Update api/ltx_server.py

api/ltx_server.py

@@ -588,14 +588,31 @@ class VideoService:
                 second_pass_args = self.config.get("second_pass", {}).copy()
                 second_pass_args["guidance_scale"] = float(guidance_scale)
 
-                multi_scale_call_kwargs = call_kwargs.copy()
-                multi_scale_call_kwargs.update(
-                    {
-                        "downscale_factor": self.config["downscale_factor"],
-                        "first_pass": first_pass_args,
-                        "second_pass": second_pass_args,
-                    }
-                )
+                #multi_scale_call_kwargs = call_kwargs.copy()
+                #multi_scale_call_kwargs.update(
+                #    {
+                #        "downscale_factor": self.config["downscale_factor"],
+                #        "first_pass": first_pass_args,
+                #        "second_pass": second_pass_args,
+                #    }
+                #)
+
+                first_pass_kwargs = call_kwargs.copy()
+                first_pass_kwargs.update(first_pass_args)
+                
+                with ctx:
+                    _apply_precision_policyresult_first = self.pipeline(**first_pass_kwargs)
+                
+                latents_first = result_first.latents
+
+                with ctx:
+                    result_second = self.latent_upsampler(
+                        latents=latents_first,
+                        **second_pass_args
+                    )
+                latents_final = result_second.latents if hasattr(result_second, "latents") else result_second
+
+                
                 print("[DEBUG] Chamando multi_scale_pipeline...")
                 t_ms = time.perf_counter()
                 ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()