| 121 | #logger.error(f"set control for LLLitePatch: {id(self)}, cn: {id(control)}") |
| 122 | |
| 123 | def clone_with_control(self, control: AdvancedControlBase): |
| 124 | #logger.error(f"clone-set control for LLLitePatch: {id(self)},{id(control)}") |
| 125 | return LLLitePatch(self.modules, self.patch_type, control) |
| 126 | |
| 127 | def cleanup(self): |
| 128 | for module in self.modules.values(): |
| 129 | module.cleanup() |
| 130 | |
| 131 | |
| 132 | # TODO: use comfy.ops to support fp8 properly |
| 133 | class LLLiteModule(torch.nn.Module): |
| 134 | def __init__( |
| 135 | self, |
| 136 | name: str, |
| 137 | is_conv2d: bool, |
| 138 | in_dim: int, |
| 139 | depth: int, |
| 140 | cond_emb_dim: int, |
| 141 | mlp_dim: int, |
| 142 | ): |
| 143 | super().__init__() |
| 144 | self.name = name |
| 145 | self.is_conv2d = is_conv2d |
| 146 | self.is_first = False |
| 147 | |
| 148 | modules = [] |
| 149 | modules.append(torch.nn.Conv2d(3, cond_emb_dim // 2, kernel_size=4, stride=4, padding=0)) # to latent (from VAE) size*2 |
| 150 | if depth == 1: |
| 151 | modules.append(torch.nn.ReLU(inplace=True)) |
| 152 | modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim, kernel_size=2, stride=2, padding=0)) |
| 153 | elif depth == 2: |
| 154 | modules.append(torch.nn.ReLU(inplace=True)) |
| 155 | modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim, kernel_size=4, stride=4, padding=0)) |
| 156 | elif depth == 3: |
| 157 | # kernel size 8 is too large, so set it to 4 |
| 158 | modules.append(torch.nn.ReLU(inplace=True)) |
| 159 | modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim // 2, kernel_size=4, stride=4, padding=0)) |
| 160 | modules.append(torch.nn.ReLU(inplace=True)) |
| 161 | modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim, kernel_size=2, stride=2, padding=0)) |
| 162 | |
| 163 | self.conditioning1 = torch.nn.Sequential(*modules) |
| 164 | |
| 165 | if self.is_conv2d: |
| 166 | self.down = torch.nn.Sequential( |
| 167 | torch.nn.Conv2d(in_dim, mlp_dim, kernel_size=1, stride=1, padding=0), |
| 168 | torch.nn.ReLU(inplace=True), |
| 169 | ) |
| 170 | self.mid = torch.nn.Sequential( |
| 171 | torch.nn.Conv2d(mlp_dim + cond_emb_dim, mlp_dim, kernel_size=1, stride=1, padding=0), |
| 172 | torch.nn.ReLU(inplace=True), |
| 173 | ) |
| 174 | self.up = torch.nn.Sequential( |
| 175 | torch.nn.Conv2d(mlp_dim, in_dim, kernel_size=1, stride=1, padding=0), |
| 176 | ) |
| 177 | else: |
| 178 | self.down = torch.nn.Sequential( |
| 179 | torch.nn.Linear(in_dim, mlp_dim), |
| 180 | torch.nn.ReLU(inplace=True), |
no outgoing calls
no test coverage detected