Skip to content

Fix PyTorch Perceiver PerceiverFourierPositionEncoding with fp16 #21787

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
amyeroberts merged 2 commits into from
Feb 27, 2023
Merged

Fix PyTorch Perceiver PerceiverFourierPositionEncoding with fp16 #21787

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3583730
fix perceiver fp16
fxmarty Feb 24, 2023
74201c6
hopefully fix tests
fxmarty Feb 24, 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
19 changes: 13 additions & 6 deletions src/transformers/models/perceiver/modeling_perceiver.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2201,7 +2201,7 @@ def decoder_query(self, inputs, modality_sizes=None, inputs_without_pos=None, su
pos_emb = self.output_position_encodings(batch_size)
elif self.position_encoding_type == "fourier":
pos_emb = self.output_position_encodings(
self.output_index_dims, batch_size=batch_size, device=inputs.device, pos=pos
self.output_index_dims, batch_size=batch_size, device=inputs.device, dtype=inputs.dtype, pos=pos
)

# Optionally project them to a target dimension.
Expand All @@ -2215,7 +2215,9 @@ def decoder_query(self, inputs, modality_sizes=None, inputs_without_pos=None, su
if self.position_encoding_type == "trainable":
pos_emb = self.output_position_encodings(batch_size)
elif self.position_encoding_type == "fourier":
pos_emb = self.output_position_encodings(index_dims, batch_size, device=inputs.device)
pos_emb = self.output_position_encodings(
index_dims, batch_size, device=inputs.device, dtype=inputs.dtype
)

# Optionally project them to a target dimension.
pos_emb = self.positions_projection(pos_emb)
Expand Down Expand Up @@ -2816,7 +2818,12 @@ def output_size(self):
return encoding_size

def forward(
self, index_dims: List[int], batch_size: int, device, pos: torch.FloatTensor = None
self,
index_dims: List[int],
batch_size: int,
device: torch.device,
dtype: torch.dtype,
pos: torch.FloatTensor = None,
) -> torch.FloatTensor:
pos = _check_or_build_spatial_positions(pos, index_dims, batch_size)
fourier_pos_enc = generate_fourier_features(
Expand All @@ -2825,7 +2832,7 @@ def forward(
max_resolution=self.max_resolution,
concat_pos=self.concat_pos,
sine_only=self.sine_only,
).to(device)
).to(device=device, dtype=dtype)
return fourier_pos_enc


Expand Down Expand Up @@ -3156,7 +3163,7 @@ def _build_network_inputs(self, inputs: torch.Tensor, network_input_is_1d: bool
if self.position_encoding_type == "trainable":
pos_enc = self.position_embeddings(batch_size)
elif self.position_encoding_type == "fourier":
pos_enc = self.position_embeddings(index_dims, batch_size, device=inputs.device)
pos_enc = self.position_embeddings(index_dims, batch_size, device=inputs.device, dtype=inputs.dtype)

# Optionally project them to a target dimension.
pos_enc = self.positions_projection(pos_enc)
Expand Down Expand Up @@ -3324,7 +3331,7 @@ def _build_network_inputs(self, inputs):
if self.position_encoding_type == "trainable":
pos_enc = self.position_embeddings(batch_size)
elif self.position_encoding_type == "fourier":
pos_enc = self.position_embeddings(index_dims, batch_size, device=inputs.device)
pos_enc = self.position_embeddings(index_dims, batch_size, device=inputs.device, dtype=inputs.dtype)

# Optionally project them to a target dimension.
pos_enc = self.positions_projection(pos_enc)
Expand Down