Skip to content

Fix BLOOM's softmax for half precisions #18185

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

Closed
NouamaneTazi wants to merge 4 commits into from
Closed

Fix BLOOM's softmax for half precisions #18185

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
Copy link
Contributor

@thomasw21 thomasw21 Jul 18, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Stupid question but when you sum torch.finfo(dtype).min with torch.finfo(dtype).min, it's not the masked value anymore?

Copy link
Member Author

@NouamaneTazi NouamaneTazi Jul 19, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

torch.finfo(dtype).min + torch.finfo(dtype).min = -inf

Copy link
Member Author

@NouamaneTazi NouamaneTazi Jul 19, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We would get something like this:

>> print(attention_mask)
tensor([[[[     0., -65504., -65504., -65504., -65504., -65504., -65504.],
          [     0.,      0., -65504., -65504., -65504., -65504., -65504.],
          [     0.,      0.,      0., -65504., -65504., -65504., -65504.],
          [     0.,      0.,      0.,      0., -65504., -65504., -65504.],
          [     0.,      0.,      0.,      0.,      0., -65504., -65504.],
          [     0.,      0.,      0.,      0.,      0.,      0., -65504.],
          [     0.,      0.,      0.,      0.,      0.,      0.,      0.]]],


        [[[-65504.,    -inf,    -inf, -65504., -65504., -65504., -65504.],
          [-65504., -65504.,    -inf, -65504., -65504., -65504., -65504.],
          [-65504., -65504., -65504., -65504., -65504., -65504., -65504.],
          [-65504., -65504., -65504.,      0., -65504., -65504., -65504.],
          [-65504., -65504., -65504.,      0.,      0., -65504., -65504.],
          [-65504., -65504., -65504.,      0.,      0.,      0., -65504.],
          [-65504., -65504., -65504.,      0.,      0.,      0.,      0.]]]],
       device='cuda:0', dtype=torch.float16)

)

return combined_attention_mask

def set_input_embeddings(self, new_embeddings):
Expand Down