up

check.py

@@ -50,10 +50,12 @@ optimize_pipeline_(
     pipe,
     conditions=[condition1],
     prompt="prompt",
+    negative_prompt="prompt",
     height=LANDSCAPE_HEIGHT,
     width=LANDSCAPE_WIDTH,
     num_frames=MAX_FRAMES_MODEL,
-    num_inference_steps=2
+    num_inference_steps=2,
+    guidance_scale=1.0,
 )
 
 default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
@@ -79,4 +81,5 @@ output_frames_list = pipe(
     guidance_scale=float(guidance_scale),
     num_inference_steps=int(steps),
     generator=torch.Generator(device="cuda").manual_seed(current_seed),
-).frames[0]
+).frames[0]
+export_to_video(output_frames_list, "output_original.mp4", fps=24)

optimization.py

@@ -19,20 +19,17 @@ P = ParamSpec("P")
 
 # Sequence packing in LTX is a bit of a pain.
 # See: https://github.com/huggingface/diffusers/blob/c052791b5fe29ce8a308bf63dda97aa205b729be/src/diffusers/pipelines/ltx/pipeline_ltx.py#L420
-# TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim("seq_len", min=4680, max=4680)
-
-# Unused currently as I don't know how to make the best use of it for LTX.
-# TRANSFORMER_DYNAMIC_SHAPES = {
-#     "hidden_states": {1: TRANSFORMER_NUM_FRAMES_DIM},
-# }
+TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim.AUTO
+TRANSFORMER_DYNAMIC_SHAPES = {
+    "hidden_states": {1: TRANSFORMER_NUM_FRAMES_DIM},
+}
 
 INDUCTOR_CONFIGS = {
     "conv_1x1_as_mm": True,
     "epilogue_fusion": False,
     "coordinate_descent_tuning": True,
     "coordinate_descent_check_all_directions": True,
-    # "max_autotune": True,
-    "max_autotune": False,
+    "max_autotune": False, # doesn't help much
     "triton.cudagraphs": True,
 }
 TRANSFORMER_SPATIAL_PATCH_SIZE = 1
@@ -54,8 +51,8 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
 
     @spaces.GPU(duration=1500)
     def compile_transformer():
-        # dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
-        # dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
+        dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
+        dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
 
         quantize_(pipeline.transformer, float8_dynamic_activation_float8_weight())
 
@@ -87,13 +84,13 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
             mod=pipeline.transformer,
             args=call.args,
             kwargs=call.kwargs | {"hidden_states": hidden_states_landscape},
-            # dynamic_shapes=dynamic_shapes,
+            dynamic_shapes=dynamic_shapes,
         )
         exported_portrait = torch.export.export(
             mod=pipeline.transformer,
             args=call.args,
             kwargs=call.kwargs | {"hidden_states": hidden_states_portrait},
-            # dynamic_shapes=dynamic_shapes,
+            dynamic_shapes=dynamic_shapes,
         )
 
         compiled_landscape = aoti_compile(exported_landscape, INDUCTOR_CONFIGS)
@@ -129,7 +126,8 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
     with torch.no_grad():
         combined_transformer(**call.kwargs)
 
-    pipeline.transformer = cudagraph(combined_transformer)
+    pipeline.transformer = combined_transformer
+    # pipeline.transformer = cudagraph(combined_transformer)
 
     with torch.no_grad():
         pipeline.transformer(**call.kwargs)