Update app.py

app.py

@@ -4,6 +4,7 @@ import spaces
 Copyright NewGenAI
 Code can't be included in commercial app used for monetary gain. No derivative code allowed.
 """
+import gc
 import json
 import torch
 import tqdm
@@ -23,10 +24,6 @@ from huggingface_hub import hf_hub_download
 STATE_FILE = "LTX091_state.json"
 queue = []
 
-torch._inductor.config.conv_1x1_as_mm = True
-torch._inductor.config.coordinate_descent_tuning = True
-torch._inductor.config.epilogue_fusion = False
-torch._inductor.config.coordinate_descent_check_all_directions = True
 
 def load_state():
     if os.path.exists(STATE_FILE):
@@ -144,9 +141,9 @@ except Exception as e:
 
 
 single_file_url = repo_id+"/ltx-video-2b-v0.9.1.safetensors"
-text_encoder = torch.compile(T5EncoderModel.from_pretrained(
+text_encoder = T5EncoderModel.from_pretrained(
   repo_id, subfolder="text_encoder", torch_dtype=torch.bfloat16
-), mode="reduce-overhead", fullgraph=True)
+)
 tokenizer = T5Tokenizer.from_pretrained(
   repo_id, subfolder="tokenizer", torch_dtype=torch.bfloat16
 )
@@ -162,9 +159,6 @@ pipe.to("cuda")
 # pipe.load_lora_weights("TODO/TODO", adapter_name="ltx-lora")
 # pipe.set_adapters(["lrx-lora"], adapter_weights=[1.0])
 
-pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True)
-pipe.vae = torch.compile(pipe.vae.decode, mode="reduce-overhead", fullgraph=True)
-
 @spaces.GPU(duration=120)
 @torch.inference_mode()
 def generate_video(prompt, negative_prompt, height, width, num_frames, num_inference_steps, fps, seed, progress=gr.Progress(track_tqdm=True)):
@@ -187,8 +181,10 @@ def generate_video(prompt, negative_prompt, height, width, num_frames, num_infer
     
     # Randomize seed if seed is 0
     if seed == 0:
-        seed = random.randint(0, 999999)
-    
+        seed = randomize_seed()
+
+    torch.cuda.empty_cache()
+    torch.cuda.synchronize()
     # Generating the video <Does not support seed :( >
     video = pipe(
         prompt=prompt,
@@ -209,7 +205,9 @@ def generate_video(prompt, negative_prompt, height, width, num_frames, num_infer
     os.makedirs("output_LTX091", exist_ok=True)
     output_path = f"./output_LTX091/{filename}"
     export_to_video(video, output_path, fps=fps)
-    
+
+    torch.cuda.empty_cache()
+    gc.collect()
     return output_path
 
 # Gradio UI setup
@@ -223,13 +221,13 @@ with gr.Blocks() as demo:
                 prompt = gr.Textbox(label="Prompt", lines=3, value=initial_state.get("prompt", "A dramatic view of the pyramids at Giza during sunset."))
                 negative_prompt = gr.Textbox(label="Negative Prompt", lines=3, value=initial_state.get("negative_prompt", "worst quality, blurry, distorted"))
             with gr.Row():
-                height = gr.Slider(label="Height", minimum=240, maximum=1080, step=1, value=initial_state.get("height", 480))
-                width = gr.Slider(label="Width", minimum=320, maximum=1920, step=1, value=initial_state.get("width", 704))
+                height = gr.Slider(label="Height", minimum=224, maximum=720, step=1, value=initial_state.get("height", 384))
+                width = gr.Slider(label="Width", minimum=320, maximum=1280, step=1, value=initial_state.get("width", 640))
             with gr.Row():
-                num_frames = gr.Slider(label="Number of Frames", minimum=1, maximum=500, step=1, value=initial_state.get("num_frames", 161))
-                num_inference_steps = gr.Slider(label="Number of Inference Steps", minimum=1, maximum=100, step=1, value=initial_state.get("num_inference_steps", 50))
+                num_frames = gr.Slider(label="Number of Frames", minimum=1, maximum=121, step=1, value=initial_state.get("num_frames", 49))
+                num_inference_steps = gr.Slider(label="Number of Inference Steps", minimum=1, maximum=30, step=1, value=initial_state.get("num_inference_steps", 20))
             with gr.Row():
-                fps = gr.Slider(label="FPS", minimum=1, maximum=60, step=1, value=initial_state.get("fps", 24))
+                fps = gr.Slider(label="FPS", minimum=1, maximum=30, step=1, value=initial_state.get("fps", 16))
                 seed = gr.Number(label="Seed", value=initial_state.get("seed", 0))
                 random_seed_button = gr.Button("Randomize Seed")
 
@@ -237,7 +235,7 @@ with gr.Blocks() as demo:
             generate_button = gr.Button("Generate Video")
             save_state_button = gr.Button("Save State")
 
-            random_seed_button.click(lambda: random.randint(0, 999999), outputs=seed)
+            random_seed_button.click(randomize_seed, outputs=seed)
             generate_button.click(
                 generate_video,
                 inputs=[prompt, negative_prompt, height, width, num_frames, num_inference_steps, fps, seed],
@@ -254,13 +252,13 @@ with gr.Blocks() as demo:
                 batch_prompt = gr.Textbox(label="Prompt", lines=3, value="A batch of videos depicting different landscapes.")
                 batch_negative_prompt = gr.Textbox(label="Negative Prompt", lines=3, value="low quality, inconsistent, jittery")
             with gr.Row():
-                batch_height = gr.Slider(label="Height", minimum=240, maximum=1080, step=1, value=480)
-                batch_width = gr.Slider(label="Width", minimum=320, maximum=1920, step=1, value=704)
+                batch_height = gr.Slider(label="Height", minimum=224, maximum=720, step=1, value=384)
+                batch_width = gr.Slider(label="Width", minimum=320, maximum=1280, step=1, value=640)
             with gr.Row():
-                batch_num_frames = gr.Slider(label="Number of Frames", minimum=1, maximum=500, step=1, value=161)
-                batch_num_inference_steps = gr.Slider(label="Number of Inference Steps", minimum=1, maximum=100, step=1, value=50)
+                batch_num_frames = gr.Slider(label="Number of Frames", minimum=1, maximum=121, step=1, value=49)
+                batch_num_inference_steps = gr.Slider(label="Number of Inference Steps", minimum=1, maximum=30, step=1, value=20)
             with gr.Row():
-                batch_fps = gr.Slider(label="FPS", minimum=1, maximum=60, step=1, value=24)
+                batch_fps = gr.Slider(label="FPS", minimum=1, maximum=30, step=1, value=16)
                 batch_seed = gr.Number(label="Seed", value=0)
                 random_seed_batch_button = gr.Button("Randomize Seed")
 
@@ -270,7 +268,7 @@ with gr.Blocks() as demo:
 
             queue_status = gr.Text(label="Queue Status")
 
-            random_seed_batch_button.click(lambda: random.randint(0, 999999), outputs=batch_seed)
+            random_seed_batch_button.click(randomize_seed, outputs=batch_seed)
             add_to_queue_button.click(
                 add_to_queue,
                 inputs=[batch_prompt, batch_negative_prompt, batch_height, batch_width, batch_num_frames, batch_num_inference_steps, batch_fps, batch_seed],