Spaces:
Running
on
Zero
Running
on
Zero
from modules.Device import Device | |
import torch | |
def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False): | |
"""#### Cast a weight tensor to a specified dtype and device. | |
#### Args: | |
- `weight` (torch.Tensor): The weight tensor. | |
- `dtype` (torch.dtype): The data type. | |
- `device` (torch.device): The device. | |
- `non_blocking` (bool): Whether to use non-blocking transfer. | |
- `copy` (bool): Whether to copy the tensor. | |
#### Returns: | |
- `torch.Tensor`: The casted weight tensor. | |
""" | |
if device is None or weight.device == device: | |
if not copy: | |
if dtype is None or weight.dtype == dtype: | |
return weight | |
return weight.to(dtype=dtype, copy=copy) | |
r = torch.empty_like(weight, dtype=dtype, device=device) | |
r.copy_(weight, non_blocking=non_blocking) | |
return r | |
def cast_to_input(weight, input, non_blocking=False, copy=True): | |
"""#### Cast a weight tensor to match the input tensor. | |
#### Args: | |
- `weight` (torch.Tensor): The weight tensor. | |
- `input` (torch.Tensor): The input tensor. | |
- `non_blocking` (bool): Whether to use non-blocking transfer. | |
- `copy` (bool): Whether to copy the tensor. | |
#### Returns: | |
- `torch.Tensor`: The casted weight tensor. | |
""" | |
return cast_to( | |
weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy | |
) | |
def cast_bias_weight(s: torch.nn.Module, input: torch.Tensor= None, dtype:torch.dtype = None, device:torch.device = None, bias_dtype:torch.dtype = None) -> tuple: | |
"""#### Cast the bias and weight of a module to match the input tensor. | |
#### Args: | |
- `s` (torch.nn.Module): The module. | |
- `input` (torch.Tensor): The input tensor. | |
#### Returns: | |
- `tuple`: The cast weight and bias. | |
""" | |
if input is not None: | |
if dtype is None: | |
dtype = input.dtype | |
if bias_dtype is None: | |
bias_dtype = dtype | |
if device is None: | |
device = input.device | |
bias = None | |
non_blocking = Device.device_supports_non_blocking(device) | |
if s.bias is not None: | |
has_function = s.bias_function is not None | |
bias = cast_to( | |
s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function | |
) | |
if has_function: | |
bias = s.bias_function(bias) | |
has_function = s.weight_function is not None | |
weight = cast_to( | |
s.weight, dtype, device, non_blocking=non_blocking, copy=has_function | |
) | |
if has_function: | |
weight = s.weight_function(weight) | |
return weight, bias | |
class CastWeightBiasOp: | |
"""#### Class representing a cast weight and bias operation.""" | |
comfy_cast_weights: bool = False | |
weight_function: callable = None | |
bias_function: callable = None | |
class disable_weight_init: | |
"""#### Class representing a module with disabled weight initialization.""" | |
class Linear(torch.nn.Linear, CastWeightBiasOp): | |
"""#### Linear layer with disabled weight initialization.""" | |
def reset_parameters(self): | |
"""#### Reset the parameters of the Linear layer.""" | |
return None | |
def forward_comfy_cast_weights(self, input): | |
"""#### Forward pass with comfy cast weights. | |
#### Args: | |
- `input` (torch.Tensor): The input tensor. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
weight, bias = cast_bias_weight(self, input) | |
return torch.nn.functional.linear(input, weight, bias) | |
def forward(self, *args, **kwargs): | |
"""#### Forward pass for the Linear layer. | |
#### Args: | |
- `*args`: Variable length argument list. | |
- `**kwargs`: Arbitrary keyword arguments. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
if self.comfy_cast_weights: | |
return self.forward_comfy_cast_weights(*args, **kwargs) | |
else: | |
return super().forward(*args, **kwargs) | |
class Conv1d(torch.nn.Conv1d, CastWeightBiasOp): | |
"""#### Conv1d layer with disabled weight initialization.""" | |
def reset_parameters(self): | |
"""#### Reset the parameters of the Conv1d layer.""" | |
return None | |
def forward_comfy_cast_weights(self, input): | |
"""#### Forward pass with comfy cast weights. | |
#### Args: | |
- `input` (torch.Tensor): The input tensor. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
weight, bias = cast_bias_weight(self, input) | |
return self._conv_forward(input, weight, bias) | |
def forward(self, *args, **kwargs): | |
"""#### Forward pass for the Conv1d layer. | |
#### Args: | |
- `*args`: Variable length argument list. | |
- `**kwargs`: Arbitrary keyword arguments. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
if self.comfy_cast_weights: | |
return self.forward_comfy_cast_weights(*args, **kwargs) | |
else: | |
return super().forward(*args, **kwargs) | |
class Conv2d(torch.nn.Conv2d, CastWeightBiasOp): | |
"""#### Conv2d layer with disabled weight initialization.""" | |
def reset_parameters(self) -> None: | |
"""#### Reset the parameters of the Conv2d layer.""" | |
return None | |
def forward_cast_weights(self, input: torch.Tensor) -> torch.Tensor: | |
"""#### Forward pass with comfy cast weights. | |
#### Args: | |
- `input` (torch.Tensor): The input tensor. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
weight, bias = cast_bias_weight(self, input) | |
return self._conv_forward(input, weight, bias) | |
def forward(self, *args, **kwargs) -> torch.Tensor: | |
"""#### Forward pass for the Conv2d layer. | |
#### Args: | |
- `*args`: Variable length argument list. | |
- `**kwargs`: Arbitrary keyword arguments. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
if self.comfy_cast_weights: | |
return self.forward_cast_weights(*args, **kwargs) | |
else: | |
return super().forward(*args, **kwargs) | |
class Conv3d(torch.nn.Conv3d, CastWeightBiasOp): | |
"""#### Conv3d layer with disabled weight initialization.""" | |
def reset_parameters(self): | |
"""#### Reset the parameters of the Conv3d layer.""" | |
return None | |
def forward_comfy_cast_weights(self, input): | |
"""#### Forward pass with comfy cast weights. | |
#### Args: | |
- `input` (torch.Tensor): The input tensor. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
weight, bias = cast_bias_weight(self, input) | |
return self._conv_forward(input, weight, bias) | |
def forward(self, *args, **kwargs): | |
"""#### Forward pass for the Conv3d layer. | |
#### Args: | |
- `*args`: Variable length argument list. | |
- `**kwargs`: Arbitrary keyword arguments. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
if self.comfy_cast_weights: | |
return self.forward_comfy_cast_weights(*args, **kwargs) | |
else: | |
return super().forward(*args, **kwargs) | |
class GroupNorm(torch.nn.GroupNorm, CastWeightBiasOp): | |
"""#### GroupNorm layer with disabled weight initialization.""" | |
def reset_parameters(self) -> None: | |
"""#### Reset the parameters of the GroupNorm layer.""" | |
return None | |
def forward_comfy_cast_weights(self, input): | |
"""#### Forward pass with comfy cast weights. | |
#### Args: | |
- `input` (torch.Tensor): The input tensor. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
weight, bias = cast_bias_weight(self, input) | |
return torch.nn.functional.group_norm( | |
input, self.num_groups, weight, bias, self.eps | |
) | |
def forward(self, *args, **kwargs): | |
"""#### Forward pass for the GroupNorm layer. | |
#### Args: | |
- `*args`: Variable length argument list. | |
- `**kwargs`: Arbitrary keyword arguments. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
if self.comfy_cast_weights: | |
return self.forward_comfy_cast_weights(*args, **kwargs) | |
else: | |
return super().forward(*args, **kwargs) | |
class LayerNorm(torch.nn.LayerNorm, CastWeightBiasOp): | |
"""#### LayerNorm layer with disabled weight initialization.""" | |
def reset_parameters(self) -> None: | |
"""#### Reset the parameters of the LayerNorm layer.""" | |
return None | |
def forward_cast_weights(self, input: torch.Tensor) -> torch.Tensor: | |
"""#### Forward pass with cast weights. | |
#### Args: | |
- `input` (torch.Tensor): The input tensor. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
if self.weight is not None: | |
weight, bias = cast_bias_weight(self, input) | |
else: | |
weight = None | |
bias = None | |
return torch.nn.functional.layer_norm( | |
input, self.normalized_shape, weight, bias, self.eps | |
) | |
def forward(self, *args, **kwargs) -> torch.Tensor: | |
"""#### Forward pass for the LayerNorm layer. | |
#### Args: | |
- `*args`: Variable length argument list. | |
- `**kwargs`: Arbitrary keyword arguments. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
if self.comfy_cast_weights: | |
return self.forward_cast_weights(*args, **kwargs) | |
else: | |
return super().forward(*args, **kwargs) | |
class ConvTranspose2d(torch.nn.ConvTranspose2d, CastWeightBiasOp): | |
"""#### ConvTranspose2d layer with disabled weight initialization.""" | |
def reset_parameters(self): | |
"""#### Reset the parameters of the ConvTranspose2d layer.""" | |
return None | |
def forward_comfy_cast_weights(self, input, output_size=None): | |
"""#### Forward pass with comfy cast weights. | |
#### Args: | |
- `input` (torch.Tensor): The input tensor. | |
- `output_size` (torch.Size): The output size. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
num_spatial_dims = 2 | |
output_padding = self._output_padding( | |
input, | |
output_size, | |
self.stride, | |
self.padding, | |
self.kernel_size, | |
num_spatial_dims, | |
self.dilation, | |
) | |
weight, bias = cast_bias_weight(self, input) | |
return torch.nn.functional.conv_transpose2d( | |
input, | |
weight, | |
bias, | |
self.stride, | |
self.padding, | |
output_padding, | |
self.groups, | |
self.dilation, | |
) | |
def forward(self, *args, **kwargs): | |
"""#### Forward pass for the ConvTranspose2d layer. | |
#### Args: | |
- `*args`: Variable length argument list. | |
- `**kwargs`: Arbitrary keyword arguments. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
if self.comfy_cast_weights: | |
return self.forward_comfy_cast_weights(*args, **kwargs) | |
else: | |
return super().forward(*args, **kwargs) | |
class ConvTranspose1d(torch.nn.ConvTranspose1d, CastWeightBiasOp): | |
"""#### ConvTranspose1d layer with disabled weight initialization.""" | |
def reset_parameters(self): | |
"""#### Reset the parameters of the ConvTranspose1d layer.""" | |
return None | |
def forward_comfy_cast_weights(self, input, output_size=None): | |
"""#### Forward pass with comfy cast weights. | |
#### Args: | |
- `input` (torch.Tensor): The input tensor. | |
- `output_size` (torch.Size): The output size. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
num_spatial_dims = 1 | |
output_padding = self._output_padding( | |
input, | |
output_size, | |
self.stride, | |
self.padding, | |
self.kernel_size, | |
num_spatial_dims, | |
self.dilation, | |
) | |
weight, bias = cast_bias_weight(self, input) | |
return torch.nn.functional.conv_transpose1d( | |
input, | |
weight, | |
bias, | |
self.stride, | |
self.padding, | |
output_padding, | |
self.groups, | |
self.dilation, | |
) | |
def forward(self, *args, **kwargs): | |
"""#### Forward pass for the ConvTranspose1d layer. | |
#### Args: | |
- `*args`: Variable length argument list. | |
- `**kwargs`: Arbitrary keyword arguments. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
if self.comfy_cast_weights: | |
return self.forward_comfy_cast_weights(*args, **kwargs) | |
else: | |
return super().forward(*args, **kwargs) | |
class Embedding(torch.nn.Embedding, CastWeightBiasOp): | |
"""#### Embedding layer with disabled weight initialization.""" | |
def reset_parameters(self): | |
"""#### Reset the parameters of the Embedding layer.""" | |
self.bias = None | |
return None | |
def forward_comfy_cast_weights(self, input, out_dtype=None): | |
"""#### Forward pass with comfy cast weights. | |
#### Args: | |
- `input` (torch.Tensor): The input tensor. | |
- `out_dtype` (torch.dtype): The output data type. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
output_dtype = out_dtype | |
if ( | |
self.weight.dtype == torch.float16 | |
or self.weight.dtype == torch.bfloat16 | |
): | |
out_dtype = None | |
weight, bias = cast_bias_weight(self, device=input.device, dtype=out_dtype) | |
return torch.nn.functional.embedding( | |
input, | |
weight, | |
self.padding_idx, | |
self.max_norm, | |
self.norm_type, | |
self.scale_grad_by_freq, | |
self.sparse, | |
).to(dtype=output_dtype) | |
def forward(self, *args, **kwargs): | |
"""#### Forward pass for the Embedding layer. | |
#### Args: | |
- `*args`: Variable length argument list. | |
- `**kwargs`: Arbitrary keyword arguments. | |
#### Returns: | |
- `torch.Tensor`: The output tensor. | |
""" | |
if self.comfy_cast_weights: | |
return self.forward_comfy_cast_weights(*args, **kwargs) | |
else: | |
if "out_dtype" in kwargs: | |
kwargs.pop("out_dtype") | |
return super().forward(*args, **kwargs) | |
def conv_nd(s, dims: int, *args, **kwargs) -> torch.nn.Conv2d: | |
"""#### Create a Conv2d layer with the specified dimensions. | |
#### Args: | |
- `dims` (int): The number of dimensions. | |
- `*args`: Variable length argument list. | |
- `**kwargs`: Arbitrary keyword arguments. | |
#### Returns: | |
- `torch.nn.Conv2d`: The Conv2d layer. | |
""" | |
if dims == 2: | |
return s.Conv2d(*args, **kwargs) | |
elif dims == 3: | |
return s.Conv3d(*args, **kwargs) | |
else: | |
raise ValueError(f"unsupported dimensions: {dims}") | |
class manual_cast(disable_weight_init): | |
"""#### Class representing a module with manual casting.""" | |
class Linear(disable_weight_init.Linear): | |
"""#### Linear layer with manual casting.""" | |
comfy_cast_weights: bool = True | |
class Conv1d(disable_weight_init.Conv1d): | |
"""#### Conv1d layer with manual casting.""" | |
comfy_cast_weights = True | |
class Conv2d(disable_weight_init.Conv2d): | |
"""#### Conv2d layer with manual casting.""" | |
comfy_cast_weights: bool = True | |
class Conv3d(disable_weight_init.Conv3d): | |
"""#### Conv3d layer with manual casting.""" | |
comfy_cast_weights = True | |
class GroupNorm(disable_weight_init.GroupNorm): | |
"""#### GroupNorm layer with manual casting.""" | |
comfy_cast_weights: bool = True | |
class LayerNorm(disable_weight_init.LayerNorm): | |
"""#### LayerNorm layer with manual casting.""" | |
comfy_cast_weights: bool = True | |
class ConvTranspose2d(disable_weight_init.ConvTranspose2d): | |
"""#### ConvTranspose2d layer with manual casting.""" | |
comfy_cast_weights = True | |
class ConvTranspose1d(disable_weight_init.ConvTranspose1d): | |
"""#### ConvTranspose1d layer with manual casting.""" | |
comfy_cast_weights = True | |
class Embedding(disable_weight_init.Embedding): | |
"""#### Embedding layer with manual casting.""" | |
comfy_cast_weights = True | |