Rename managers/ltx_manager_helpers.py to managers/ltx_manager.py

managers/{ltx_manager_helpers.py → ltx_manager.py}

@@ -1,4 +1,4 @@
-# ltx_manager_helpers.py
+# managers/ltx_manager.py
 #
 # Copyright (C) August 4, 2025  Carlos Rodrigues dos Santos
 #
@@ -25,6 +25,7 @@ from inference import create_ltx_video_pipeline, calculate_padding
 # We need these types for our patch
 from ltx_video.pipelines.pipeline_ltx_video import LTXVideoPipeline, ConditioningItem, LatentConditioningItem
 from ltx_video.models.autoencoders.vae_encode import vae_encode, latent_to_pixel_coords
+from ltx_video.pipelines.pipeline_ltx_video import LTXMultiScalePipeline
 from diffusers.utils.torch_utils import randn_tensor
 
 logger = logging.getLogger(__name__)
@@ -70,8 +71,7 @@ def _aduc_prepare_conditioning_patch(
             media_item_latents = item.latent_tensor.to(dtype=init_latents.dtype, device=init_latents.device)
             media_frame_number = item.media_frame_number
             strength = item.conditioning_strength
-            n_latent_frames = media_item_latents.shape[2]
-
+            
             if media_frame_number == 0:
                 f_l, h_l, w_l = media_item_latents.shape[-3:]
                 init_latents[:, :, :f_l, :h_l, :w_l] = torch.lerp(init_latents[:, :, :f_l, :h_l, :w_l], media_item_latents, strength)
@@ -87,7 +87,7 @@ def _aduc_prepare_conditioning_patch(
                 extra_conditioning_latents.append(patched_latents)
                 extra_conditioning_pixel_coords.append(pixel_coords)
                 extra_conditioning_mask.append(new_mask)
-    else: # Original pixel-based logic
+    else: # Original pixel-based logic for fallback
         for item in conditioning_items:
             if not isinstance(item, ConditioningItem): continue
             item = self._resize_conditioning_item(item, height, width)
@@ -97,15 +97,14 @@ def _aduc_prepare_conditioning_patch(
             ).to(dtype=init_latents.dtype)
             media_frame_number = item.media_frame_number
             strength = item.conditioning_strength
-            n_pixel_frames = item.media_item.shape[2]
             if media_frame_number == 0:
                 media_item_latents, l_x, l_y = self._get_latent_spatial_position(media_item_latents, item, height, width, strip_latent_border=True)
                 f_l, h_l, w_l = media_item_latents.shape[-3:]
                 init_latents[:, :, :f_l, l_y:l_y+h_l, l_x:l_x+w_l] = torch.lerp(init_latents[:, :, :f_l, l_y:l_y+h_l, l_x:l_x+w_l], media_item_latents, strength)
                 init_conditioning_mask[:, :f_l, l_y:l_y+h_l, l_x:l_x+w_l] = strength
             else:
-                # ... (this part of the original logic can be included if needed) ...
-                pass # For ADUC, we primarily use latent anchors for non-zero frames.
+                logger.warning("Pixel-based conditioning for non-zero frames is not fully implemented in this patch.")
+                pass
 
     init_latents, init_latent_coords = self.patchifier.patchify(latents=init_latents)
     init_pixel_coords = latent_to_pixel_coords(init_latent_coords, self.vae, causal_fix=self.transformer.config.causal_temporal_positioning)
@@ -194,7 +193,6 @@ class LtxPoolManager:
         self.current_worker_index = 0
         self.lock = threading.Lock()
 
-        # <<< NEW: APPLY PATCH AFTER INITIALIZATION >>>
         self._apply_ltx_pipeline_patches()
 
         if all(w.device.type == 'cuda' for w in self.workers):
@@ -208,12 +206,9 @@ class LtxPoolManager:
     def _apply_ltx_pipeline_patches(self):
         """
         Applies runtime patches to the LTX pipeline for ADUC-SDR compatibility.
-        This is where the monkey patching happens.
         """
         logger.info("LTX POOL MANAGER: Applying ADUC-SDR patches to LTX pipeline...")
         for worker in self.workers:
-            # The __get__ method binds our standalone function to the pipeline instance,
-            # making it behave like a regular method (so 'self' works correctly).
             worker.pipeline.prepare_conditioning = _aduc_prepare_conditioning_patch.__get__(worker.pipeline, LTXVideoPipeline)
         logger.info("LTX POOL MANAGER: All pipeline instances have been patched successfully.")
 
@@ -224,7 +219,6 @@ class LtxPoolManager:
             return worker
     
     def _prepare_pipeline_params(self, worker: LtxWorker, **kwargs) -> dict:
-        # This function remains unchanged
         pipeline_params = {
             "height": kwargs['height'], "width": kwargs['width'], "num_frames": kwargs['video_total_frames'],
             "frame_rate": kwargs.get('video_fps', 24),
@@ -254,7 +248,6 @@ class LtxPoolManager:
         return pipeline_params
 
     def generate_latent_fragment(self, **kwargs) -> (torch.Tensor, tuple):
-        # This function remains unchanged
         worker_to_use = self._get_next_worker()
         try:
             height, width = kwargs['height'], kwargs['width']
@@ -277,7 +270,6 @@ class LtxPoolManager:
                     gc.collect(); torch.cuda.empty_cache()
 
     def refine_latents(self, latents_to_refine: torch.Tensor, **kwargs) -> (torch.Tensor, tuple):
-        # This function remains unchanged
         worker_to_use = self._get_next_worker()
         try:
             _b, _c, _f, latent_h, latent_w = latents_to_refine.shape