finetrainers is broken, reverting

finetrainers/patches/__init__.py

@@ -17,12 +17,7 @@ def perform_patches_for_training(args: "BaseArgs", parallel_backend: "ParallelBa
         if parallel_backend.tensor_parallel_enabled:
             patch.patch_apply_rotary_emb_for_tp_compatibility()
 
-    if args.model_name == ModelType.WAN:
-        from .models.wan import patch
-
-        patch.patch_time_text_image_embedding_forward()
-
     if args.training_type == TrainingType.LORA and len(args.layerwise_upcasting_modules) > 0:
-        from .dependencies.peft import patch
+        from dependencies.peft import patch
 
         patch.patch_peft_move_adapter_to_device_of_base_layer()

finetrainers/patches/models/ltx_video/patch.py

@@ -16,7 +16,7 @@ def patch_apply_rotary_emb_for_tp_compatibility() -> None:
 
 
 def _perform_ltx_transformer_forward_patch() -> None:
-    LTXVideoTransformer3DModel.forward = _patched_LTXVideoTransformer3D_forward
+    LTXVideoTransformer3DModel.forward = _patched_LTXVideoTransformer3Dforward
 
 
 def _perform_ltx_apply_rotary_emb_tensor_parallel_compatibility_patch() -> None:
@@ -35,7 +35,7 @@ def _perform_ltx_apply_rotary_emb_tensor_parallel_compatibility_patch() -> None:
     diffusers.models.transformers.transformer_ltx.apply_rotary_emb = apply_rotary_emb
 
 
-def _patched_LTXVideoTransformer3D_forward(
+def _patched_LTXVideoTransformer3Dforward(
     self,
     hidden_states: torch.Tensor,
     encoder_hidden_states: torch.Tensor,

finetrainers/patches/models/wan/patch.py

@@ -1,33 +0,0 @@
-from typing import Optional
-
-import diffusers
-import torch
-
-
-def patch_time_text_image_embedding_forward() -> None:
-    _patch_time_text_image_embedding_forward()
-
-
-def _patch_time_text_image_embedding_forward() -> None:
-    diffusers.models.transformers.transformer_wan.WanTimeTextImageEmbedding.forward = (
-        _patched_WanTimeTextImageEmbedding_forward
-    )
-
-
-def _patched_WanTimeTextImageEmbedding_forward(
-    self,
-    timestep: torch.Tensor,
-    encoder_hidden_states: torch.Tensor,
-    encoder_hidden_states_image: Optional[torch.Tensor] = None,
-):
-    # Some code has been removed compared to original implementation in Diffusers
-    # Also, timestep is typed as that of encoder_hidden_states
-    timestep = self.timesteps_proj(timestep).type_as(encoder_hidden_states)
-    temb = self.time_embedder(timestep).type_as(encoder_hidden_states)
-    timestep_proj = self.time_proj(self.act_fn(temb))
-
-    encoder_hidden_states = self.text_embedder(encoder_hidden_states)
-    if encoder_hidden_states_image is not None:
-        encoder_hidden_states_image = self.image_embedder(encoder_hidden_states_image)
-
-    return temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image

finetrainers/trainer/sft_trainer/trainer.py

@@ -334,7 +334,6 @@ class SFTTrainer:
         parallel_backend = self.state.parallel_backend
         train_state = self.state.train_state
         device = parallel_backend.device
-        dtype = self.args.transformer_dtype
 
         memory_statistics = utils.get_memory_statistics()
         logger.info(f"Memory before training start: {json.dumps(memory_statistics, indent=4)}")
@@ -448,8 +447,8 @@ class SFTTrainer:
 
             logger.debug(f"Starting training step ({train_state.step}/{self.args.train_steps})")
 
-            latent_model_conditions = utils.align_device_and_dtype(latent_model_conditions, device, dtype)
-            condition_model_conditions = utils.align_device_and_dtype(condition_model_conditions, device, dtype)
+            utils.align_device_and_dtype(latent_model_conditions, device, self.args.transformer_dtype)
+            utils.align_device_and_dtype(condition_model_conditions, device, self.args.transformer_dtype)
             latent_model_conditions = utils.make_contiguous(latent_model_conditions)
             condition_model_conditions = utils.make_contiguous(condition_model_conditions)
 

requirements.txt

@@ -19,7 +19,8 @@ eva-decord==0.6.1
 wandb
 pandas
 sentencepiece>=0.2.0
-imageio-ffmpeg>=0.5.1
+imageio
+imageio-ffmpeg
 torchdata==0.11.0
 
 flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

requirements_without_flash_attention.txt

@@ -20,7 +20,8 @@ eva-decord==0.6.1
 wandb
 pandas
 sentencepiece>=0.2.0
-imageio-ffmpeg>=0.5.1
+imageio
+imageio-ffmpeg
 torchdata==0.11.0
 
 # for youtube video download

vms/ui/project/tabs/preview_tab.py

@@ -33,6 +33,25 @@ class PreviewTab(BaseTab):
             
             with gr.Row():
                 with gr.Column(scale=2):
+                                        
+                    # Add dropdown to choose between LoRA and original model
+                    has_lora = self.check_lora_model_exists()
+                    lora_choices = []
+                    default_lora_choice = ""
+
+                    if has_lora:
+                        lora_choices = ["Use LoRA model", "Use original model"]
+                        default_lora_choice = "Use LoRA model"
+                    else:
+                        lora_choices = ["Cannot find LoRA model", "Use original model"]
+                        default_lora_choice = "Use original model"
+
+                    self.components["use_lora"] = gr.Dropdown(
+                        choices=lora_choices,
+                        label="Model Selection",
+                        value=default_lora_choice
+                    )
+                    
                     self.components["prompt"] = gr.Textbox(
                         label="Prompt",
                         placeholder="Enter your prompt here...",
@@ -82,25 +101,7 @@ class PreviewTab(BaseTab):
                                 choices=self.get_model_version_choices(default_model),
                                 value=self.get_default_model_version(default_model)
                             )
-                    
-                    # Add dropdown to choose between LoRA and original model
-                    has_lora = self.check_lora_model_exists()
-                    lora_choices = []
-                    default_lora_choice = ""
-
-                    if has_lora:
-                        lora_choices = ["Use LoRA model", "Use original model"]
-                        default_lora_choice = "Use LoRA model"
-                    else:
-                        lora_choices = ["Cannot find LoRA model", "Use original model"]
-                        default_lora_choice = "Use original model"
 
-                    self.components["use_lora"] = gr.Dropdown(
-                        choices=lora_choices,
-                        label="Model Selection",
-                        value=default_lora_choice
-                    )
-                    
                     # Add image input for image-to-video models
                     self.components["conditioning_image"] = gr.Image(
                         label="Conditioning Image (for Image-to-Video models)",