Skip to content

Fix downcasting and upcasting #67

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
merged 1 commit into from
Jun 30, 2024
+10 −5
Merged

Fix downcasting and upcasting #67

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 10 additions & 5 deletions gemma/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -179,12 +179,14 @@ def _norm(self, x):
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

def forward(self, x):
x = self._norm(x.float()).type_as(x)
# Llama does x.to(float16) * w whilst Gemma2 is (x * w).to(float16)
# See https://github.com/huggingface/transformers/pull/29402
output = self._norm(x.float())
if self.add_unit_offset:
output = x * (1 + self.weight)
output = output * (1 + self.weight.float())
else:
output = x * self.weight
return output
output = output * self.weight.float()
return output.type_as(x)


class GemmaMLP(nn.Module):
Expand Down Expand Up @@ -546,7 +548,10 @@ def forward(
# [batch_size, input_len, hidden_size]
hidden_states = self.embedder(input_token_ids)
# Gemma normalizes the embedding by sqrt(hidden_size).
hidden_states = hidden_states * (self.config.hidden_size**0.5)
# Gemma2 downcasts the below to float16, causing sqrt(3072)=55.4256 to become 55.5
# See https://github.com/huggingface/transformers/pull/29402
normalizer = torch.tensor(self.config.hidden_size**0.5, dtype=hidden_states.dtype)
hidden_states = hidden_states * normalizer

hidden_states = self.model(
hidden_states=hidden_states,
Expand Down