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

Bloom att mask #1

Merged
merged 4 commits into from
Jul 20, 2022
Merged

Bloom att mask #1

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
6870de6
fix causal_mask
NouamaneTazi Jul 18, 2022
69e72e1
fix softmax for float16
NouamaneTazi Jul 18, 2022
6815664
styling
NouamaneTazi Jul 18, 2022
93eaaff
revert back to using addition for some speed
NouamaneTazi Jul 18, 2022
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: 5 additions & 4 deletions src/transformers/models/bloom/modeling_bloom.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,7 +56,7 @@ def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_
Make causal mask used for bi-directional self-attention.
"""
batch_size, target_length = input_ids_shape
mask = torch.full((target_length, target_length), torch.finfo(dtype).min)
mask = torch.full((target_length, target_length), -torch.inf)
mask_cond = torch.arange(mask.size(-1))
intermediate_mask = mask_cond < (mask_cond + 1).view(mask.size(-1), 1)
mask.masked_fill_(intermediate_mask, 0)
Expand All @@ -79,7 +79,7 @@ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: int = None):

inverted_mask = 1.0 - expanded_mask

return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), -torch.inf)


def build_alibi_tensor(attention_mask: torch.Tensor, n_head: int, dtype, device) -> torch.Tensor:
Expand Down Expand Up @@ -303,7 +303,9 @@ def forward(
# We replace the scaled softmax by just a few line of code - [batch_size, num_heads, q_length, k_length]
input_dtype = attention_scores.dtype
attn_weights = (attention_scores * self.layer_number) + attention_mask
attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
attn_weights = torch.clip(
attn_weights, torch.finfo(attn_weights.dtype).min, torch.finfo(attn_weights.dtype).max
)
attention_probs = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(input_dtype)
attention_probs = attention_probs * (~attention_mask.bool())
# [batch_size, num_heads, q_length, k_length]
Expand Down Expand Up @@ -599,7 +601,6 @@ def _prepare_attn_mask(self, attention_mask, input_shape, inputs_embeds, past_ke
combined_attention_mask = (
expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
)

return combined_attention_mask

def set_input_embeddings(self, new_embeddings):
Expand Down