Memory optimizations for Zero GPU

app.py

@@ -155,7 +155,7 @@ input[type="range"] {
 # ============================================
 # VIDEO GENERATION FUNCTION
 # ============================================
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=180)
 def generate_video(
     prompt,
     negative_prompt,
@@ -168,45 +168,62 @@ def generate_video(
     fps,
     progress=gr.Progress(track_tqdm=True)
 ):
+    import gc
+
     if not prompt or prompt.strip() == "":
         gr.Warning("Please enter a prompt to generate a video.")
         return None
 
     progress(0, desc="Loading model...")
 
-    # Load model inside GPU context
+    # Clear memory before loading
+    gc.collect()
+    torch.cuda.empty_cache() if torch.cuda.is_available() else None
+
+    # Load model with memory optimizations
     pipe = LTXPipeline.from_pretrained(
         "Lightricks/LTX-Video",
-        torch_dtype=torch.bfloat16
+        torch_dtype=torch.float16,
+        low_cpu_mem_usage=True,
+        use_safetensors=True,
     )
-    pipe.to("cuda")
-    pipe.enable_model_cpu_offload()
+
+    # Use CPU offload for memory efficiency (don't call .to("cuda") with this)
+    pipe.enable_sequential_cpu_offload()
+    pipe.vae.enable_slicing()
+    pipe.vae.enable_tiling()
 
     # Handle seed
     if seed == -1:
         seed = random.randint(0, 2**32 - 1)
 
-    generator = torch.Generator(device="cuda").manual_seed(int(seed))
+    generator = torch.Generator(device="cpu").manual_seed(int(seed))
 
     progress(0.1, desc="Generating video frames...")
 
-    # Generate video
-    output = pipe(
-        prompt=prompt,
-        negative_prompt=negative_prompt if negative_prompt else None,
-        num_frames=int(num_frames),
-        guidance_scale=float(guidance_scale),
-        num_inference_steps=int(num_inference_steps),
-        generator=generator,
-        height=int(height),
-        width=int(width),
-    )
+    # Generate video with memory-efficient settings
+    with torch.inference_mode():
+        output = pipe(
+            prompt=prompt,
+            negative_prompt=negative_prompt if negative_prompt else None,
+            num_frames=int(num_frames),
+            guidance_scale=float(guidance_scale),
+            num_inference_steps=int(num_inference_steps),
+            generator=generator,
+            height=int(height),
+            width=int(width),
+        )
 
     progress(0.9, desc="Exporting video...")
 
     # Export to video file
     video_path = export_to_video(output.frames[0], fps=int(fps))
 
+    # Cleanup
+    del pipe
+    gc.collect()
+    torch.cuda.empty_cache() if torch.cuda.is_available() else None
+
     progress(1.0, desc="Complete!")
 
     return video_path
@@ -285,8 +302,8 @@ def create_ui():
                             num_frames = gr.Slider(
                                 label="Frames",
                                 minimum=9,
-                                maximum=97,
-                                value=41,
+                                maximum=57,
+                                value=25,
                                 step=8,
                                 info="More frames = longer video"
                             )
@@ -305,8 +322,8 @@ def create_ui():
                             num_inference_steps = gr.Slider(
                                 label="Inference Steps",
                                 minimum=10,
-                                maximum=50,
-                                value=30,
+                                maximum=40,
+                                value=20,
                                 step=5,
                                 info="More steps = higher quality"
                             )