2024-02-16 17:56:11 +00:00
|
|
|
"""
|
|
|
|
This file is part of ComfyUI.
|
|
|
|
Copyright (C) 2024 Stability AI
|
|
|
|
|
|
|
|
This program is free software: you can redistribute it and/or modify
|
|
|
|
it under the terms of the GNU General Public License as published by
|
|
|
|
the Free Software Foundation, either version 3 of the License, or
|
|
|
|
(at your option) any later version.
|
|
|
|
|
|
|
|
This program is distributed in the hope that it will be useful,
|
|
|
|
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
|
|
GNU General Public License for more details.
|
|
|
|
|
|
|
|
You should have received a copy of the GNU General Public License
|
|
|
|
along with this program. If not, see <https://www.gnu.org/licenses/>.
|
|
|
|
"""
|
|
|
|
|
2023-06-14 15:17:59 +00:00
|
|
|
import torch
|
2023-12-22 19:24:04 +00:00
|
|
|
import comfy.model_management
|
2024-08-20 15:49:33 +00:00
|
|
|
from comfy.cli_args import args
|
2024-10-20 03:47:42 +00:00
|
|
|
import comfy.float
|
2024-07-31 04:52:34 +00:00
|
|
|
|
2024-10-17 21:25:56 +00:00
|
|
|
cast_to = comfy.model_management.cast_to #TODO: remove once no more references
|
2024-07-31 04:52:34 +00:00
|
|
|
|
2024-08-22 19:15:47 +00:00
|
|
|
def cast_to_input(weight, input, non_blocking=False, copy=True):
|
2024-10-17 21:25:56 +00:00
|
|
|
return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
|
2024-07-31 04:52:34 +00:00
|
|
|
|
2024-08-22 19:15:47 +00:00
|
|
|
def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
|
2024-07-31 04:52:34 +00:00
|
|
|
if input is not None:
|
|
|
|
if dtype is None:
|
|
|
|
dtype = input.dtype
|
2024-08-22 19:15:47 +00:00
|
|
|
if bias_dtype is None:
|
|
|
|
bias_dtype = dtype
|
2024-07-31 04:52:34 +00:00
|
|
|
if device is None:
|
|
|
|
device = input.device
|
2024-07-30 09:03:20 +00:00
|
|
|
|
2023-12-22 19:24:04 +00:00
|
|
|
bias = None
|
2024-08-18 14:29:33 +00:00
|
|
|
non_blocking = comfy.model_management.device_supports_non_blocking(device)
|
2023-12-22 19:24:04 +00:00
|
|
|
if s.bias is not None:
|
2024-08-22 19:15:47 +00:00
|
|
|
has_function = s.bias_function is not None
|
2024-10-17 21:25:56 +00:00
|
|
|
bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function)
|
2024-08-22 19:15:47 +00:00
|
|
|
if has_function:
|
2024-03-13 23:04:41 +00:00
|
|
|
bias = s.bias_function(bias)
|
2024-08-22 19:15:47 +00:00
|
|
|
|
|
|
|
has_function = s.weight_function is not None
|
2024-10-17 21:25:56 +00:00
|
|
|
weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function)
|
2024-08-22 19:15:47 +00:00
|
|
|
if has_function:
|
2024-03-13 23:04:41 +00:00
|
|
|
weight = s.weight_function(weight)
|
2023-12-22 19:24:04 +00:00
|
|
|
return weight, bias
|
|
|
|
|
2024-03-14 13:30:21 +00:00
|
|
|
class CastWeightBiasOp:
|
|
|
|
comfy_cast_weights = False
|
|
|
|
weight_function = None
|
|
|
|
bias_function = None
|
2023-06-14 15:17:59 +00:00
|
|
|
|
2023-12-12 04:27:13 +00:00
|
|
|
class disable_weight_init:
|
2024-03-14 13:30:21 +00:00
|
|
|
class Linear(torch.nn.Linear, CastWeightBiasOp):
|
2023-12-12 04:27:13 +00:00
|
|
|
def reset_parameters(self):
|
|
|
|
return None
|
2023-06-14 23:46:08 +00:00
|
|
|
|
2023-12-22 19:24:04 +00:00
|
|
|
def forward_comfy_cast_weights(self, input):
|
|
|
|
weight, bias = cast_bias_weight(self, input)
|
|
|
|
return torch.nn.functional.linear(input, weight, bias)
|
|
|
|
|
|
|
|
def forward(self, *args, **kwargs):
|
|
|
|
if self.comfy_cast_weights:
|
|
|
|
return self.forward_comfy_cast_weights(*args, **kwargs)
|
|
|
|
else:
|
|
|
|
return super().forward(*args, **kwargs)
|
|
|
|
|
2024-06-15 16:14:56 +00:00
|
|
|
class Conv1d(torch.nn.Conv1d, CastWeightBiasOp):
|
|
|
|
def reset_parameters(self):
|
|
|
|
return None
|
|
|
|
|
|
|
|
def forward_comfy_cast_weights(self, input):
|
|
|
|
weight, bias = cast_bias_weight(self, input)
|
|
|
|
return self._conv_forward(input, weight, bias)
|
|
|
|
|
|
|
|
def forward(self, *args, **kwargs):
|
|
|
|
if self.comfy_cast_weights:
|
|
|
|
return self.forward_comfy_cast_weights(*args, **kwargs)
|
|
|
|
else:
|
|
|
|
return super().forward(*args, **kwargs)
|
|
|
|
|
2024-03-14 13:30:21 +00:00
|
|
|
class Conv2d(torch.nn.Conv2d, CastWeightBiasOp):
|
2023-12-12 04:27:13 +00:00
|
|
|
def reset_parameters(self):
|
|
|
|
return None
|
2023-06-15 00:13:08 +00:00
|
|
|
|
2023-12-22 19:24:04 +00:00
|
|
|
def forward_comfy_cast_weights(self, input):
|
|
|
|
weight, bias = cast_bias_weight(self, input)
|
|
|
|
return self._conv_forward(input, weight, bias)
|
|
|
|
|
|
|
|
def forward(self, *args, **kwargs):
|
|
|
|
if self.comfy_cast_weights:
|
|
|
|
return self.forward_comfy_cast_weights(*args, **kwargs)
|
|
|
|
else:
|
|
|
|
return super().forward(*args, **kwargs)
|
|
|
|
|
2024-03-14 13:30:21 +00:00
|
|
|
class Conv3d(torch.nn.Conv3d, CastWeightBiasOp):
|
2023-12-12 04:27:13 +00:00
|
|
|
def reset_parameters(self):
|
|
|
|
return None
|
2023-11-11 06:00:43 +00:00
|
|
|
|
2023-12-22 19:24:04 +00:00
|
|
|
def forward_comfy_cast_weights(self, input):
|
|
|
|
weight, bias = cast_bias_weight(self, input)
|
|
|
|
return self._conv_forward(input, weight, bias)
|
|
|
|
|
|
|
|
def forward(self, *args, **kwargs):
|
|
|
|
if self.comfy_cast_weights:
|
|
|
|
return self.forward_comfy_cast_weights(*args, **kwargs)
|
|
|
|
else:
|
|
|
|
return super().forward(*args, **kwargs)
|
|
|
|
|
2024-03-14 13:30:21 +00:00
|
|
|
class GroupNorm(torch.nn.GroupNorm, CastWeightBiasOp):
|
2023-12-12 04:27:13 +00:00
|
|
|
def reset_parameters(self):
|
|
|
|
return None
|
2023-12-04 08:12:18 +00:00
|
|
|
|
2023-12-22 19:24:04 +00:00
|
|
|
def forward_comfy_cast_weights(self, input):
|
|
|
|
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):
|
|
|
|
if self.comfy_cast_weights:
|
|
|
|
return self.forward_comfy_cast_weights(*args, **kwargs)
|
|
|
|
else:
|
|
|
|
return super().forward(*args, **kwargs)
|
|
|
|
|
|
|
|
|
2024-03-14 13:30:21 +00:00
|
|
|
class LayerNorm(torch.nn.LayerNorm, CastWeightBiasOp):
|
2023-12-12 04:27:13 +00:00
|
|
|
def reset_parameters(self):
|
|
|
|
return None
|
2023-12-04 08:12:18 +00:00
|
|
|
|
2023-12-22 19:24:04 +00:00
|
|
|
def forward_comfy_cast_weights(self, input):
|
2024-02-16 17:56:11 +00:00
|
|
|
if self.weight is not None:
|
|
|
|
weight, bias = cast_bias_weight(self, input)
|
|
|
|
else:
|
|
|
|
weight = None
|
|
|
|
bias = None
|
2023-12-22 19:24:04 +00:00
|
|
|
return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
|
|
|
|
|
|
|
|
def forward(self, *args, **kwargs):
|
|
|
|
if self.comfy_cast_weights:
|
|
|
|
return self.forward_comfy_cast_weights(*args, **kwargs)
|
|
|
|
else:
|
|
|
|
return super().forward(*args, **kwargs)
|
|
|
|
|
2024-03-14 13:30:21 +00:00
|
|
|
class ConvTranspose2d(torch.nn.ConvTranspose2d, CastWeightBiasOp):
|
2024-02-16 17:56:11 +00:00
|
|
|
def reset_parameters(self):
|
|
|
|
return None
|
|
|
|
|
|
|
|
def forward_comfy_cast_weights(self, input, output_size=None):
|
|
|
|
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):
|
|
|
|
if self.comfy_cast_weights:
|
|
|
|
return self.forward_comfy_cast_weights(*args, **kwargs)
|
|
|
|
else:
|
|
|
|
return super().forward(*args, **kwargs)
|
|
|
|
|
2024-06-15 16:14:56 +00:00
|
|
|
class ConvTranspose1d(torch.nn.ConvTranspose1d, CastWeightBiasOp):
|
|
|
|
def reset_parameters(self):
|
|
|
|
return None
|
|
|
|
|
|
|
|
def forward_comfy_cast_weights(self, input, output_size=None):
|
|
|
|
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):
|
|
|
|
if self.comfy_cast_weights:
|
|
|
|
return self.forward_comfy_cast_weights(*args, **kwargs)
|
|
|
|
else:
|
|
|
|
return super().forward(*args, **kwargs)
|
|
|
|
|
2024-07-30 09:03:20 +00:00
|
|
|
class Embedding(torch.nn.Embedding, CastWeightBiasOp):
|
|
|
|
def reset_parameters(self):
|
|
|
|
self.bias = None
|
|
|
|
return None
|
|
|
|
|
2024-07-31 04:52:34 +00:00
|
|
|
def forward_comfy_cast_weights(self, input, out_dtype=None):
|
|
|
|
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)
|
2024-07-30 09:03:20 +00:00
|
|
|
|
|
|
|
def forward(self, *args, **kwargs):
|
|
|
|
if self.comfy_cast_weights:
|
|
|
|
return self.forward_comfy_cast_weights(*args, **kwargs)
|
|
|
|
else:
|
2024-07-31 04:52:34 +00:00
|
|
|
if "out_dtype" in kwargs:
|
|
|
|
kwargs.pop("out_dtype")
|
2024-07-30 09:03:20 +00:00
|
|
|
return super().forward(*args, **kwargs)
|
|
|
|
|
2023-12-12 04:27:13 +00:00
|
|
|
@classmethod
|
|
|
|
def conv_nd(s, dims, *args, **kwargs):
|
|
|
|
if dims == 2:
|
|
|
|
return s.Conv2d(*args, **kwargs)
|
|
|
|
elif dims == 3:
|
|
|
|
return s.Conv3d(*args, **kwargs)
|
|
|
|
else:
|
|
|
|
raise ValueError(f"unsupported dimensions: {dims}")
|
2023-06-15 00:13:08 +00:00
|
|
|
|
2023-12-11 04:00:54 +00:00
|
|
|
|
2023-12-12 04:27:13 +00:00
|
|
|
class manual_cast(disable_weight_init):
|
|
|
|
class Linear(disable_weight_init.Linear):
|
2023-12-22 19:24:04 +00:00
|
|
|
comfy_cast_weights = True
|
2023-12-11 04:00:54 +00:00
|
|
|
|
2024-06-15 16:14:56 +00:00
|
|
|
class Conv1d(disable_weight_init.Conv1d):
|
|
|
|
comfy_cast_weights = True
|
|
|
|
|
2023-12-12 04:27:13 +00:00
|
|
|
class Conv2d(disable_weight_init.Conv2d):
|
2023-12-22 19:24:04 +00:00
|
|
|
comfy_cast_weights = True
|
2023-12-11 04:00:54 +00:00
|
|
|
|
2023-12-12 04:27:13 +00:00
|
|
|
class Conv3d(disable_weight_init.Conv3d):
|
2023-12-22 19:24:04 +00:00
|
|
|
comfy_cast_weights = True
|
2023-12-11 04:00:54 +00:00
|
|
|
|
2023-12-12 04:27:13 +00:00
|
|
|
class GroupNorm(disable_weight_init.GroupNorm):
|
2023-12-22 19:24:04 +00:00
|
|
|
comfy_cast_weights = True
|
2023-12-11 04:00:54 +00:00
|
|
|
|
2023-12-12 04:27:13 +00:00
|
|
|
class LayerNorm(disable_weight_init.LayerNorm):
|
2023-12-22 19:24:04 +00:00
|
|
|
comfy_cast_weights = True
|
2024-02-16 17:56:11 +00:00
|
|
|
|
|
|
|
class ConvTranspose2d(disable_weight_init.ConvTranspose2d):
|
|
|
|
comfy_cast_weights = True
|
2024-06-15 16:14:56 +00:00
|
|
|
|
|
|
|
class ConvTranspose1d(disable_weight_init.ConvTranspose1d):
|
|
|
|
comfy_cast_weights = True
|
2024-07-30 09:03:20 +00:00
|
|
|
|
|
|
|
class Embedding(disable_weight_init.Embedding):
|
|
|
|
comfy_cast_weights = True
|
2024-08-20 15:49:33 +00:00
|
|
|
|
|
|
|
|
|
|
|
def fp8_linear(self, input):
|
|
|
|
dtype = self.weight.dtype
|
|
|
|
if dtype not in [torch.float8_e4m3fn]:
|
|
|
|
return None
|
|
|
|
|
|
|
|
if len(input.shape) == 3:
|
2024-08-22 19:15:47 +00:00
|
|
|
w, bias = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input.dtype)
|
|
|
|
w = w.t()
|
2024-08-21 02:53:26 +00:00
|
|
|
|
2024-08-21 18:01:41 +00:00
|
|
|
scale_weight = self.scale_weight
|
|
|
|
scale_input = self.scale_input
|
|
|
|
if scale_weight is None:
|
2024-10-20 03:47:42 +00:00
|
|
|
scale_weight = torch.ones((), device=input.device, dtype=torch.float32)
|
2024-08-21 18:01:41 +00:00
|
|
|
if scale_input is None:
|
2024-10-20 03:47:42 +00:00
|
|
|
scale_input = torch.ones((), device=input.device, dtype=torch.float32)
|
|
|
|
inn = input.reshape(-1, input.shape[2]).to(dtype)
|
|
|
|
else:
|
|
|
|
inn = (input * (1.0 / scale_input).to(input.dtype)).reshape(-1, input.shape[2]).to(dtype)
|
2024-08-21 18:01:41 +00:00
|
|
|
|
2024-08-22 19:15:47 +00:00
|
|
|
if bias is not None:
|
|
|
|
o = torch._scaled_mm(inn, w, out_dtype=input.dtype, bias=bias, scale_a=scale_input, scale_b=scale_weight)
|
2024-08-21 02:53:26 +00:00
|
|
|
else:
|
2024-08-21 18:01:41 +00:00
|
|
|
o = torch._scaled_mm(inn, w, out_dtype=input.dtype, scale_a=scale_input, scale_b=scale_weight)
|
|
|
|
|
|
|
|
if isinstance(o, tuple):
|
|
|
|
o = o[0]
|
2024-08-21 02:53:26 +00:00
|
|
|
|
2024-08-21 15:21:48 +00:00
|
|
|
return o.reshape((-1, input.shape[1], self.weight.shape[0]))
|
2024-08-20 15:49:33 +00:00
|
|
|
return None
|
|
|
|
|
|
|
|
class fp8_ops(manual_cast):
|
|
|
|
class Linear(manual_cast.Linear):
|
2024-08-21 18:01:41 +00:00
|
|
|
def reset_parameters(self):
|
|
|
|
self.scale_weight = None
|
|
|
|
self.scale_input = None
|
|
|
|
return None
|
|
|
|
|
2024-08-20 15:49:33 +00:00
|
|
|
def forward_comfy_cast_weights(self, input):
|
|
|
|
out = fp8_linear(self, input)
|
|
|
|
if out is not None:
|
|
|
|
return out
|
|
|
|
|
|
|
|
weight, bias = cast_bias_weight(self, input)
|
|
|
|
return torch.nn.functional.linear(input, weight, bias)
|
|
|
|
|
2024-10-20 03:47:42 +00:00
|
|
|
def scaled_fp8_ops(fp8_matrix_mult=False):
|
|
|
|
class scaled_fp8_op(manual_cast):
|
|
|
|
class Linear(manual_cast.Linear):
|
|
|
|
def reset_parameters(self):
|
|
|
|
if not hasattr(self, 'scale_weight'):
|
|
|
|
self.scale_weight = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False)
|
|
|
|
if not hasattr(self, 'scale_input'):
|
|
|
|
self.scale_input = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False)
|
|
|
|
return None
|
|
|
|
|
|
|
|
def forward_comfy_cast_weights(self, input):
|
|
|
|
if fp8_matrix_mult:
|
|
|
|
out = fp8_linear(self, input)
|
|
|
|
if out is not None:
|
|
|
|
return out
|
|
|
|
|
|
|
|
weight, bias = cast_bias_weight(self, input)
|
|
|
|
return torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias)
|
|
|
|
|
|
|
|
def convert_weight(self, weight):
|
|
|
|
return weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype)
|
2024-08-20 15:49:33 +00:00
|
|
|
|
2024-10-20 03:47:42 +00:00
|
|
|
def set_weight(self, weight, inplace_update=False, seed=None, **kwargs):
|
|
|
|
weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed)
|
|
|
|
if inplace_update:
|
|
|
|
self.weight.data.copy_(weight)
|
|
|
|
else:
|
|
|
|
self.weight = torch.nn.Parameter(weight, requires_grad=False)
|
|
|
|
|
|
|
|
return scaled_fp8_op
|
|
|
|
|
|
|
|
def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=False):
|
|
|
|
fp8_compute = comfy.model_management.supports_fp8_compute(load_device)
|
|
|
|
if scaled_fp8:
|
|
|
|
return scaled_fp8_ops(fp8_matrix_mult=fp8_compute)
|
|
|
|
|
|
|
|
if fp8_compute and (fp8_optimizations or args.fast) and not disable_fast_fp8:
|
|
|
|
return fp8_ops
|
2024-10-09 23:43:17 +00:00
|
|
|
|
2024-08-20 15:49:33 +00:00
|
|
|
if compute_dtype is None or weight_dtype == compute_dtype:
|
|
|
|
return disable_weight_init
|
2024-10-20 03:47:42 +00:00
|
|
|
|
2024-08-20 15:49:33 +00:00
|
|
|
return manual_cast
|