jinx/ComfyUI
1
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-08-02 15:04:50 +08:00
Code Issues Activity

Support for Control Loras.

Browse Source 
Control loras are controlnets where some of the weights are stored in
"lora" format: an up and a down low rank matrice that when multiplied
together and added to the unet weight give the controlnet weight.

This allows a much smaller memory footprint depending on the rank of the
matrices.

These controlnets are used just like regular ones.
This commit is contained in:
comfyanonymous
2023-08-18 02:46:11 -04:00
parent 39ac856a33
commit d6e4b342e6
6 changed files with 216 additions and 92 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
comfy/ldm/modules/diffusionmodules/openaimodel.py
View File

@@ -8,8 +8,6 @@ import torch.nn.functional as F
from .util import (
checkpoint,
conv_nd,
linear,
avg_pool_nd,
zero_module,
normalization,
@@ -17,7 +15,7 @@ from .util import (
)
from ..attention import SpatialTransformer
from comfy.ldm.util import exists
import comfy.ops
class TimestepBlock(nn.Module):
"""
@@ -72,14 +70,14 @@ class Upsample(nn.Module):
upsampling occurs in the inner-two dimensions.
"""
def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None):
def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=None):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.dims = dims
if use_conv:
self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding, dtype=dtype, device=device)
self.conv = operations.conv_nd(dims, self.channels, self.out_channels, 3, padding=padding, dtype=dtype, device=device)
def forward(self, x, output_shape=None):
assert x.shape[1] == self.channels
@@ -108,7 +106,7 @@ class Downsample(nn.Module):
downsampling occurs in the inner-two dimensions.
"""
def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None):
def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=None):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
@@ -116,7 +114,7 @@ class Downsample(nn.Module):
self.dims = dims
stride = 2 if dims != 3 else (1, 2, 2)
if use_conv:
self.op = conv_nd(
self.op = operations.conv_nd(
dims, self.channels, self.out_channels, 3, stride=stride, padding=padding, dtype=dtype, device=device
)
else:
@@ -158,6 +156,7 @@ class ResBlock(TimestepBlock):
down=False,
dtype=None,
device=None,
operations=None
):
super().__init__()
self.channels = channels
@@ -171,7 +170,7 @@ class ResBlock(TimestepBlock):
self.in_layers = nn.Sequential(
nn.GroupNorm(32, channels, dtype=dtype, device=device),
nn.SiLU(),
conv_nd(dims, channels, self.out_channels, 3, padding=1, dtype=dtype, device=device),
operations.conv_nd(dims, channels, self.out_channels, 3, padding=1, dtype=dtype, device=device),
)
self.updown = up or down
@@ -187,7 +186,7 @@ class ResBlock(TimestepBlock):
self.emb_layers = nn.Sequential(
nn.SiLU(),
linear(
operations.Linear(
emb_channels,
2 * self.out_channels if use_scale_shift_norm else self.out_channels, dtype=dtype, device=device
),
@@ -197,18 +196,18 @@ class ResBlock(TimestepBlock):
nn.SiLU(),
nn.Dropout(p=dropout),
zero_module(
conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1, dtype=dtype, device=device)
operations.conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1, dtype=dtype, device=device)
),
)
if self.out_channels == channels:
self.skip_connection = nn.Identity()
elif use_conv:
self.skip_connection = conv_nd(
self.skip_connection = operations.conv_nd(
dims, channels, self.out_channels, 3, padding=1, dtype=dtype, device=device
)
else:
self.skip_connection = conv_nd(dims, channels, self.out_channels, 1, dtype=dtype, device=device)
self.skip_connection = operations.conv_nd(dims, channels, self.out_channels, 1, dtype=dtype, device=device)
def forward(self, x, emb):
"""
@@ -317,6 +316,7 @@ class UNetModel(nn.Module):
adm_in_channels=None,
transformer_depth_middle=None,
device=None,
operations=comfy.ops,
):
super().__init__()
assert use_spatial_transformer == True, "use_spatial_transformer has to be true"
@@ -379,9 +379,9 @@ class UNetModel(nn.Module):
time_embed_dim = model_channels * 4
self.time_embed = nn.Sequential(
linear(model_channels, time_embed_dim, dtype=self.dtype, device=device),
operations.Linear(model_channels, time_embed_dim, dtype=self.dtype, device=device),
nn.SiLU(),
linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
)
if self.num_classes is not None:
@@ -394,9 +394,9 @@ class UNetModel(nn.Module):
assert adm_in_channels is not None
self.label_emb = nn.Sequential(
nn.Sequential(
linear(adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
operations.Linear(adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
nn.SiLU(),
linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
)
)
else:
@@ -405,7 +405,7 @@ class UNetModel(nn.Module):
self.input_blocks = nn.ModuleList(
[
TimestepEmbedSequential(
conv_nd(dims, in_channels, model_channels, 3, padding=1, dtype=self.dtype, device=device)
operations.conv_nd(dims, in_channels, model_channels, 3, padding=1, dtype=self.dtype, device=device)
)
]
)
@@ -426,6 +426,7 @@ class UNetModel(nn.Module):
use_scale_shift_norm=use_scale_shift_norm,
dtype=self.dtype,
device=device,
operations=operations,
)
]
ch = mult * model_channels
@@ -447,7 +448,7 @@ class UNetModel(nn.Module):
layers.append(SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth[level], context_dim=context_dim,
disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
use_checkpoint=use_checkpoint, dtype=self.dtype, device=device
use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
)
)
self.input_blocks.append(TimestepEmbedSequential(*layers))
@@ -468,10 +469,11 @@ class UNetModel(nn.Module):
down=True,
dtype=self.dtype,
device=device,
operations=operations
)
if resblock_updown
else Downsample(
ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device
ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations
)
)
)
@@ -498,11 +500,12 @@ class UNetModel(nn.Module):
use_scale_shift_norm=use_scale_shift_norm,
dtype=self.dtype,
device=device,
operations=operations
),
SpatialTransformer( # always uses a self-attn
ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim,
disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
use_checkpoint=use_checkpoint, dtype=self.dtype, device=device
use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
),
ResBlock(
ch,
@@ -513,6 +516,7 @@ class UNetModel(nn.Module):
use_scale_shift_norm=use_scale_shift_norm,
dtype=self.dtype,
device=device,
operations=operations
),
)
self._feature_size += ch
@@ -532,6 +536,7 @@ class UNetModel(nn.Module):
use_scale_shift_norm=use_scale_shift_norm,
dtype=self.dtype,
device=device,
operations=operations
)
]
ch = model_channels * mult
@@ -554,7 +559,7 @@ class UNetModel(nn.Module):
SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth[level], context_dim=context_dim,
disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
use_checkpoint=use_checkpoint, dtype=self.dtype, device=device
use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
)
)
if level and i == self.num_res_blocks[level]:
@@ -571,9 +576,10 @@ class UNetModel(nn.Module):
up=True,
dtype=self.dtype,
device=device,
operations=operations
)
if resblock_updown
else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device)
else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations)
)
ds //= 2
self.output_blocks.append(TimestepEmbedSequential(*layers))
@@ -582,12 +588,12 @@ class UNetModel(nn.Module):
self.out = nn.Sequential(
nn.GroupNorm(32, ch, dtype=self.dtype, device=device),
nn.SiLU(),
zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1, dtype=self.dtype, device=device)),
zero_module(operations.conv_nd(dims, model_channels, out_channels, 3, padding=1, dtype=self.dtype, device=device)),
)
if self.predict_codebook_ids:
self.id_predictor = nn.Sequential(
nn.GroupNorm(32, ch, dtype=self.dtype, device=device),
conv_nd(dims, model_channels, n_embed, 1, dtype=self.dtype, device=device),
operations.conv_nd(dims, model_channels, n_embed, 1, dtype=self.dtype, device=device),
#nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits
)