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

torch.flip issue is fixed in SynapseAI 1.16, so remove the WA in opti… #1092

Merged
merged 2 commits into from
Jul 22, 2024
Merged

torch.flip issue is fixed in SynapseAI 1.16, so remove the WA in opti… #1092

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9cc3f75
torch.flip issue is fixed in SynapseAI 1.16, so remove the WA in opti…
sywangyi Jun 24, 2024
9f5e6fb
Apply suggestions from code review
sywangyi Jul 11, 2024
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
2 changes: 0 additions & 2 deletions optimum/habana/transformers/modeling_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -177,7 +177,6 @@
gaudi_unconstrained_rational_quadratic_spline,
gaudi_VisionEncoderDecoderModel_prepare_inputs_for_generation,
gaudi_vit_self_attention_forward,
gaudi_VitsResidualCouplingLayer_forward,
gaudi_wav2vec2_encoder_forward,
gaudi_wav2vec2_forward,
gaudi_wav2vec2_tdnnlayer_forward,
Expand Down Expand Up @@ -490,7 +489,6 @@ def adapt_transformers_to_gaudi():
transformers.models.vits.modeling_vits._unconstrained_rational_quadratic_spline = (
gaudi_unconstrained_rational_quadratic_spline
)
transformers.models.vits.modeling_vits.VitsResidualCouplingLayer.forward = gaudi_VitsResidualCouplingLayer_forward

# Optimization for starcoder2 on Gaudi
transformers.models.starcoder2.modeling_starcoder2.Starcoder2ForCausalLM = GaudiStarcoder2ForCausalLM
Expand Down
5 changes: 1 addition & 4 deletions optimum/habana/transformers/models/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -193,10 +193,7 @@
gaudi_VisionEncoderDecoderModel_prepare_inputs_for_generation,
)
from .vit import gaudi_vit_self_attention_forward
from .vits import (
gaudi_unconstrained_rational_quadratic_spline,
gaudi_VitsResidualCouplingLayer_forward,
)
from .vits import gaudi_unconstrained_rational_quadratic_spline
from .wav2vec2 import (
_gaudi_wav2vec2_compute_mask_indices,
_gaudi_wav2vec2_mask_hidden_states,
Expand Down
5 changes: 1 addition & 4 deletions optimum/habana/transformers/models/vits/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1 @@
from .modeling_vits import (
gaudi_unconstrained_rational_quadratic_spline,
gaudi_VitsResidualCouplingLayer_forward,
)
from .modeling_vits import gaudi_unconstrained_rational_quadratic_spline
23 changes: 0 additions & 23 deletions optimum/habana/transformers/models/vits/modeling_vits.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,26 +52,3 @@ def gaudi_unconstrained_rational_quadratic_spline(
outputs = outputs_i * inside_interval_mask + outputs * outside_interval_mask
log_abs_det = log_abs_det_i * inside_interval_mask + log_abs_det * outside_interval_mask
return outputs, log_abs_det


def gaudi_VitsResidualCouplingLayer_forward(self, inputs, padding_mask, global_conditioning=None, reverse=False):
"""
Copied from VitsResidualCouplingLayer:forward: https://github.com/huggingface/transformers/blob/v4.38.2/src/transformers/models/vits/modeling_vits.py
The only differences are:
- WA to fix torch.flip issue after conv1d
"""
first_half, second_half = torch.split(inputs, [self.half_channels] * 2, dim=1)
hidden_states = self.conv_pre(first_half) * padding_mask
hidden_states = self.wavenet(hidden_states, padding_mask, global_conditioning)
mean = self.conv_post(hidden_states) * padding_mask
log_stddev = torch.zeros_like(mean)

if not reverse:
second_half = mean.cpu() + second_half * torch.exp(log_stddev) * padding_mask
outputs = torch.cat([first_half, second_half], dim=1)
log_determinant = torch.sum(log_stddev, [1, 2])
return outputs, log_determinant
else:
second_half = (second_half - mean.cpu()) * torch.exp(-log_stddev) * padding_mask
outputs = torch.cat([first_half, second_half], dim=1)
return outputs, None
Loading