Update skyreelsinfer/pipelines/pipeline_skyreels_video.py

skyreelsinfer/pipelines/pipeline_skyreels_video.py

@@ -14,7 +14,7 @@ from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video import MultiPipeli
 from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video import PipelineCallback
 from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video import retrieve_timesteps
 from PIL import Image
-
+import gc
 
 def resizecrop(image, th, tw):
     w, h = image.size
@@ -240,7 +240,7 @@ class SkyreelsVideoPipeline(HunyuanVideoPipeline):
             batch_size = len(prompt)
         else:
             batch_size = prompt_embeds.shape[0]
-        pipe.text_encoder.to("cuda")
+        #pipe.text_encoder.to("cuda")
 
         # 3. Encode input prompt
         (
@@ -339,8 +339,8 @@ class SkyreelsVideoPipeline(HunyuanVideoPipeline):
         if hasattr(self, "text_encoder_to_cpu"):
             self.text_encoder_to_cpu()
         pipe.text_encoder.to("cpu")
-        pipe.vae.to("cpu")
-        torch.cuda.empty_cache()
+        #pipe.vae.to("cpu")
+        #torch.cuda.empty_cache()
         
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
@@ -414,7 +414,7 @@ class SkyreelsVideoPipeline(HunyuanVideoPipeline):
                     progress_bar.update()
 
         if not output_type == "latent":
-            pipe.vae.to("cuda")
+        #    pipe.vae.to("cuda")
             latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor
             video = self.vae.decode(latents, return_dict=False)[0]
             video = self.video_processor.postprocess_video(video, output_type=output_type)