2022-12-10 14:14:30 +08:00
|
|
|
import torch
|
2023-01-25 12:51:45 +08:00
|
|
|
from packaging import version
|
|
|
|
|
|
|
|
from modules import devices
|
|
|
|
from modules.sd_hijack_utils import CondFunc
|
2022-12-10 14:14:30 +08:00
|
|
|
|
|
|
|
|
|
|
|
class TorchHijackForUnet:
|
|
|
|
"""
|
|
|
|
This is torch, but with cat that resizes tensors to appropriate dimensions if they do not match;
|
2022-12-15 10:01:32 +08:00
|
|
|
this makes it possible to create pictures with dimensions that are multiples of 8 rather than 64
|
2022-12-10 14:14:30 +08:00
|
|
|
"""
|
|
|
|
|
|
|
|
def __getattr__(self, item):
|
|
|
|
if item == 'cat':
|
|
|
|
return self.cat
|
|
|
|
|
|
|
|
if hasattr(torch, item):
|
|
|
|
return getattr(torch, item)
|
|
|
|
|
|
|
|
raise AttributeError("'{}' object has no attribute '{}'".format(type(self).__name__, item))
|
|
|
|
|
|
|
|
def cat(self, tensors, *args, **kwargs):
|
|
|
|
if len(tensors) == 2:
|
|
|
|
a, b = tensors
|
|
|
|
if a.shape[-2:] != b.shape[-2:]:
|
|
|
|
a = torch.nn.functional.interpolate(a, b.shape[-2:], mode="nearest")
|
|
|
|
|
|
|
|
tensors = (a, b)
|
|
|
|
|
|
|
|
return torch.cat(tensors, *args, **kwargs)
|
|
|
|
|
|
|
|
|
|
|
|
th = TorchHijackForUnet()
|
2023-01-25 12:51:45 +08:00
|
|
|
|
|
|
|
|
|
|
|
# Below are monkey patches to enable upcasting a float16 UNet for float32 sampling
|
|
|
|
def apply_model(orig_func, self, x_noisy, t, cond, **kwargs):
|
2023-01-26 01:11:01 +08:00
|
|
|
|
|
|
|
if isinstance(cond, dict):
|
|
|
|
for y in cond.keys():
|
|
|
|
cond[y] = [x.to(devices.dtype_unet) if isinstance(x, torch.Tensor) else x for x in cond[y]]
|
|
|
|
|
2023-01-25 12:51:45 +08:00
|
|
|
with devices.autocast():
|
|
|
|
return orig_func(self, x_noisy.to(devices.dtype_unet), t.to(devices.dtype_unet), cond, **kwargs).float()
|
|
|
|
|
2023-02-07 13:05:54 +08:00
|
|
|
|
2023-01-25 12:51:45 +08:00
|
|
|
class GELUHijack(torch.nn.GELU, torch.nn.Module):
|
|
|
|
def __init__(self, *args, **kwargs):
|
|
|
|
torch.nn.GELU.__init__(self, *args, **kwargs)
|
|
|
|
def forward(self, x):
|
|
|
|
if devices.unet_needs_upcast:
|
|
|
|
return torch.nn.GELU.forward(self.float(), x.float()).to(devices.dtype_unet)
|
|
|
|
else:
|
|
|
|
return torch.nn.GELU.forward(self, x)
|
|
|
|
|
2023-02-07 13:05:54 +08:00
|
|
|
|
|
|
|
ddpm_edit_hijack = None
|
|
|
|
def hijack_ddpm_edit():
|
|
|
|
global ddpm_edit_hijack
|
|
|
|
if not ddpm_edit_hijack:
|
|
|
|
CondFunc('modules.models.diffusion.ddpm_edit.LatentDiffusion.decode_first_stage', first_stage_sub, first_stage_cond)
|
|
|
|
CondFunc('modules.models.diffusion.ddpm_edit.LatentDiffusion.encode_first_stage', first_stage_sub, first_stage_cond)
|
|
|
|
ddpm_edit_hijack = CondFunc('modules.models.diffusion.ddpm_edit.LatentDiffusion.apply_model', apply_model, unet_needs_upcast)
|
|
|
|
|
|
|
|
|
2023-01-25 12:51:45 +08:00
|
|
|
unet_needs_upcast = lambda *args, **kwargs: devices.unet_needs_upcast
|
|
|
|
CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.apply_model', apply_model, unet_needs_upcast)
|
2023-01-27 23:19:43 +08:00
|
|
|
CondFunc('ldm.modules.diffusionmodules.openaimodel.timestep_embedding', lambda orig_func, timesteps, *args, **kwargs: orig_func(timesteps, *args, **kwargs).to(torch.float32 if timesteps.dtype == torch.int64 else devices.dtype_unet), unet_needs_upcast)
|
2023-03-24 20:25:42 +08:00
|
|
|
if version.parse(torch.__version__) <= version.parse("1.13.2") or torch.cuda.is_available():
|
2023-01-25 12:51:45 +08:00
|
|
|
CondFunc('ldm.modules.diffusionmodules.util.GroupNorm32.forward', lambda orig_func, self, *args, **kwargs: orig_func(self.float(), *args, **kwargs), unet_needs_upcast)
|
|
|
|
CondFunc('ldm.modules.attention.GEGLU.forward', lambda orig_func, self, x: orig_func(self.float(), x.float()).to(devices.dtype_unet), unet_needs_upcast)
|
|
|
|
CondFunc('open_clip.transformer.ResidualAttentionBlock.__init__', lambda orig_func, *args, **kwargs: kwargs.update({'act_layer': GELUHijack}) and False or orig_func(*args, **kwargs), lambda _, *args, **kwargs: kwargs.get('act_layer') is None or kwargs['act_layer'] == torch.nn.GELU)
|
2023-01-27 23:19:43 +08:00
|
|
|
|
|
|
|
first_stage_cond = lambda _, self, *args, **kwargs: devices.unet_needs_upcast and self.model.diffusion_model.dtype == torch.float16
|
|
|
|
first_stage_sub = lambda orig_func, self, x, **kwargs: orig_func(self, x.to(devices.dtype_vae), **kwargs)
|
|
|
|
CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.decode_first_stage', first_stage_sub, first_stage_cond)
|
|
|
|
CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.encode_first_stage', first_stage_sub, first_stage_cond)
|
|
|
|
CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.get_first_stage_encoding', lambda orig_func, *args, **kwargs: orig_func(*args, **kwargs).float(), first_stage_cond)
|