Fix default duration + remove timings

app.py

@@ -40,7 +40,10 @@ MAX_SEED = np.iinfo(np.int32).max
 
 FIXED_FPS = 24
 MIN_FRAMES_MODEL = 8
-MAX_FRAMES_MODEL = 81 
+MAX_FRAMES_MODEL = 81
+
+MIN_DURATION = round(MIN_FRAMES_MODEL/FIXED_FPS,1)
+MAX_DURATION = round(MAX_FRAMES_MODEL/FIXED_FPS,1)
 
 optimize_pipeline_(pipe,
     image=Image.new('RGB', (LANDSCAPE_WIDTH, LANDSCAPE_HEIGHT)),
@@ -99,7 +102,7 @@ def generate_video(
     input_image,
     prompt,
     negative_prompt=default_negative_prompt,
-    duration_seconds = 2,
+    duration_seconds = MAX_DURATION,
     guidance_scale = 1,
     steps = 4,
     seed = 42,
@@ -178,7 +181,7 @@ with gr.Blocks() as demo:
         with gr.Column():
             input_image_component = gr.Image(type="pil", label="Input Image (auto-resized to target H/W)")
             prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
-            duration_seconds_input = gr.Slider(minimum=round(MIN_FRAMES_MODEL/FIXED_FPS,1), maximum=round(MAX_FRAMES_MODEL/FIXED_FPS,1), step=0.1, value=round(MAX_FRAMES_MODEL/FIXED_FPS,1), label="Duration (seconds)", info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps.")
+            duration_seconds_input = gr.Slider(minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=MAX_DURATION, label="Duration (seconds)", info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps.")
             
             with gr.Accordion("Advanced Settings", open=False):
                 negative_prompt_input = gr.Textbox(label="Negative Prompt", value=default_negative_prompt, lines=3)

optimization.py

@@ -1,7 +1,6 @@
 """
 """
 
-from datetime import datetime
 from typing import Any
 from typing import Callable
 from typing import ParamSpec
@@ -39,25 +38,16 @@ INDUCTOR_CONFIGS = {
 
 def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
 
-    t0 = datetime.now()
-
     @spaces.GPU(duration=1500)
     def compile_transformer():
-        nonlocal t0
-        
-        print('compile_transformer', -(t0 - (t0 := datetime.now())))
 
         with capture_component_call(pipeline, 'transformer') as call:
             pipeline(*args, **kwargs)
-
-        print('capture_component_call', -(t0 - (t0 := datetime.now())))
         
         dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
         dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
 
         quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
-
-        print('quantize_', -(t0 - (t0 := datetime.now())))
         
         hidden_states: torch.Tensor = call.kwargs['hidden_states']
         hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
@@ -74,8 +64,6 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
             kwargs=call.kwargs | {'hidden_states': hidden_states_landscape},
             dynamic_shapes=dynamic_shapes,
         )
-
-        print('exported_landscape', -(t0 - (t0 := datetime.now())))
         
         exported_portrait = torch.export.export(
             mod=pipeline.transformer,
@@ -84,23 +72,13 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
             dynamic_shapes=dynamic_shapes,
         )
 
-        print('exported_portrait', -(t0 - (t0 := datetime.now())))
-
         compiled_landscape = aoti_compile(exported_landscape, INDUCTOR_CONFIGS)
-        print('compiled_landscape', -(t0 - (t0 := datetime.now())))
-
         compiled_portrait = aoti_compile(exported_portrait, INDUCTOR_CONFIGS)
-        print('compiled_portrait', -(t0 - (t0 := datetime.now())))
-
-        # Avoid weights duplication when serializing back to main process
-        compiled_portrait.weights = compiled_landscape.weights
+        compiled_portrait.weights = compiled_landscape.weights # Avoid weights duplication when serializing back to main process
 
         return compiled_landscape, compiled_portrait
 
     compiled_landscape, compiled_portrait = compile_transformer()
-    print('compiled', -(t0 - (t0 := datetime.now())))
-
-    compiled_portrait.weights = compiled_landscape.weights
 
     def combined_transformer(*args, **kwargs):
         hidden_states: torch.Tensor = kwargs['hidden_states']