addition of booleans is currently wrong in iree-compile #341

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1268ee5
Addition of booleans is currently wrong in iree-compile
rsuderman Oct 26, 2024
1857ebe
move change to decode_attention_mask too
rsuderman Oct 26, 2024
4482cf1
Merge branch 'main' into fix_masking
renxida Oct 28, 2024
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
14 changes: 7 additions & 7 deletions sharktank/sharktank/layers/causal_llm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,10 +95,9 @@ def input_mask(

def decode_attention_mask(self, boolean_input_mask: torch.Tensor):
dtype = self.attention_dtype
numeric_mask = torch.zeros_like(boolean_input_mask, dtype=dtype)
numeric_mask.masked_fill_(
boolean_input_mask, self._maximally_negative_value(dtype)
)
numeric_mask = torch.where(
renxida marked this conversation as resolved.
Show resolved Hide resolved
boolean_input_mask, self._maximally_negative_value(dtype), 0
).to(dtype)
return numeric_mask.unsqueeze(1).unsqueeze(1).to(self.device)

def attention_mask(
Expand Down Expand Up @@ -127,9 +126,10 @@ def attention_mask(
dtype = self.attention_dtype
_, batch_seq_len = input_mask.shape
causal_mask = causal_context_mask[:, :, :batch_seq_len, :batch_seq_len]
boolean_mask = causal_mask + input_mask[:, None, None, :]
numeric_mask = torch.zeros_like(boolean_mask, dtype=dtype)
numeric_mask.masked_fill_(boolean_mask, self._maximally_negative_value(dtype))
boolean_mask = torch.logical_or(causal_mask, input_mask[:, None, None, :])
numeric_mask = torch.where(
boolean_mask, self._maximally_negative_value(dtype), 0
).to(dtype)
return numeric_mask.to(self.device)

def extract_tokens_from_logits(
Expand Down
Loading