From 7c128bbdac0da1767c239174e91af6f327845372 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Thu, 19 Oct 2023 13:56:17 +0800 Subject: Add fp8 for sd unet --- extensions-builtin/Lora/network.py | 2 +- extensions-builtin/Lora/network_full.py | 4 ++-- extensions-builtin/Lora/network_glora.py | 10 +++++----- extensions-builtin/Lora/network_hada.py | 12 ++++++------ extensions-builtin/Lora/network_ia3.py | 2 +- extensions-builtin/Lora/network_lokr.py | 18 +++++++++--------- extensions-builtin/Lora/network_lora.py | 6 +++--- extensions-builtin/Lora/network_norm.py | 4 ++-- extensions-builtin/Lora/networks.py | 6 +++--- 9 files changed, 32 insertions(+), 32 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network.py b/extensions-builtin/Lora/network.py index 6021fd8d..a62e5eff 100644 --- a/extensions-builtin/Lora/network.py +++ b/extensions-builtin/Lora/network.py @@ -137,7 +137,7 @@ class NetworkModule: def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None): if self.bias is not None: updown = updown.reshape(self.bias.shape) - updown += self.bias.to(orig_weight.device, dtype=orig_weight.dtype) + updown += self.bias.to(orig_weight.device, dtype=updown.dtype) updown = updown.reshape(output_shape) if len(output_shape) == 4: diff --git a/extensions-builtin/Lora/network_full.py b/extensions-builtin/Lora/network_full.py index bf6930e9..f221c95f 100644 --- a/extensions-builtin/Lora/network_full.py +++ b/extensions-builtin/Lora/network_full.py @@ -18,9 +18,9 @@ class NetworkModuleFull(network.NetworkModule): def calc_updown(self, orig_weight): output_shape = self.weight.shape - updown = self.weight.to(orig_weight.device, dtype=orig_weight.dtype) + updown = self.weight.to(orig_weight.device) if self.ex_bias is not None: - ex_bias = self.ex_bias.to(orig_weight.device, dtype=orig_weight.dtype) + ex_bias = self.ex_bias.to(orig_weight.device) else: ex_bias = None diff --git a/extensions-builtin/Lora/network_glora.py b/extensions-builtin/Lora/network_glora.py index 492d4870..efe5c681 100644 --- a/extensions-builtin/Lora/network_glora.py +++ b/extensions-builtin/Lora/network_glora.py @@ -22,12 +22,12 @@ class NetworkModuleGLora(network.NetworkModule): self.w2b = weights.w["b2.weight"] def calc_updown(self, orig_weight): - w1a = self.w1a.to(orig_weight.device, dtype=orig_weight.dtype) - w1b = self.w1b.to(orig_weight.device, dtype=orig_weight.dtype) - w2a = self.w2a.to(orig_weight.device, dtype=orig_weight.dtype) - w2b = self.w2b.to(orig_weight.device, dtype=orig_weight.dtype) + w1a = self.w1a.to(orig_weight.device) + w1b = self.w1b.to(orig_weight.device) + w2a = self.w2a.to(orig_weight.device) + w2b = self.w2b.to(orig_weight.device) output_shape = [w1a.size(0), w1b.size(1)] - updown = ((w2b @ w1b) + ((orig_weight @ w2a) @ w1a)) + updown = ((w2b @ w1b) + ((orig_weight.to(dtype = w1a.dtype) @ w2a) @ w1a)) return self.finalize_updown(updown, orig_weight, output_shape) diff --git a/extensions-builtin/Lora/network_hada.py b/extensions-builtin/Lora/network_hada.py index 5fcb0695..d95a0fd1 100644 --- a/extensions-builtin/Lora/network_hada.py +++ b/extensions-builtin/Lora/network_hada.py @@ -27,16 +27,16 @@ class NetworkModuleHada(network.NetworkModule): self.t2 = weights.w.get("hada_t2") def calc_updown(self, orig_weight): - w1a = self.w1a.to(orig_weight.device, dtype=orig_weight.dtype) - w1b = self.w1b.to(orig_weight.device, dtype=orig_weight.dtype) - w2a = self.w2a.to(orig_weight.device, dtype=orig_weight.dtype) - w2b = self.w2b.to(orig_weight.device, dtype=orig_weight.dtype) + w1a = self.w1a.to(orig_weight.device) + w1b = self.w1b.to(orig_weight.device) + w2a = self.w2a.to(orig_weight.device) + w2b = self.w2b.to(orig_weight.device) output_shape = [w1a.size(0), w1b.size(1)] if self.t1 is not None: output_shape = [w1a.size(1), w1b.size(1)] - t1 = self.t1.to(orig_weight.device, dtype=orig_weight.dtype) + t1 = self.t1.to(orig_weight.device) updown1 = lyco_helpers.make_weight_cp(t1, w1a, w1b) output_shape += t1.shape[2:] else: @@ -45,7 +45,7 @@ class NetworkModuleHada(network.NetworkModule): updown1 = lyco_helpers.rebuild_conventional(w1a, w1b, output_shape) if self.t2 is not None: - t2 = self.t2.to(orig_weight.device, dtype=orig_weight.dtype) + t2 = self.t2.to(orig_weight.device) updown2 = lyco_helpers.make_weight_cp(t2, w2a, w2b) else: updown2 = lyco_helpers.rebuild_conventional(w2a, w2b, output_shape) diff --git a/extensions-builtin/Lora/network_ia3.py b/extensions-builtin/Lora/network_ia3.py index 7edc4249..96faeaf3 100644 --- a/extensions-builtin/Lora/network_ia3.py +++ b/extensions-builtin/Lora/network_ia3.py @@ -17,7 +17,7 @@ class NetworkModuleIa3(network.NetworkModule): self.on_input = weights.w["on_input"].item() def calc_updown(self, orig_weight): - w = self.w.to(orig_weight.device, dtype=orig_weight.dtype) + w = self.w.to(orig_weight.device) output_shape = [w.size(0), orig_weight.size(1)] if self.on_input: diff --git a/extensions-builtin/Lora/network_lokr.py b/extensions-builtin/Lora/network_lokr.py index 340acdab..fcdaeafd 100644 --- a/extensions-builtin/Lora/network_lokr.py +++ b/extensions-builtin/Lora/network_lokr.py @@ -37,22 +37,22 @@ class NetworkModuleLokr(network.NetworkModule): def calc_updown(self, orig_weight): if self.w1 is not None: - w1 = self.w1.to(orig_weight.device, dtype=orig_weight.dtype) + w1 = self.w1.to(orig_weight.device) else: - w1a = self.w1a.to(orig_weight.device, dtype=orig_weight.dtype) - w1b = self.w1b.to(orig_weight.device, dtype=orig_weight.dtype) + w1a = self.w1a.to(orig_weight.device) + w1b = self.w1b.to(orig_weight.device) w1 = w1a @ w1b if self.w2 is not None: - w2 = self.w2.to(orig_weight.device, dtype=orig_weight.dtype) + w2 = self.w2.to(orig_weight.device) elif self.t2 is None: - w2a = self.w2a.to(orig_weight.device, dtype=orig_weight.dtype) - w2b = self.w2b.to(orig_weight.device, dtype=orig_weight.dtype) + w2a = self.w2a.to(orig_weight.device) + w2b = self.w2b.to(orig_weight.device) w2 = w2a @ w2b else: - t2 = self.t2.to(orig_weight.device, dtype=orig_weight.dtype) - w2a = self.w2a.to(orig_weight.device, dtype=orig_weight.dtype) - w2b = self.w2b.to(orig_weight.device, dtype=orig_weight.dtype) + t2 = self.t2.to(orig_weight.device) + w2a = self.w2a.to(orig_weight.device) + w2b = self.w2b.to(orig_weight.device) w2 = lyco_helpers.make_weight_cp(t2, w2a, w2b) output_shape = [w1.size(0) * w2.size(0), w1.size(1) * w2.size(1)] diff --git a/extensions-builtin/Lora/network_lora.py b/extensions-builtin/Lora/network_lora.py index 26c0a72c..4cc40295 100644 --- a/extensions-builtin/Lora/network_lora.py +++ b/extensions-builtin/Lora/network_lora.py @@ -61,13 +61,13 @@ class NetworkModuleLora(network.NetworkModule): return module def calc_updown(self, orig_weight): - up = self.up_model.weight.to(orig_weight.device, dtype=orig_weight.dtype) - down = self.down_model.weight.to(orig_weight.device, dtype=orig_weight.dtype) + up = self.up_model.weight.to(orig_weight.device) + down = self.down_model.weight.to(orig_weight.device) output_shape = [up.size(0), down.size(1)] if self.mid_model is not None: # cp-decomposition - mid = self.mid_model.weight.to(orig_weight.device, dtype=orig_weight.dtype) + mid = self.mid_model.weight.to(orig_weight.device) updown = lyco_helpers.rebuild_cp_decomposition(up, down, mid) output_shape += mid.shape[2:] else: diff --git a/extensions-builtin/Lora/network_norm.py b/extensions-builtin/Lora/network_norm.py index ce450158..d25afcbb 100644 --- a/extensions-builtin/Lora/network_norm.py +++ b/extensions-builtin/Lora/network_norm.py @@ -18,10 +18,10 @@ class NetworkModuleNorm(network.NetworkModule): def calc_updown(self, orig_weight): output_shape = self.w_norm.shape - updown = self.w_norm.to(orig_weight.device, dtype=orig_weight.dtype) + updown = self.w_norm.to(orig_weight.device) if self.b_norm is not None: - ex_bias = self.b_norm.to(orig_weight.device, dtype=orig_weight.dtype) + ex_bias = self.b_norm.to(orig_weight.device) else: ex_bias = None diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 60d8dec4..8ea4ea60 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -381,12 +381,12 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn # inpainting model. zero pad updown to make channel[1] 4 to 9 updown = torch.nn.functional.pad(updown, (0, 0, 0, 0, 0, 5)) - self.weight += updown + self.weight.copy_((self.weight.to(dtype=updown.dtype) + updown).to(dtype=self.weight.dtype)) if ex_bias is not None and hasattr(self, 'bias'): if self.bias is None: - self.bias = torch.nn.Parameter(ex_bias) + self.bias = torch.nn.Parameter(ex_bias).to(self.weight.dtype) else: - self.bias += ex_bias + self.bias.copy_((self.bias.to(dtype=ex_bias.dtype) + ex_bias).to(dtype=self.bias.dtype)) except RuntimeError as e: logging.debug(f"Network {net.name} layer {network_layer_name}: {e}") extra_network_lora.errors[net.name] = extra_network_lora.errors.get(net.name, 0) + 1 -- cgit v1.2.3 From 370a77f8e78e65a8a1339289d684cb43df142f70 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Tue, 21 Nov 2023 19:59:34 +0800 Subject: Option for using fp16 weight when apply lora --- extensions-builtin/Lora/networks.py | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 0170dbfb..d22ed843 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -388,18 +388,26 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn if module is not None and hasattr(self, 'weight'): try: with torch.no_grad(): - updown, ex_bias = module.calc_updown(self.weight) + if getattr(self, 'fp16_weight', None) is None: + weight = self.weight + bias = self.bias + else: + weight = self.fp16_weight.clone().to(self.weight.device) + bias = getattr(self, 'fp16_bias', None) + if bias is not None: + bias = bias.clone().to(self.bias.device) + updown, ex_bias = module.calc_updown(weight) - if len(self.weight.shape) == 4 and self.weight.shape[1] == 9: + if len(weight.shape) == 4 and weight.shape[1] == 9: # inpainting model. zero pad updown to make channel[1] 4 to 9 updown = torch.nn.functional.pad(updown, (0, 0, 0, 0, 0, 5)) - self.weight.copy_((self.weight.to(dtype=updown.dtype) + updown).to(dtype=self.weight.dtype)) + self.weight.copy_((weight.to(dtype=updown.dtype) + updown).to(dtype=self.weight.dtype)) if ex_bias is not None and hasattr(self, 'bias'): if self.bias is None: self.bias = torch.nn.Parameter(ex_bias).to(self.weight.dtype) else: - self.bias.copy_((self.bias.to(dtype=ex_bias.dtype) + ex_bias).to(dtype=self.bias.dtype)) + self.bias.copy_((bias + ex_bias).to(dtype=self.bias.dtype)) except RuntimeError as e: logging.debug(f"Network {net.name} layer {network_layer_name}: {e}") extra_network_lora.errors[net.name] = extra_network_lora.errors.get(net.name, 0) + 1 -- cgit v1.2.3 From 16bdcce92d5b482d50cdc32a8f308040d320b6c9 Mon Sep 17 00:00:00 2001 From: Rene Kroon Date: Fri, 8 Dec 2023 21:19:29 +0100 Subject: #13354: solve lora loading issue --- extensions-builtin/Lora/networks.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 7f814706..629bf853 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -159,7 +159,8 @@ def load_network(name, network_on_disk): bundle_embeddings = {} for key_network, weight in sd.items(): - key_network_without_network_parts, network_part = key_network.split(".", 1) + key_network_without_network_parts, _, network_part = key_network.partition(".") + if key_network_without_network_parts == "bundle_emb": emb_name, vec_name = network_part.split(".", 1) emb_dict = bundle_embeddings.get(emb_name, {}) -- cgit v1.2.3 From 735c9e8059384d4f640e5582413c30871f83eac5 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Thu, 14 Dec 2023 01:38:32 +0800 Subject: Fix network_oft --- extensions-builtin/Lora/network_oft.py | 21 +++++++++++---------- 1 file changed, 11 insertions(+), 10 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index 05c37811..44465f7a 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -53,12 +53,17 @@ class NetworkModuleOFT(network.NetworkModule): self.constraint = None self.block_size, self.num_blocks = factorization(self.out_dim, self.dim) - def calc_updown_kb(self, orig_weight, multiplier): + def calc_updown(self, orig_weight): + I = torch.eye(self.block_size, device=self.oft_blocks.device) oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype) - oft_blocks = oft_blocks - oft_blocks.transpose(1, 2) # ensure skew-symmetric orthogonal matrix + 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) + block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8)) + oft_blocks = torch.matmul(I + block_Q, (I - block_Q).float().inverse()) R = oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype) - R = R * multiplier + torch.eye(self.block_size, device=orig_weight.device) # This errors out for MultiheadAttention, might need to be handled up-stream merged_weight = rearrange(orig_weight, '(k n) ... -> k n ...', k=self.num_blocks, n=self.block_size) @@ -70,15 +75,10 @@ class NetworkModuleOFT(network.NetworkModule): merged_weight = rearrange(merged_weight, 'k m ... -> (k m) ...') updown = merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) - orig_weight + print(torch.norm(updown)) output_shape = orig_weight.shape return self.finalize_updown(updown, orig_weight, output_shape) - def calc_updown(self, orig_weight): - # if alpha is a very small number as in coft, calc_scale() will return a almost zero number so we ignore it - multiplier = self.multiplier() - return self.calc_updown_kb(orig_weight, multiplier) - - # override to remove the multiplier/scale factor; it's already multiplied in get_weight def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None): if self.bias is not None: updown = updown.reshape(self.bias.shape) @@ -94,4 +94,5 @@ class NetworkModuleOFT(network.NetworkModule): if ex_bias is not None: ex_bias = ex_bias * self.multiplier() - return updown, ex_bias + # Ignore calc_scale, which is not used in OFT. + return updown * self.multiplier(), ex_bias -- cgit v1.2.3 From 265bc26c21264d63956e8f30f1ce31dec917fc76 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Thu, 14 Dec 2023 01:43:24 +0800 Subject: Use self.scale instead of custom finalize --- extensions-builtin/Lora/network_oft.py | 20 ++------------------ 1 file changed, 2 insertions(+), 18 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index 44465f7a..e3ae61a2 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -21,6 +21,8 @@ class NetworkModuleOFT(network.NetworkModule): self.lin_module = None self.org_module: list[torch.Module] = [self.sd_module] + self.scale = 1.0 + # kohya-ss if "oft_blocks" in weights.w.keys(): self.is_kohya = True @@ -78,21 +80,3 @@ class NetworkModuleOFT(network.NetworkModule): print(torch.norm(updown)) output_shape = orig_weight.shape return self.finalize_updown(updown, orig_weight, output_shape) - - def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None): - if self.bias is not None: - updown = updown.reshape(self.bias.shape) - updown += self.bias.to(orig_weight.device, dtype=orig_weight.dtype) - updown = updown.reshape(output_shape) - - if len(output_shape) == 4: - updown = updown.reshape(output_shape) - - if orig_weight.size().numel() == updown.size().numel(): - updown = updown.reshape(orig_weight.shape) - - if ex_bias is not None: - ex_bias = ex_bias * self.multiplier() - - # Ignore calc_scale, which is not used in OFT. - return updown * self.multiplier(), ex_bias -- cgit v1.2.3 From 8fc67f3851babd4575d3312b931d5e7c2b0c78c6 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Thu, 14 Dec 2023 01:44:49 +0800 Subject: remove debug print --- extensions-builtin/Lora/network_oft.py | 1 - 1 file changed, 1 deletion(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index e3ae61a2..ff4eb59b 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -77,6 +77,5 @@ class NetworkModuleOFT(network.NetworkModule): merged_weight = rearrange(merged_weight, 'k m ... -> (k m) ...') updown = merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) - orig_weight - print(torch.norm(updown)) output_shape = orig_weight.shape return self.finalize_updown(updown, orig_weight, output_shape) -- cgit v1.2.3 From 3772a82a70769fe1aac884a75bf5a3313fb83328 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Thu, 14 Dec 2023 01:47:13 +0800 Subject: better naming and correct order for device. --- extensions-builtin/Lora/network_oft.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index ff4eb59b..fa647020 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -56,14 +56,15 @@ class NetworkModuleOFT(network.NetworkModule): self.block_size, self.num_blocks = factorization(self.out_dim, self.dim) def calc_updown(self, orig_weight): - I = torch.eye(self.block_size, device=self.oft_blocks.device) oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype) + eye = torch.eye(self.block_size, device=self.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) block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8)) - oft_blocks = torch.matmul(I + block_Q, (I - block_Q).float().inverse()) + oft_blocks = torch.matmul(eye + block_Q, (eye - block_Q).float().inverse()) R = oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype) -- cgit v1.2.3 From 59d060fd5ea93fcc3fdbfbd13b6e20fda06ecf94 Mon Sep 17 00:00:00 2001 From: w-e-w <40751091+w-e-w@users.noreply.github.com> Date: Sat, 30 Dec 2023 17:11:03 +0900 Subject: More lora not found warning --- extensions-builtin/Lora/networks.py | 8 +++++++- extensions-builtin/Lora/scripts/lora_script.py | 2 ++ 2 files changed, 9 insertions(+), 1 deletion(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 985b2753..72ebd624 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -1,3 +1,4 @@ +import gradio as gr import logging import os import re @@ -314,7 +315,12 @@ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=No emb_db.skipped_embeddings[name] = embedding if failed_to_load_networks: - sd_hijack.model_hijack.comments.append("Networks not found: " + ", ".join(failed_to_load_networks)) + lora_not_found_message = f'Lora not found: {", ".join(failed_to_load_networks)}' + sd_hijack.model_hijack.comments.append(lora_not_found_message) + if shared.opts.lora_not_found_warning_console: + print(f'\n{lora_not_found_message}\n') + if shared.opts.lora_not_found_gradio_warning: + gr.Warning(lora_not_found_message) purge_networks_from_memory() diff --git a/extensions-builtin/Lora/scripts/lora_script.py b/extensions-builtin/Lora/scripts/lora_script.py index ef23968c..1518f7e5 100644 --- a/extensions-builtin/Lora/scripts/lora_script.py +++ b/extensions-builtin/Lora/scripts/lora_script.py @@ -39,6 +39,8 @@ shared.options_templates.update(shared.options_section(('extra_networks', "Extra "lora_show_all": shared.OptionInfo(False, "Always show all networks on the Lora page").info("otherwise, those detected as for incompatible version of Stable Diffusion will be hidden"), "lora_hide_unknown_for_versions": shared.OptionInfo([], "Hide networks of unknown versions for model versions", gr.CheckboxGroup, {"choices": ["SD1", "SD2", "SDXL"]}), "lora_in_memory_limit": shared.OptionInfo(0, "Number of Lora networks to keep cached in memory", gr.Number, {"precision": 0}), + "lora_not_found_warning_console": shared.OptionInfo(False, "Lora not found warning in console"), + "lora_not_found_gradio_warning": shared.OptionInfo(False, "Lora not found warning popup in webui"), })) -- cgit v1.2.3 From bc5ae74c7d8949bab37e260b16e76889b9968099 Mon Sep 17 00:00:00 2001 From: Learwin <6223515+Learwin@users.noreply.github.com> Date: Sat, 30 Dec 2023 21:52:27 +0100 Subject: Added negative prompts to extra networks lora --- extensions-builtin/Lora/ui_edit_user_metadata.py | 14 ++++++++++++-- extensions-builtin/Lora/ui_extra_networks_lora.py | 9 +++++++++ 2 files changed, 21 insertions(+), 2 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/ui_edit_user_metadata.py b/extensions-builtin/Lora/ui_edit_user_metadata.py index c7011909..f7859b21 100644 --- a/extensions-builtin/Lora/ui_edit_user_metadata.py +++ b/extensions-builtin/Lora/ui_edit_user_metadata.py @@ -54,12 +54,14 @@ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor) self.slider_preferred_weight = None self.edit_notes = None - def save_lora_user_metadata(self, name, desc, sd_version, activation_text, preferred_weight, notes): + def save_lora_user_metadata(self, name, desc, sd_version, activation_text, preferred_weight, negative_text, negative_weight, notes): user_metadata = self.get_user_metadata(name) user_metadata["description"] = desc user_metadata["sd version"] = sd_version user_metadata["activation text"] = activation_text user_metadata["preferred weight"] = preferred_weight + user_metadata["negative text"] = negative_text + user_metadata["negative weight"] = negative_weight user_metadata["notes"] = notes self.write_user_metadata(name, user_metadata) @@ -127,6 +129,8 @@ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor) gr.HighlightedText.update(value=gradio_tags, visible=True if tags else False), user_metadata.get('activation text', ''), float(user_metadata.get('preferred weight', 0.0)), + user_metadata.get('negative text', ''), + float(user_metadata.get('negative weight', 0.0)), gr.update(visible=True if tags else False), gr.update(value=self.generate_random_prompt_from_tags(tags), visible=True if tags else False), ] @@ -162,7 +166,8 @@ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor) self.taginfo = gr.HighlightedText(label="Training dataset tags") self.edit_activation_text = gr.Text(label='Activation text', info="Will be added to prompt along with Lora") self.slider_preferred_weight = gr.Slider(label='Preferred weight', info="Set to 0 to disable", minimum=0.0, maximum=2.0, step=0.01) - + self.edit_negative_text = gr.Text(label='Negative prompt', info="Will be added to negative prompts") + self.slider_negative_weight = gr.Slider(label='Preferred negative weight', info="Set to 0 to disable", minimum=0.0, maximum=2.0, step=0.01) with gr.Row() as row_random_prompt: with gr.Column(scale=8): random_prompt = gr.Textbox(label='Random prompt', lines=4, max_lines=4, interactive=False) @@ -198,6 +203,8 @@ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor) self.taginfo, self.edit_activation_text, self.slider_preferred_weight, + self.edit_negative_text, + self.slider_negative_weight, row_random_prompt, random_prompt, ] @@ -211,7 +218,10 @@ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor) self.select_sd_version, self.edit_activation_text, self.slider_preferred_weight, + self.edit_negative_text, + self.slider_negative_weight, self.edit_notes, ] + self.setup_save_handler(self.button_save, self.save_lora_user_metadata, edited_components) diff --git a/extensions-builtin/Lora/ui_extra_networks_lora.py b/extensions-builtin/Lora/ui_extra_networks_lora.py index df02c663..09ce2a05 100644 --- a/extensions-builtin/Lora/ui_extra_networks_lora.py +++ b/extensions-builtin/Lora/ui_extra_networks_lora.py @@ -45,6 +45,15 @@ class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage): if activation_text: item["prompt"] += " + " + quote_js(" " + activation_text) + negative_prompt = item["user_metadata"].get("negative text") + preferred_negative_weight = item["user_metadata"].get("negative weight") + item["negative_prompt"] = quote_js("") + if negative_prompt: + neg_prompt = negative_prompt + if (preferred_negative_weight > 0): + neg_prompt = '(' + negative_prompt + ':' + str(preferred_negative_weight) + ')' + item["negative_prompt"] = quote_js(neg_prompt) + sd_version = item["user_metadata"].get("sd version") if sd_version in network.SdVersion.__members__: item["sd_version"] = sd_version -- cgit v1.2.3 From a2f23f9d22dde87bf2529dcb2854a6a5d3d44278 Mon Sep 17 00:00:00 2001 From: Learwin <6223515+Learwin@users.noreply.github.com> Date: Sat, 30 Dec 2023 22:16:51 +0100 Subject: Code Style fixes --- extensions-builtin/Lora/ui_extra_networks_lora.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/ui_extra_networks_lora.py b/extensions-builtin/Lora/ui_extra_networks_lora.py index 09ce2a05..9a6624e3 100644 --- a/extensions-builtin/Lora/ui_extra_networks_lora.py +++ b/extensions-builtin/Lora/ui_extra_networks_lora.py @@ -52,8 +52,8 @@ class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage): neg_prompt = negative_prompt if (preferred_negative_weight > 0): neg_prompt = '(' + negative_prompt + ':' + str(preferred_negative_weight) + ')' - item["negative_prompt"] = quote_js(neg_prompt) - + item["negative_prompt"] = quote_js(neg_prompt) + sd_version = item["user_metadata"].get("sd version") if sd_version in network.SdVersion.__members__: item["sd_version"] = sd_version -- cgit v1.2.3 From d4945f4422e5a0bf31a6dbe4c1aeedd78c09eacb Mon Sep 17 00:00:00 2001 From: Learwin <6223515+Learwin@users.noreply.github.com> Date: Sun, 31 Dec 2023 13:22:30 +0100 Subject: Removed weight slider for negative prompts --- extensions-builtin/Lora/ui_edit_user_metadata.py | 7 +------ extensions-builtin/Lora/ui_extra_networks_lora.py | 6 +----- 2 files changed, 2 insertions(+), 11 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/ui_edit_user_metadata.py b/extensions-builtin/Lora/ui_edit_user_metadata.py index f7859b21..3160aecf 100644 --- a/extensions-builtin/Lora/ui_edit_user_metadata.py +++ b/extensions-builtin/Lora/ui_edit_user_metadata.py @@ -54,14 +54,13 @@ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor) self.slider_preferred_weight = None self.edit_notes = None - def save_lora_user_metadata(self, name, desc, sd_version, activation_text, preferred_weight, negative_text, negative_weight, notes): + def save_lora_user_metadata(self, name, desc, sd_version, activation_text, preferred_weight, negative_text, notes): user_metadata = self.get_user_metadata(name) user_metadata["description"] = desc user_metadata["sd version"] = sd_version user_metadata["activation text"] = activation_text user_metadata["preferred weight"] = preferred_weight user_metadata["negative text"] = negative_text - user_metadata["negative weight"] = negative_weight user_metadata["notes"] = notes self.write_user_metadata(name, user_metadata) @@ -130,7 +129,6 @@ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor) user_metadata.get('activation text', ''), float(user_metadata.get('preferred weight', 0.0)), user_metadata.get('negative text', ''), - float(user_metadata.get('negative weight', 0.0)), gr.update(visible=True if tags else False), gr.update(value=self.generate_random_prompt_from_tags(tags), visible=True if tags else False), ] @@ -167,7 +165,6 @@ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor) self.edit_activation_text = gr.Text(label='Activation text', info="Will be added to prompt along with Lora") self.slider_preferred_weight = gr.Slider(label='Preferred weight', info="Set to 0 to disable", minimum=0.0, maximum=2.0, step=0.01) self.edit_negative_text = gr.Text(label='Negative prompt', info="Will be added to negative prompts") - self.slider_negative_weight = gr.Slider(label='Preferred negative weight', info="Set to 0 to disable", minimum=0.0, maximum=2.0, step=0.01) with gr.Row() as row_random_prompt: with gr.Column(scale=8): random_prompt = gr.Textbox(label='Random prompt', lines=4, max_lines=4, interactive=False) @@ -204,7 +201,6 @@ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor) self.edit_activation_text, self.slider_preferred_weight, self.edit_negative_text, - self.slider_negative_weight, row_random_prompt, random_prompt, ] @@ -219,7 +215,6 @@ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor) self.edit_activation_text, self.slider_preferred_weight, self.edit_negative_text, - self.slider_negative_weight, self.edit_notes, ] diff --git a/extensions-builtin/Lora/ui_extra_networks_lora.py b/extensions-builtin/Lora/ui_extra_networks_lora.py index 9a6624e3..e714fac4 100644 --- a/extensions-builtin/Lora/ui_extra_networks_lora.py +++ b/extensions-builtin/Lora/ui_extra_networks_lora.py @@ -46,13 +46,9 @@ class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage): item["prompt"] += " + " + quote_js(" " + activation_text) negative_prompt = item["user_metadata"].get("negative text") - preferred_negative_weight = item["user_metadata"].get("negative weight") item["negative_prompt"] = quote_js("") if negative_prompt: - neg_prompt = negative_prompt - if (preferred_negative_weight > 0): - neg_prompt = '(' + negative_prompt + ':' + str(preferred_negative_weight) + ')' - item["negative_prompt"] = quote_js(neg_prompt) + item["negative_prompt"] = quote_js('(' + negative_prompt + ':1)') sd_version = item["user_metadata"].get("sd version") if sd_version in network.SdVersion.__members__: -- cgit v1.2.3 From f8f38c7c28e48f9f79225c969e3e82b1adcfb910 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Fri, 5 Jan 2024 16:31:48 +0800 Subject: Fix dtype casting for OFT module --- extensions-builtin/Lora/network_oft.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index fa647020..342fcd0d 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -56,7 +56,7 @@ class NetworkModuleOFT(network.NetworkModule): 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, dtype=orig_weight.dtype) + oft_blocks = self.oft_blocks.to(orig_weight.device) eye = torch.eye(self.block_size, device=self.oft_blocks.device) if self.is_kohya: @@ -66,7 +66,7 @@ class NetworkModuleOFT(network.NetworkModule): 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, dtype=orig_weight.dtype) + R = oft_blocks.to(orig_weight.device) # This errors out for MultiheadAttention, might need to be handled up-stream merged_weight = rearrange(orig_weight, '(k n) ... -> k n ...', k=self.num_blocks, n=self.block_size) @@ -77,6 +77,6 @@ class NetworkModuleOFT(network.NetworkModule): ) merged_weight = rearrange(merged_weight, 'k m ... -> (k m) ...') - updown = merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) - orig_weight + updown = merged_weight.to(orig_weight.device) - orig_weight.to(merged_weight.dtype) output_shape = orig_weight.shape return self.finalize_updown(updown, orig_weight, output_shape) -- cgit v1.2.3 From 18ca987c92f52690daec43a6c67363c341bb6008 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Fri, 5 Jan 2024 16:32:19 +0800 Subject: Add general forward method for all modules. --- extensions-builtin/Lora/network.py | 34 +++++++++++++++++++++++++++++++++- extensions-builtin/Lora/networks.py | 12 ++++++------ 2 files changed, 39 insertions(+), 7 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network.py b/extensions-builtin/Lora/network.py index a62e5eff..f9b571b5 100644 --- a/extensions-builtin/Lora/network.py +++ b/extensions-builtin/Lora/network.py @@ -3,6 +3,10 @@ import os from collections import namedtuple import enum +import torch +import torch.nn as nn +import torch.nn.functional as F + from modules import sd_models, cache, errors, hashes, shared NetworkWeights = namedtuple('NetworkWeights', ['network_key', 'sd_key', 'w', 'sd_module']) @@ -115,6 +119,29 @@ class NetworkModule: if hasattr(self.sd_module, 'weight'): self.shape = self.sd_module.weight.shape + self.ops = None + self.extra_kwargs = {} + if isinstance(self.sd_module, nn.Conv2d): + self.ops = F.conv2d + self.extra_kwargs = { + 'stride': self.sd_module.stride, + 'padding': self.sd_module.padding + } + elif isinstance(self.sd_module, nn.Linear): + self.ops = F.linear + elif isinstance(self.sd_module, nn.LayerNorm): + self.ops = F.layer_norm + self.extra_kwargs = { + 'normalized_shape': self.sd_module.normalized_shape, + 'eps': self.sd_module.eps + } + elif isinstance(self.sd_module, nn.GroupNorm): + self.ops = F.group_norm + self.extra_kwargs = { + 'num_groups': self.sd_module.num_groups, + 'eps': self.sd_module.eps + } + self.dim = None self.bias = weights.w.get("bias") self.alpha = weights.w["alpha"].item() if "alpha" in weights.w else None @@ -155,5 +182,10 @@ class NetworkModule: raise NotImplementedError() def forward(self, x, y): - raise NotImplementedError() + """A general forward implementation for all modules""" + if self.ops is None: + raise NotImplementedError() + else: + updown, ex_bias = self.calc_updown(self.sd_module.weight) + return y + self.ops(x, weight=updown, bias=ex_bias, **self.extra_kwargs) diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 72ebd624..32e10b62 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -458,23 +458,23 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn self.network_current_names = wanted_names -def network_forward(module, input, original_forward): +def network_forward(org_module, input, original_forward): """ Old way of applying Lora by executing operations during layer's forward. Stacking many loras this way results in big performance degradation. """ if len(loaded_networks) == 0: - return original_forward(module, input) + return original_forward(org_module, input) input = devices.cond_cast_unet(input) - network_restore_weights_from_backup(module) - network_reset_cached_weight(module) + network_restore_weights_from_backup(org_module) + network_reset_cached_weight(org_module) - y = original_forward(module, input) + y = original_forward(org_module, input) - network_layer_name = getattr(module, 'network_layer_name', None) + network_layer_name = getattr(org_module, 'network_layer_name', None) for lora in loaded_networks: module = lora.modules.get(network_layer_name, None) if module is None: -- cgit v1.2.3 From 44744d6005da5e424267698ee3279caa597dfebc Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Fri, 5 Jan 2024 16:38:05 +0800 Subject: linting --- extensions-builtin/Lora/network.py | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network.py b/extensions-builtin/Lora/network.py index f9b571b5..b8fd9194 100644 --- a/extensions-builtin/Lora/network.py +++ b/extensions-builtin/Lora/network.py @@ -3,7 +3,6 @@ import os from collections import namedtuple import enum -import torch import torch.nn as nn import torch.nn.functional as F @@ -124,7 +123,7 @@ class NetworkModule: if isinstance(self.sd_module, nn.Conv2d): self.ops = F.conv2d self.extra_kwargs = { - 'stride': self.sd_module.stride, + 'stride': self.sd_module.stride, 'padding': self.sd_module.padding } elif isinstance(self.sd_module, nn.Linear): -- cgit v1.2.3 From 0726a6e12e85a37d1e514f5603acf9f058c11783 Mon Sep 17 00:00:00 2001 From: Sj-Si Date: Thu, 11 Jan 2024 15:06:57 -0500 Subject: Finish base layout. Fix bugs. Need to test for stability and clean up. --- extensions-builtin/Lora/ui_extra_networks_lora.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/ui_extra_networks_lora.py b/extensions-builtin/Lora/ui_extra_networks_lora.py index df02c663..db612fa2 100644 --- a/extensions-builtin/Lora/ui_extra_networks_lora.py +++ b/extensions-builtin/Lora/ui_extra_networks_lora.py @@ -24,13 +24,16 @@ class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage): alias = lora_on_disk.get_alias() + search_terms = [self.search_terms_from_path(lora_on_disk.filename)] + if lora_on_disk.hash: + search_terms.append(lora_on_disk.hash) item = { "name": name, "filename": lora_on_disk.filename, "shorthash": lora_on_disk.shorthash, "preview": self.find_preview(path), "description": self.find_description(path), - "search_term": self.search_terms_from_path(lora_on_disk.filename) + " " + (lora_on_disk.hash or ""), + "search_terms": search_terms, "local_preview": f"{path}.{shared.opts.samples_format}", "metadata": lora_on_disk.metadata, "sort_keys": {'default': index, **self.get_sort_keys(lora_on_disk.filename)}, -- cgit v1.2.3 From e280eb40556d881c5fc9c8941e388b8f6fe1f7e1 Mon Sep 17 00:00:00 2001 From: chi2nagisa Date: Tue, 16 Jan 2024 03:45:19 +0800 Subject: fix using wrong model caused by alias --- extensions-builtin/Lora/networks.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 32e10b62..83ea2802 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -260,11 +260,11 @@ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=No loaded_networks.clear() - networks_on_disk = [available_network_aliases.get(name, None) for name in names] + networks_on_disk = [available_networks.get(name, None) if name.lower() in forbidden_network_aliases else available_network_aliases.get(name, None) for name in names] if any(x is None for x in networks_on_disk): list_available_networks() - networks_on_disk = [available_network_aliases.get(name, None) for name in names] + networks_on_disk = [available_networks.get(name, None) if name.lower() in forbidden_network_aliases else available_network_aliases.get(name, None) for name in names] failed_to_load_networks = [] -- cgit v1.2.3 From fd383140cf405100f3c619f106472273a7545beb Mon Sep 17 00:00:00 2001 From: v0xie <28695009+v0xie@users.noreply.github.com> Date: Mon, 22 Jan 2024 02:52:34 -0800 Subject: fix: wrong devices for eye and constraint --- extensions-builtin/Lora/network_oft.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index 342fcd0d..d1c46a4b 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -57,12 +57,12 @@ class NetworkModuleOFT(network.NetworkModule): def calc_updown(self, orig_weight): oft_blocks = self.oft_blocks.to(orig_weight.device) - eye = torch.eye(self.block_size, device=self.oft_blocks.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) + 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()) -- cgit v1.2.3 From 06ab10a1be812036605e1472f054228562ea08d9 Mon Sep 17 00:00:00 2001 From: catboxanon <122327233+catboxanon@users.noreply.github.com> Date: Thu, 15 Feb 2024 14:22:13 -0500 Subject: Normalize cmd arg paths In particular, this fixes an issue on Windows where some functions will misbehave if forward slashes are provided rather than double backslashes. --- extensions-builtin/Lora/preload.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/preload.py b/extensions-builtin/Lora/preload.py index 50961be3..52fab29b 100644 --- a/extensions-builtin/Lora/preload.py +++ b/extensions-builtin/Lora/preload.py @@ -1,7 +1,8 @@ import os from modules import paths +from modules.paths_internal import normalized_filepath def preload(parser): - parser.add_argument("--lora-dir", type=str, help="Path to directory with Lora networks.", default=os.path.join(paths.models_path, 'Lora')) - parser.add_argument("--lyco-dir-backcompat", type=str, help="Path to directory with LyCORIS networks (for backawards compatibility; can also use --lyco-dir).", default=os.path.join(paths.models_path, 'LyCORIS')) + parser.add_argument("--lora-dir", type=normalized_filepath, help="Path to directory with Lora networks.", default=os.path.join(paths.models_path, 'Lora')) + parser.add_argument("--lyco-dir-backcompat", type=normalized_filepath, help="Path to directory with LyCORIS networks (for backawards compatibility; can also use --lyco-dir).", default=os.path.join(paths.models_path, 'LyCORIS')) -- cgit v1.2.3 From 92ab0ef7d65ededa758f81e52cf4f48f72d13564 Mon Sep 17 00:00:00 2001 From: AUTOMATIC1111 <16777216c@gmail.com> Date: Mon, 19 Feb 2024 10:05:30 +0300 Subject: Merge pull request #14871 from v0xie/boft Support inference with LyCORIS BOFT networks --- extensions-builtin/Lora/network_oft.py | 58 ++++++++++++++++++++++++++++------ 1 file changed, 48 insertions(+), 10 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index d1c46a4b..d658ad10 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -22,6 +22,8 @@ class NetworkModuleOFT(network.NetworkModule): self.org_module: list[torch.Module] = [self.sd_module] self.scale = 1.0 + self.is_kohya = False + self.is_boft = False # kohya-ss if "oft_blocks" in weights.w.keys(): @@ -29,13 +31,19 @@ class NetworkModuleOFT(network.NetworkModule): self.oft_blocks = weights.w["oft_blocks"] # (num_blocks, block_size, block_size) self.alpha = weights.w["alpha"] # alpha is constraint self.dim = self.oft_blocks.shape[0] # lora dim - # LyCORIS + # LyCORIS OFT elif "oft_diag" in weights.w.keys(): - self.is_kohya = False 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 + 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)) + is_linear = type(self.sd_module) in [torch.nn.Linear, torch.nn.modules.linear.NonDynamicallyQuantizableLinear] is_conv = type(self.sd_module) in [torch.nn.Conv2d] is_other_linear = type(self.sd_module) in [torch.nn.MultiheadAttention] # unsupported @@ -51,6 +59,13 @@ class NetworkModuleOFT(network.NetworkModule): self.constraint = self.alpha * self.out_dim self.num_blocks = self.dim self.block_size = 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_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) @@ -68,14 +83,37 @@ class NetworkModuleOFT(network.NetworkModule): R = oft_blocks.to(orig_weight.device) - # This errors out for MultiheadAttention, might need to be handled up-stream - merged_weight = rearrange(orig_weight, '(k n) ... -> k n ...', k=self.num_blocks, n=self.block_size) - merged_weight = torch.einsum( - 'k n m, k n ... -> k m ...', - R, - merged_weight - ) - merged_weight = rearrange(merged_weight, 'k m ... -> (k m) ...') + if not self.is_boft: + # This errors out for MultiheadAttention, might need to be handled up-stream + merged_weight = rearrange(orig_weight, '(k n) ... -> k n ...', k=self.num_blocks, n=self.block_size) + merged_weight = torch.einsum( + 'k n m, k n ... -> k m ...', + R, + merged_weight + ) + merged_weight = rearrange(merged_weight, 'k m ... -> (k m) ...') + else: + # TODO: determine correct value for scale + scale = 1.0 + m = self.boft_m + b = self.boft_b + r_b = b // 2 + inp = orig_weight + for i in range(m): + bi = R[i] # b_num, b_size, b_size + if i == 0: + # Apply multiplier/scale and rescale into first weight + bi = bi * scale + (1 - scale) * eye + inp = rearrange(inp, "(c g k) ... -> (c k g) ...", g=2, k=2**i * r_b) + inp = rearrange(inp, "(d b) ... -> d b ...", b=b) + inp = torch.einsum("b i j, b j ... -> b i ...", bi, inp) + inp = rearrange(inp, "d b ... -> (d b) ...") + inp = rearrange(inp, "(c k g) ... -> (c g k) ...", g=2, k=2**i * r_b) + merged_weight = inp + + # Rescale mechanism + if self.rescale is not None: + merged_weight = self.rescale.to(merged_weight) * merged_weight updown = merged_weight.to(orig_weight.device) - orig_weight.to(merged_weight.dtype) output_shape = orig_weight.shape -- cgit v1.2.3 From a10c8df8761c01801bac60d7977ae7e997ab51b0 Mon Sep 17 00:00:00 2001 From: AUTOMATIC1111 <16777216c@gmail.com> Date: Mon, 26 Feb 2024 07:12:12 +0300 Subject: 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(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index d658ad10..7821a8a7 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 - elif self.is_boft: + 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.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.block_size = self.dim + self.num_blocks = self.out_dim // self.dim + elif self.is_boft: + 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) -- cgit v1.2.3