Update app.py

app.py

@@ -6,6 +6,7 @@ Code can't be included in commercial app used for monetary gain. No derivative c
 """
 import json
 import torch
+import tqdm
 import gradio as gr
 import random
 import time
@@ -22,6 +23,11 @@ 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):
         with open(STATE_FILE, "r") as file:
@@ -58,11 +64,12 @@ def process_queue():
     if not queue:
         return "Queue is empty."
 
-    for i, task in enumerate(queue):
+    for i, task in tqdm(enumerate(queue)):
         generate_video(**task)
         time.sleep(1)  # Simulate processing time
 
     queue.clear()
+    tqdm.close()
     return "All tasks in the queue have been processed."
 
 def save_ui_state(prompt, negative_prompt, height, width, num_frames, num_inference_steps, fps, seed):
@@ -149,11 +156,39 @@ pipe = LTXPipeline.from_single_file(
     tokenizer=tokenizer, 
     torch_dtype=torch.bfloat16
 )
-pipe.enable_model_cpu_offload()
+pipe.vae.enable_tiling()
+pipe.to("cuda")
+
+# pipe.load_lora_weights("TODO/TODO", adapter_name="ltx-lora")
+# pipe.set_adapters(["lrx-lora"], adapter_weights=[1.0])
+
+pipe.fuse_qkv_projections()
+
+pipe.unet.to(memory_format=torch.channels_last)
+pipe.vae.to(memory_format=torch.channels_last)
+pipe.unet = torch.compile(pipe.unet, mode="max-autotune", fullgraph=True)
+pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True)
 
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=120, progress=gr.Progress(track_tqdm=True))
 @torch.inference_mode()
 def generate_video(prompt, negative_prompt, height, width, num_frames, num_inference_steps, fps, seed):
+    progress_steps = []
+
+    def setup_progressbar_length(_num_steps=num_inference_steps):
+        for _step in bytes(range(_num_steps)):
+            progress_steps.append(_step) # one step one byte - fq the logic
+    
+    def progress_step():
+        if len(progress_steps) == 0:
+            return
+        for done_step in tqdm(enumerate(range(len(progress_steps)))):
+            progress_steps.pop()
+            if len(progress_steps) == 0:        
+                tqdm.close()
+            break
+
+    setup_progressbar_length()
+    
     # Randomize seed if seed is 0
     if seed == 0:
         seed = random.randint(0, 999999)
@@ -167,6 +202,7 @@ def generate_video(prompt, negative_prompt, height, width, num_frames, num_infer
         num_frames=num_frames,
         num_inference_steps=num_inference_steps,
         generator=torch.Generator(device='cuda').manual_seed(seed),
+        callback_on_step_end=progress_step
     ).frames[0]
 
     # Create output filename based on prompt and timestamp