From a10c8df8761c01801bac60d7977ae7e997ab51b0 Mon Sep 17 00:00:00 2001 From: AUTOMATIC1111 <16777216c@gmail.com> Date: Mon, 26 Feb 2024 07:12:12 +0300 Subject: [PATCH] Merge pull request #14973 from AUTOMATIC1111/Fix-new-oft-boft Fix the OFT/BOFT bugs when using new LyCORIS implementation --- extensions-builtin/Lora/network_oft.py | 50 +++++++++++++------------- 1 file changed, 24 insertions(+), 26 deletions(-) diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index d658ad109..7821a8a7d 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -1,6 +1,5 @@ import torch import network -from lyco_helpers import factorization from einops import rearrange @@ -22,24 +21,24 @@ class NetworkModuleOFT(network.NetworkModule): self.org_module: list[torch.Module] = [self.sd_module] self.scale = 1.0 - self.is_kohya = False + self.is_R = False self.is_boft = False - # kohya-ss + # kohya-ss/New LyCORIS OFT/BOFT if "oft_blocks" in weights.w.keys(): - self.is_kohya = True self.oft_blocks = weights.w["oft_blocks"] # (num_blocks, block_size, block_size) - self.alpha = weights.w["alpha"] # alpha is constraint + self.alpha = weights.w.get("alpha", None) # alpha is constraint self.dim = self.oft_blocks.shape[0] # lora dim - # LyCORIS OFT + # Old LyCORIS OFT elif "oft_diag" in weights.w.keys(): + self.is_R = True self.oft_blocks = weights.w["oft_diag"] # self.alpha is unused self.dim = self.oft_blocks.shape[1] # (num_blocks, block_size, block_size) - # LyCORIS BOFT - if weights.w["oft_diag"].dim() == 4: - self.is_boft = True + # LyCORIS BOFT + if self.oft_blocks.dim() == 4: + self.is_boft = True self.rescale = weights.w.get('rescale', None) if self.rescale is not None: self.rescale = self.rescale.reshape(-1, *[1]*(self.org_module[0].weight.dim() - 1)) @@ -55,30 +54,29 @@ class NetworkModuleOFT(network.NetworkModule): elif is_other_linear: self.out_dim = self.sd_module.embed_dim - if self.is_kohya: - self.constraint = self.alpha * self.out_dim - self.num_blocks = self.dim - self.block_size = self.out_dim // self.dim + self.num_blocks = self.dim + self.block_size = self.out_dim // self.dim + self.constraint = (0 if self.alpha is None else self.alpha) * self.out_dim + if self.is_R: + self.constraint = None + self.block_size = self.dim + self.num_blocks = self.out_dim // self.dim elif self.is_boft: - self.constraint = None - self.boft_m = weights.w["oft_diag"].shape[0] - self.block_num = weights.w["oft_diag"].shape[1] - self.block_size = weights.w["oft_diag"].shape[2] + self.boft_m = self.oft_blocks.shape[0] + self.num_blocks = self.oft_blocks.shape[1] + self.block_size = self.oft_blocks.shape[2] self.boft_b = self.block_size - #self.block_size, self.block_num = butterfly_factor(self.out_dim, self.dim) - else: - self.constraint = None - self.block_size, self.num_blocks = factorization(self.out_dim, self.dim) def calc_updown(self, orig_weight): oft_blocks = self.oft_blocks.to(orig_weight.device) eye = torch.eye(self.block_size, device=oft_blocks.device) - if self.is_kohya: - block_Q = oft_blocks - oft_blocks.transpose(1, 2) # ensure skew-symmetric orthogonal matrix - norm_Q = torch.norm(block_Q.flatten()) - new_norm_Q = torch.clamp(norm_Q, max=self.constraint.to(oft_blocks.device)) - block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8)) + if not self.is_R: + block_Q = oft_blocks - oft_blocks.transpose(-1, -2) # ensure skew-symmetric orthogonal matrix + if self.constraint != 0: + norm_Q = torch.norm(block_Q.flatten()) + new_norm_Q = torch.clamp(norm_Q, max=self.constraint.to(oft_blocks.device)) + block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8)) oft_blocks = torch.matmul(eye + block_Q, (eye - block_Q).float().inverse()) R = oft_blocks.to(orig_weight.device)