From 534e938a62a136b96a2cdf64b31ae4a27a03ecf2 Mon Sep 17 00:00:00 2001 From: Ryan Dick Date: Mon, 9 Sep 2024 18:44:36 +0000 Subject: [PATCH] Add unit test for lora_model_from_flux_diffusers_state_dict(...). --- .../flux_diffusers_lora_conversion_utils.py | 208 +----------------- ...st_flux_diffusers_lora_conversion_utils.py | 29 +++ 2 files changed, 33 insertions(+), 204 deletions(-) create mode 100644 tests/backend/lora/conversions/test_flux_diffusers_lora_conversion_utils.py diff --git a/invokeai/backend/lora/conversions/flux_diffusers_lora_conversion_utils.py b/invokeai/backend/lora/conversions/flux_diffusers_lora_conversion_utils.py index c4bb969c2f..9c13183158 100644 --- a/invokeai/backend/lora/conversions/flux_diffusers_lora_conversion_utils.py +++ b/invokeai/backend/lora/conversions/flux_diffusers_lora_conversion_utils.py @@ -8,211 +8,8 @@ from invokeai.backend.lora.layers.lora_layer import LoRALayer from invokeai.backend.lora.layers.lora_layer_base import LoRALayerBase from invokeai.backend.lora.lora_model_raw import LoRAModelRaw -# def convert_flux_transformer_checkpoint_to_diffusers( -# original_state_dict, num_layers, num_single_layers, inner_dim, mlp_ratio=4.0 -# ): -# converted_state_dict = {} -# ## time_text_embed.timestep_embedder <- time_in -# converted_state_dict["time_text_embed.timestep_embedder.linear_1.weight"] = original_state_dict.pop( -# "time_in.in_layer.weight" -# ) -# converted_state_dict["time_text_embed.timestep_embedder.linear_1.bias"] = original_state_dict.pop( -# "time_in.in_layer.bias" -# ) -# converted_state_dict["time_text_embed.timestep_embedder.linear_2.weight"] = original_state_dict.pop( -# "time_in.out_layer.weight" -# ) -# converted_state_dict["time_text_embed.timestep_embedder.linear_2.bias"] = original_state_dict.pop( -# "time_in.out_layer.bias" -# ) - -# ## time_text_embed.text_embedder <- vector_in -# converted_state_dict["time_text_embed.text_embedder.linear_1.weight"] = original_state_dict.pop( -# "vector_in.in_layer.weight" -# ) -# converted_state_dict["time_text_embed.text_embedder.linear_1.bias"] = original_state_dict.pop( -# "vector_in.in_layer.bias" -# ) -# converted_state_dict["time_text_embed.text_embedder.linear_2.weight"] = original_state_dict.pop( -# "vector_in.out_layer.weight" -# ) -# converted_state_dict["time_text_embed.text_embedder.linear_2.bias"] = original_state_dict.pop( -# "vector_in.out_layer.bias" -# ) - -# # guidance -# has_guidance = any("guidance" in k for k in original_state_dict) -# if has_guidance: -# converted_state_dict["time_text_embed.guidance_embedder.linear_1.weight"] = original_state_dict.pop( -# "guidance_in.in_layer.weight" -# ) -# converted_state_dict["time_text_embed.guidance_embedder.linear_1.bias"] = original_state_dict.pop( -# "guidance_in.in_layer.bias" -# ) -# converted_state_dict["time_text_embed.guidance_embedder.linear_2.weight"] = original_state_dict.pop( -# "guidance_in.out_layer.weight" -# ) -# converted_state_dict["time_text_embed.guidance_embedder.linear_2.bias"] = original_state_dict.pop( -# "guidance_in.out_layer.bias" -# ) - -# # context_embedder -# converted_state_dict["context_embedder.weight"] = original_state_dict.pop("txt_in.weight") -# converted_state_dict["context_embedder.bias"] = original_state_dict.pop("txt_in.bias") - -# # x_embedder -# converted_state_dict["x_embedder.weight"] = original_state_dict.pop("img_in.weight") -# converted_state_dict["x_embedder.bias"] = original_state_dict.pop("img_in.bias") - -# # double transformer blocks -# for i in range(num_layers): -# block_prefix = f"transformer_blocks.{i}." -# # norms. -# ## norm1 -# converted_state_dict[f"{block_prefix}norm1.linear.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.img_mod.lin.weight" -# ) -# converted_state_dict[f"{block_prefix}norm1.linear.bias"] = original_state_dict.pop( -# f"double_blocks.{i}.img_mod.lin.bias" -# ) -# ## norm1_context -# converted_state_dict[f"{block_prefix}norm1_context.linear.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.txt_mod.lin.weight" -# ) -# converted_state_dict[f"{block_prefix}norm1_context.linear.bias"] = original_state_dict.pop( -# f"double_blocks.{i}.txt_mod.lin.bias" -# ) -# # Q, K, V -# sample_q, sample_k, sample_v = torch.chunk( -# original_state_dict.pop(f"double_blocks.{i}.img_attn.qkv.weight"), 3, dim=0 -# ) -# context_q, context_k, context_v = torch.chunk( -# original_state_dict.pop(f"double_blocks.{i}.txt_attn.qkv.weight"), 3, dim=0 -# ) -# sample_q_bias, sample_k_bias, sample_v_bias = torch.chunk( -# original_state_dict.pop(f"double_blocks.{i}.img_attn.qkv.bias"), 3, dim=0 -# ) -# context_q_bias, context_k_bias, context_v_bias = torch.chunk( -# original_state_dict.pop(f"double_blocks.{i}.txt_attn.qkv.bias"), 3, dim=0 -# ) -# converted_state_dict[f"{block_prefix}attn.to_q.weight"] = torch.cat([sample_q]) -# converted_state_dict[f"{block_prefix}attn.to_q.bias"] = torch.cat([sample_q_bias]) -# converted_state_dict[f"{block_prefix}attn.to_k.weight"] = torch.cat([sample_k]) -# converted_state_dict[f"{block_prefix}attn.to_k.bias"] = torch.cat([sample_k_bias]) -# converted_state_dict[f"{block_prefix}attn.to_v.weight"] = torch.cat([sample_v]) -# converted_state_dict[f"{block_prefix}attn.to_v.bias"] = torch.cat([sample_v_bias]) -# converted_state_dict[f"{block_prefix}attn.add_q_proj.weight"] = torch.cat([context_q]) -# converted_state_dict[f"{block_prefix}attn.add_q_proj.bias"] = torch.cat([context_q_bias]) -# converted_state_dict[f"{block_prefix}attn.add_k_proj.weight"] = torch.cat([context_k]) -# converted_state_dict[f"{block_prefix}attn.add_k_proj.bias"] = torch.cat([context_k_bias]) -# converted_state_dict[f"{block_prefix}attn.add_v_proj.weight"] = torch.cat([context_v]) -# converted_state_dict[f"{block_prefix}attn.add_v_proj.bias"] = torch.cat([context_v_bias]) -# # qk_norm -# converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.img_attn.norm.query_norm.scale" -# ) -# converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.img_attn.norm.key_norm.scale" -# ) -# converted_state_dict[f"{block_prefix}attn.norm_added_q.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.txt_attn.norm.query_norm.scale" -# ) -# converted_state_dict[f"{block_prefix}attn.norm_added_k.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.txt_attn.norm.key_norm.scale" -# ) -# # ff img_mlp -# converted_state_dict[f"{block_prefix}ff.net.0.proj.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.img_mlp.0.weight" -# ) -# converted_state_dict[f"{block_prefix}ff.net.0.proj.bias"] = original_state_dict.pop( -# f"double_blocks.{i}.img_mlp.0.bias" -# ) -# converted_state_dict[f"{block_prefix}ff.net.2.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.img_mlp.2.weight" -# ) -# converted_state_dict[f"{block_prefix}ff.net.2.bias"] = original_state_dict.pop( -# f"double_blocks.{i}.img_mlp.2.bias" -# ) -# converted_state_dict[f"{block_prefix}ff_context.net.0.proj.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.txt_mlp.0.weight" -# ) -# converted_state_dict[f"{block_prefix}ff_context.net.0.proj.bias"] = original_state_dict.pop( -# f"double_blocks.{i}.txt_mlp.0.bias" -# ) -# converted_state_dict[f"{block_prefix}ff_context.net.2.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.txt_mlp.2.weight" -# ) -# converted_state_dict[f"{block_prefix}ff_context.net.2.bias"] = original_state_dict.pop( -# f"double_blocks.{i}.txt_mlp.2.bias" -# ) -# # output projections. -# converted_state_dict[f"{block_prefix}attn.to_out.0.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.img_attn.proj.weight" -# ) -# converted_state_dict[f"{block_prefix}attn.to_out.0.bias"] = original_state_dict.pop( -# f"double_blocks.{i}.img_attn.proj.bias" -# ) -# converted_state_dict[f"{block_prefix}attn.to_add_out.weight"] = original_state_dict.pop( -# f"double_blocks.{i}.txt_attn.proj.weight" -# ) -# converted_state_dict[f"{block_prefix}attn.to_add_out.bias"] = original_state_dict.pop( -# f"double_blocks.{i}.txt_attn.proj.bias" -# ) - -# # single transfomer blocks -# for i in range(num_single_layers): -# block_prefix = f"single_transformer_blocks.{i}." -# # norm.linear <- single_blocks.0.modulation.lin -# converted_state_dict[f"{block_prefix}norm.linear.weight"] = original_state_dict.pop( -# f"single_blocks.{i}.modulation.lin.weight" -# ) -# converted_state_dict[f"{block_prefix}norm.linear.bias"] = original_state_dict.pop( -# f"single_blocks.{i}.modulation.lin.bias" -# ) -# # Q, K, V, mlp -# mlp_hidden_dim = int(inner_dim * mlp_ratio) -# split_size = (inner_dim, inner_dim, inner_dim, mlp_hidden_dim) -# q, k, v, mlp = torch.split(original_state_dict.pop(f"single_blocks.{i}.linear1.weight"), split_size, dim=0) -# q_bias, k_bias, v_bias, mlp_bias = torch.split( -# original_state_dict.pop(f"single_blocks.{i}.linear1.bias"), split_size, dim=0 -# ) -# converted_state_dict[f"{block_prefix}attn.to_q.weight"] = torch.cat([q]) -# converted_state_dict[f"{block_prefix}attn.to_q.bias"] = torch.cat([q_bias]) -# converted_state_dict[f"{block_prefix}attn.to_k.weight"] = torch.cat([k]) -# converted_state_dict[f"{block_prefix}attn.to_k.bias"] = torch.cat([k_bias]) -# converted_state_dict[f"{block_prefix}attn.to_v.weight"] = torch.cat([v]) -# converted_state_dict[f"{block_prefix}attn.to_v.bias"] = torch.cat([v_bias]) -# converted_state_dict[f"{block_prefix}proj_mlp.weight"] = torch.cat([mlp]) -# converted_state_dict[f"{block_prefix}proj_mlp.bias"] = torch.cat([mlp_bias]) -# # qk norm -# converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = original_state_dict.pop( -# f"single_blocks.{i}.norm.query_norm.scale" -# ) -# converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = original_state_dict.pop( -# f"single_blocks.{i}.norm.key_norm.scale" -# ) -# # output projections. -# converted_state_dict[f"{block_prefix}proj_out.weight"] = original_state_dict.pop( -# f"single_blocks.{i}.linear2.weight" -# ) -# converted_state_dict[f"{block_prefix}proj_out.bias"] = original_state_dict.pop( -# f"single_blocks.{i}.linear2.bias" -# ) - -# converted_state_dict["proj_out.weight"] = original_state_dict.pop("final_layer.linear.weight") -# converted_state_dict["proj_out.bias"] = original_state_dict.pop("final_layer.linear.bias") -# converted_state_dict["norm_out.linear.weight"] = swap_scale_shift( -# original_state_dict.pop("final_layer.adaLN_modulation.1.weight") -# ) -# converted_state_dict["norm_out.linear.bias"] = swap_scale_shift( -# original_state_dict.pop("final_layer.adaLN_modulation.1.bias") -# ) - -# return converted_state_dict - - -# TODO(ryand): What alpha should we use? 1.0? Rank of the matrix? +# TODO(ryand): What alpha should we use? 1.0? Rank of the LoRA? def lora_model_from_flux_diffusers_state_dict(state_dict: Dict[str, torch.Tensor], alpha: float = 1.0) -> LoRAModelRaw: # pyright: ignore[reportRedeclaration] (state_dict is intentionally re-declared) """Loads a state dict in the Diffusers FLUX LoRA format into a LoRAModelRaw object. @@ -222,6 +19,9 @@ def lora_model_from_flux_diffusers_state_dict(state_dict: Dict[str, torch.Tensor # Group keys by layer. grouped_state_dict: dict[str, dict[str, torch.Tensor]] = _group_by_layer(state_dict) + # Remove the "transformer." prefix from all keys. + grouped_state_dict = {k.replace("transformer.", ""): v for k, v in grouped_state_dict.items()} + # Constants for FLUX.1 num_double_layers = 19 num_single_layers = 38 diff --git a/tests/backend/lora/conversions/test_flux_diffusers_lora_conversion_utils.py b/tests/backend/lora/conversions/test_flux_diffusers_lora_conversion_utils.py new file mode 100644 index 0000000000..43982b5478 --- /dev/null +++ b/tests/backend/lora/conversions/test_flux_diffusers_lora_conversion_utils.py @@ -0,0 +1,29 @@ +import torch + +from invokeai.backend.lora.conversions.flux_diffusers_lora_conversion_utils import ( + lora_model_from_flux_diffusers_state_dict, +) +from tests.backend.lora.conversions.lora_state_dicts.flux_lora_diffusers_format import state_dict_keys + + +def test_lora_model_from_flux_diffusers_state_dict(): + """Test that lora_model_from_flux_diffusers_state_dict() can load a state dict in the Diffusers FLUX LoRA format.""" + # Construct a state dict that is in the Diffusers FLUX LoRA format. + state_dict: dict[str, torch.Tensor] = {} + for k in state_dict_keys: + state_dict[k] = torch.empty(1) + + # Load the state dict into a LoRAModelRaw object. + model = lora_model_from_flux_diffusers_state_dict(state_dict) + + # Check that the model has the correct number of LoRA layers. + expected_lora_layers: set[str] = set() + for k in state_dict_keys: + k = k.replace("lora_A.weight", "") + k = k.replace("lora_B.weight", "") + expected_lora_layers.add(k) + # Drop the K/V/proj_mlp weights because these are all concatenated into a single layer in the BFL format (we keep + # the Q weights so that we count these layers once). + concatenated_weights = ["to_k", "to_v", "proj_mlp", "add_k_proj", "add_v_proj"] + expected_lora_layers = {k for k in expected_lora_layers if not any(w in k for w in concatenated_weights)} + assert len(model.layers) == len(expected_lora_layers)