Skip to content

Wider support for gelu. Remove unused activation. Use same torch layer for silu and swish #3302

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
wants to merge 5 commits into from
Closed

Wider support for gelu. Remove unused activation. Use same torch layer for silu and swish #3302

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
25f5342
wider support for gelu. use same torch layer for silu and swish
hypnopump May 1, 2023
937ee18
remove useless import
hypnopump May 1, 2023
67ff580
remove unused activation
hypnopump May 1, 2023
1f26ebe
raise error in construction if param is needed at runtime.
hypnopump May 1, 2023
b88e7fd
Merge branch 'main' into support_gelu_fuse_silu_swish
hypnopump May 4, 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
6 changes: 2 additions & 4 deletions src/diffusers/models/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -520,12 +520,10 @@ def __init__(
self.num_groups = num_groups
self.eps = eps
self.act = None
if act_fn == "swish":
self.act = lambda x: F.silu(x)
if act_fn in {"swish", "silu"}:
self.act = nn.SiLU()
elif act_fn == "mish":
self.act = nn.Mish()
elif act_fn == "silu":
self.act = nn.SiLU()
elif act_fn == "gelu":
self.act = nn.GELU()

Expand Down
6 changes: 2 additions & 4 deletions src/diffusers/models/resnet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -513,12 +513,10 @@ def __init__(
conv_2d_out_channels = conv_2d_out_channels or out_channels
self.conv2 = torch.nn.Conv2d(out_channels, conv_2d_out_channels, kernel_size=3, stride=1, padding=1)

if non_linearity == "swish":
self.nonlinearity = lambda x: F.silu(x)
if non_linearity in {"swish", "silu"}:
self.nonlinearity = nn.SiLU()
elif non_linearity == "mish":
self.nonlinearity = nn.Mish()
elif non_linearity == "silu":
self.nonlinearity = nn.SiLU()
elif non_linearity == "gelu":
self.nonlinearity = nn.GELU()

Expand Down
34 changes: 14 additions & 20 deletions src/diffusers/models/unet_1d_blocks.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,12 +55,12 @@ def __init__(

self.resnets = nn.ModuleList(resnets)

if non_linearity == "swish":
self.nonlinearity = lambda x: F.silu(x)
if non_linearity in {"swish", "silu"}:
self.nonlinearity = nn.SiLU()
elif non_linearity == "mish":
self.nonlinearity = nn.Mish()
elif non_linearity == "silu":
self.nonlinearity = nn.SiLU()
elif non_linearity == "gelu":
self.nonlinearity = nn.GELU()
else:
self.nonlinearity = None

Expand Down Expand Up @@ -119,12 +119,12 @@ def __init__(

self.resnets = nn.ModuleList(resnets)

if non_linearity == "swish":
self.nonlinearity = lambda x: F.silu(x)
if non_linearity in {"swish", "silu"}:
self.nonlinearity = nn.SiLU()
elif non_linearity == "mish":
self.nonlinearity = nn.Mish()
elif non_linearity == "silu":
self.nonlinearity = nn.SiLU()
elif non_linearity == "gelu":
self.nonlinearity = nn.GELU()
else:
self.nonlinearity = None

Expand Down Expand Up @@ -179,7 +179,6 @@ def __init__(
num_layers: int = 1,
add_downsample: bool = False,
add_upsample: bool = False,
non_linearity=None,
):
super().__init__()
self.in_channels = in_channels
Expand All @@ -194,15 +193,6 @@ def __init__(

self.resnets = nn.ModuleList(resnets)

if non_linearity == "swish":
self.nonlinearity = lambda x: F.silu(x)
elif non_linearity == "mish":
self.nonlinearity = nn.Mish()
elif non_linearity == "silu":
self.nonlinearity = nn.SiLU()
else:
self.nonlinearity = None

self.upsample = None
if add_upsample:
self.upsample = Downsample1D(out_channels, use_conv=True)
Expand Down Expand Up @@ -232,10 +222,14 @@ def __init__(self, num_groups_out, out_channels, embed_dim, act_fn):
super().__init__()
self.final_conv1d_1 = nn.Conv1d(embed_dim, embed_dim, 5, padding=2)
self.final_conv1d_gn = nn.GroupNorm(num_groups_out, embed_dim)
if act_fn == "silu":
if act_fn in {"silu", "swish"}:
self.final_conv1d_act = nn.SiLU()
if act_fn == "mish":
elif act_fn == "mish":
self.final_conv1d_act = nn.Mish()
elif act_fn == "gelu":
self.final_conv1d_act = nn.GELU()
else:
raise ValueError(f"Act_fn {act_fn} must be one of silu, mish, gelu")
self.final_conv1d_2 = nn.Conv1d(embed_dim, out_channels, 1)

def forward(self, hidden_states, temb=None):
Expand Down
13 changes: 4 additions & 9 deletions src/diffusers/models/unet_2d_condition.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,6 @@

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint

from ..configuration_utils import ConfigMixin, register_to_config
Expand Down Expand Up @@ -295,12 +294,10 @@ def __init__(

if time_embedding_act_fn is None:
self.time_embed_act = None
elif time_embedding_act_fn == "swish":
self.time_embed_act = lambda x: F.silu(x)
elif time_embedding_act_fn in {"swish", "silu"}:
self.time_embed_act = nn.SiLU()
elif time_embedding_act_fn == "mish":
self.time_embed_act = nn.Mish()
elif time_embedding_act_fn == "silu":
self.time_embed_act = nn.SiLU()
elif time_embedding_act_fn == "gelu":
self.time_embed_act = nn.GELU()
else:
Expand Down Expand Up @@ -458,12 +455,10 @@ def __init__(
num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
)

if act_fn == "swish":
self.conv_act = lambda x: F.silu(x)
if act_fn in {"swish", "silu"}:
self.conv_act = nn.SiLU()
elif act_fn == "mish":
self.conv_act = nn.Mish()
elif act_fn == "silu":
self.conv_act = nn.SiLU()
elif act_fn == "gelu":
self.conv_act = nn.GELU()
else:
Expand Down
12 changes: 4 additions & 8 deletions src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -398,12 +398,10 @@ def __init__(

if time_embedding_act_fn is None:
self.time_embed_act = None
elif time_embedding_act_fn == "swish":
self.time_embed_act = lambda x: F.silu(x)
elif time_embedding_act_fn in {"swish", "silu"}:
self.time_embed_act = nn.SiLU()
elif time_embedding_act_fn == "mish":
self.time_embed_act = nn.Mish()
elif time_embedding_act_fn == "silu":
self.time_embed_act = nn.SiLU()
elif time_embedding_act_fn == "gelu":
self.time_embed_act = nn.GELU()
else:
Expand Down Expand Up @@ -561,12 +559,10 @@ def __init__(
num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
)

if act_fn == "swish":
self.conv_act = lambda x: F.silu(x)
if act_fn in {"swish", "silu"}:
self.conv_act = nn.SiLU()
elif act_fn == "mish":
self.conv_act = nn.Mish()
elif act_fn == "silu":
self.conv_act = nn.SiLU()
elif act_fn == "gelu":
self.conv_act = nn.GELU()
else:
Expand Down