stable-diffusion-webui

1
import torch
2
from packaging import version
3

4
from modules import devices
5
from modules.sd_hijack_utils import CondFunc
6

7

8
class TorchHijackForUnet:
9
    """
10
    This is torch, but with cat that resizes tensors to appropriate dimensions if they do not match;
11
    this makes it possible to create pictures with dimensions that are multiples of 8 rather than 64
12
    """
13

14
    def __getattr__(self, item):
15
        if item == 'cat':
16
            return self.cat
17

18
        if hasattr(torch, item):
19
            return getattr(torch, item)
20

21
        raise AttributeError(f"'{type(self).__name__}' object has no attribute '{item}'")
22

23
    def cat(self, tensors, *args, **kwargs):
24
        if len(tensors) == 2:
25
            a, b = tensors
26
            if a.shape[-2:] != b.shape[-2:]:
27
                a = torch.nn.functional.interpolate(a, b.shape[-2:], mode="nearest")
28

29
            tensors = (a, b)
30

31
        return torch.cat(tensors, *args, **kwargs)
32

33

34
th = TorchHijackForUnet()
35

36

37
# Below are monkey patches to enable upcasting a float16 UNet for float32 sampling
38
def apply_model(orig_func, self, x_noisy, t, cond, **kwargs):
39

40
    if isinstance(cond, dict):
41
        for y in cond.keys():
42
            if isinstance(cond[y], list):
43
                cond[y] = [x.to(devices.dtype_unet) if isinstance(x, torch.Tensor) else x for x in cond[y]]
44
            else:
45
                cond[y] = cond[y].to(devices.dtype_unet) if isinstance(cond[y], torch.Tensor) else cond[y]
46

47
    with devices.autocast():
48
        return orig_func(self, x_noisy.to(devices.dtype_unet), t.to(devices.dtype_unet), cond, **kwargs).float()
49

50

51
class GELUHijack(torch.nn.GELU, torch.nn.Module):
52
    def __init__(self, *args, **kwargs):
53
        torch.nn.GELU.__init__(self, *args, **kwargs)
54
    def forward(self, x):
55
        if devices.unet_needs_upcast:
56
            return torch.nn.GELU.forward(self.float(), x.float()).to(devices.dtype_unet)
57
        else:
58
            return torch.nn.GELU.forward(self, x)
59

60

61
ddpm_edit_hijack = None
62
def hijack_ddpm_edit():
63
    global ddpm_edit_hijack
64
    if not ddpm_edit_hijack:
65
        CondFunc('modules.models.diffusion.ddpm_edit.LatentDiffusion.decode_first_stage', first_stage_sub, first_stage_cond)
66
        CondFunc('modules.models.diffusion.ddpm_edit.LatentDiffusion.encode_first_stage', first_stage_sub, first_stage_cond)
67
        ddpm_edit_hijack = CondFunc('modules.models.diffusion.ddpm_edit.LatentDiffusion.apply_model', apply_model, unet_needs_upcast)
68

69

70
unet_needs_upcast = lambda *args, **kwargs: devices.unet_needs_upcast
71
CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.apply_model', apply_model, unet_needs_upcast)
72
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)
73
if version.parse(torch.__version__) <= version.parse("1.13.2") or torch.cuda.is_available():
74
    CondFunc('ldm.modules.diffusionmodules.util.GroupNorm32.forward', lambda orig_func, self, *args, **kwargs: orig_func(self.float(), *args, **kwargs), unet_needs_upcast)
75
    CondFunc('ldm.modules.attention.GEGLU.forward', lambda orig_func, self, x: orig_func(self.float(), x.float()).to(devices.dtype_unet), unet_needs_upcast)
76
    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)
77

78
first_stage_cond = lambda _, self, *args, **kwargs: devices.unet_needs_upcast and self.model.diffusion_model.dtype == torch.float16
79
first_stage_sub = lambda orig_func, self, x, **kwargs: orig_func(self, x.to(devices.dtype_vae), **kwargs)
80
CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.decode_first_stage', first_stage_sub, first_stage_cond)
81
CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.encode_first_stage', first_stage_sub, first_stage_cond)
82
CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.get_first_stage_encoding', lambda orig_func, *args, **kwargs: orig_func(*args, **kwargs).float(), first_stage_cond)
83

84
CondFunc('sgm.modules.diffusionmodules.wrappers.OpenAIWrapper.forward', apply_model, unet_needs_upcast)
85
CondFunc('sgm.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)
86