Update api/ltx_server_refactored.py

api/ltx_server_refactored.py

@@ -341,13 +341,13 @@ class VideoService:
         try: 
             with torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype, enabled=self.device.type == 'cuda'):
                 latents_bruto = self.pipeline(**first_pass_kwargs).images
-                latents_cpu_bruto = latents_bruto.detach().to("cpu")
-                tensor_path_cpu = os.path.join(results_dir, f"latents_low_res.pt")
-                torch.save(latents_cpu_bruto, tensor_path_cpu)
+                #latents_cpu_bruto = latents_bruto.detach().to("cpu")
+                #tensor_path_cpu = os.path.join(results_dir, f"latents_low_res.pt")
+                #torch.save(latents_cpu_bruto, tensor_path_cpu)
                 log_tensor_info(latents_bruto, f"Latente Bruto Gerado para: '{prompt[:40]}...'")
         
             print("-" * 20 + " FIM: _generate_single_chunk_low " + "-"*20)
-            return latents_cpu_bruto
+            return latents_bruto
         except Exception as e:
             print("-" * 20 + f" ERRO: _generate_single_chunk_low {e} " + "-"*20)
         finally:
@@ -430,34 +430,38 @@ class VideoService:
 
             frames_per_chunk = ((frames_per_chunk//8)*8)+1
             
-            latent_path = self._generate_single_chunk_low(
+            latentes_bruto_r = self._generate_single_chunk_low(
                 prompt=chunk_prompt, negative_prompt=negative_prompt, height=height, width=width,
                 num_frames=frames_per_chunk, guidance_scale=guidance_scale, seed=used_seed + i,
                 initial_latent_condition=condition_item_latent_overlap, image_conditions=current_image_conditions,
                 ltx_configs_override=ltx_configs_override
             )
 
-
-
+            log_tensor_info(latentes_bruto_r, f"latentes_bruto_r recebidk: {i}...'")
+        
             #latent_path_bufer = load_tensor(latent_path)
             #final_latents = torch.cat(lista_tensores, dim=2).to(self.device)
 
         
-            latentes_bruto = latent_path #torch.load(latent_path_bufer).to("cpu")
 
             #poda inicio overlap
-            if i > 0 and poda_latents_num > 0:
-                 latentes_bruto = latentes_bruto[:, :, poda_latents_num:, :, :]
-
-
+            if i > 0 and poda_latents_num > 0 and latentes_bruto_r.shape[2]>poda_latents_num:
+                 latentes_bruto = latentes_bruto_r[:, :, poda_latents_num:, :, :].clone()
+            else: 
+                latentes_bruto = latentes_bruto_r[:, :, :, :, :].clone()
+    
+            log_tensor_info(latentes_bruto, f"latentes_bruto recebidk: {i}...'")
+        
             # cria estado overlap para proximo
             if i < num_chunks - 1 and poda_latents_num > 0:
-                overlap_latents = latentes_bruto[:, :, -poda_latents_num:, :, :].clone()
+                overlap_latents = latentes_bruto_r[:, :, -poda_latents_num:, :, :].clone()
+                log_tensor_info(overlap_latents, f"overlap_latents recebidk: {i}...'")
                 overlap_latents = overlap_latents.detach().to(self.device)
                 condition_item_latent_overlap = ConditioningItem(
                     media_item=overlap_latents, media_frame_number=0, conditioning_strength=1.0
                 )
 
+
             #adiciona a lista
             tensor_path_podado = os.path.join(results_dir, f"latents_poda{i}_res.pt")
             torch.save(latentes_bruto, tensor_path_podado)