def forward(
self,
hidden_states,
attention_mask=None,
output_attentions=False,
output_hidden_states=False,
return_dict=True,
):
all_hidden_states = () if output_hidden_states else None
all_self_attentions = () if output_attentions else None
if attention_mask is not None:
# make sure padded tokens output 0
hidden_states[~attention_mask] = 0.0
# extend attention_mask
attention_mask = (1.0 - attention_mask[:, None, None, :].to(
dtype=hidden_states.dtype)) * -10000.0
attention_mask = attention_mask.expand(attention_mask.shape[0], 1,
attention_mask.shape[-1],
attention_mask.shape[-1])
position_embeddings = self.pos_conv_embed(hidden_states)
hidden_states = hidden_states + position_embeddings
hidden_states = self.layer_norm(hidden_states)
hidden_states = self.dropout(hidden_states)
deepspeed_zero3_is_enabled = is_deepspeed_zero3_enabled()
for layer in self.layers:
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states, )
# add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
dropout_probability = np.random.uniform(0, 1)
skip_the_layer = True if self.training and (
dropout_probability < self.config.layerdrop) else False
if not skip_the_layer or deepspeed_zero3_is_enabled:
# under deepspeed zero3 all gpus must run in sync
if getattr(self.config, "gradient_checkpointing",
False) and self.training:
# create gradient checkpointing function
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs, output_attentions)
return custom_forward
layer_outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(layer),
hidden_states,
attention_mask,
)
else:
layer_outputs = layer(hidden_states,
attention_mask=attention_mask,
output_attentions=output_attentions)
hidden_states = layer_outputs[0]
if skip_the_layer:
layer_outputs = (None, None)
if output_attentions:
all_self_attentions = all_self_attentions + (
layer_outputs[1], )
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states, )
if not return_dict:
return tuple(
v for v in
[hidden_states, all_hidden_states, all_self_attentions]
if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
)
def forward(
self,
hidden_states,
attention_mask=None,
output_attentions=False,
output_hidden_states=False,
return_dict=True,
):
all_hidden_states = () if output_hidden_states else None
all_self_attentions = () if output_attentions else None
if attention_mask is not None:
# make sure padded tokens output 0
hidden_states[~attention_mask] = 0.0
input_lengths = (attention_mask.long()).sum(-1)
# apply pooling formula to get real output_lengths
output_lengths = input_lengths // self.config.squeeze_factor
max_encoder_length = hidden_states.shape[1] // self.config.squeeze_factor
attention_ids = (
torch.arange(0, max_encoder_length, device=output_lengths.device)
.view(1, -1)
.expand(output_lengths.shape[0], -1)
)
attention_mask = (attention_ids < output_lengths.view(-1, 1)).long()
# extend attention_mask
attention_mask = (1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)) * -10000.0
attention_mask = attention_mask.expand(
attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1]
)
n_input_timesteps = hidden_states.shape[1]
hidden_states = hidden_states.transpose(1, 2)
position_embeddings = self.pos_conv_embed(hidden_states)
pooled_hidden_states = self.pool(hidden_states)
min_length = min(position_embeddings.size(-1), pooled_hidden_states.size(-1))
hidden_states = pooled_hidden_states[..., :min_length] + position_embeddings[..., :min_length]
hidden_states = hidden_states.transpose(1, 2)
hidden_states = self.layer_norm(hidden_states)
hidden_states = self.dropout(hidden_states)
deepspeed_zero3_is_enabled = is_deepspeed_zero3_enabled()
for layer in self.layers:
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
# add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
dropout_probability = np.random.uniform(0, 1)
skip_the_layer = True if self.training and (dropout_probability < self.config.layerdrop) else False
if not skip_the_layer or deepspeed_zero3_is_enabled:
# under deepspeed zero3 all gpus must run in sync
if self.gradient_checkpointing and self.training:
# create gradient checkpointing function
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs, output_attentions)
return custom_forward
layer_outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(layer),
hidden_states,
attention_mask,
)
else:
layer_outputs = layer(
hidden_states, attention_mask=attention_mask, output_attentions=output_attentions
)
hidden_states = layer_outputs[0]
if skip_the_layer:
layer_outputs = (None, None)
if output_attentions:
all_self_attentions = all_self_attentions + (layer_outputs[1],)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
hidden_states = self.upsample(hidden_states)
if hidden_states.shape[1] < n_input_timesteps:
hidden_states = nn.functional.pad(hidden_states, (0, 0, 0, n_input_timesteps - hidden_states.shape[1]))
if not return_dict:
return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
)
