Update optimization.py

optimization.py

@@ -40,57 +40,128 @@ INDUCTOR_CONFIGS = {
 
 def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
 
- @spaces.GPU(duration=1500)
- def compile_transformer():
-    # --- LoRA 1: lightning (loads into default transformer) ---
-    pipeline.load_lora_weights(
-        "Kijai/WanVideo_comfy",
-        weight_name="Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors",
-        adapter_name="lightning",
-    )
-
-    # --- LoRA 2: lightning_2 (loads into transformer_2) ---
-    kwargs_lora = {"load_into_transformer_2": True}
-    pipeline.load_lora_weights(
-        "deadman44/Wan2.2_T2i_T2v_LoRA",
-        weight_name="lora_wan2.2_myjd_Low_v01.safetensors",
-        adapter_name="lightning_2",
-        **kwargs_lora,
-    )
-
-    # --- LoRA 3: orbit_shot (the ostris repo you asked for) ---
-    # Load into transformer_2 as well (set load_into_transformer_2 True if this adapter targets transformer_2)
-    # pipeline.load_lora_weights(
-    #     "ostris/wan22_i2v_14b_orbit_shot_lora",
-    #     weight_name="wan22_14b_i2v_orbit_low_noise.safetensors",
-    #     adapter_name="orbit_shot",
-    #     **kwargs_lora,
-    # )
-
-    # Register adapters and their relative weights
-    # (adjust adapter_weights to taste; here each is weight 1.0)
-    #pipeline.set_adapters(["lightning", "lightning_2", "orbit_shot"], adapter_weights=[1.0, 1.0, 1.0])
-    pipeline.set_adapters(["lightning", "lightning_2", "orbit_shot"], adapter_weights=[1., 1.])
-
-
-    # Fuse each adapter into the correct component with chosen lora_scale:
-    # - lightning -> transformer (boosted by 3x in your original)
-    # - lightning_2 -> transformer_2
-    # - orbit_shot -> transformer_2 (or transformer depending on the LoRA design)
-    pipeline.fuse_lora(adapter_names=["lightning"], lora_scale=3.0, components=["transformer"])
-    pipeline.fuse_lora(adapter_names=["lightning_2"], lora_scale=1.0, components=["transformer_2"])
-    #pipeline.fuse_lora(adapter_names=["orbit_shot"], lora_scale=1.0, components=["transformer_2"])
-
-    # After fusing, you can unload LoRA weights to free memory (fused weights remain applied)
-    pipeline.unload_lora_weights()
-
-    # --- then continue with capture_component_call / export / compile logic as you already have ---
-    with capture_component_call(pipeline, 'transformer') as call:
-        pipeline(*args, **kwargs)
-
-    # ... rest of your function unchanged ...
- 
-   
+    @spaces.GPU(duration=1500)
+    def compile_transformer():
+        
+        # pipeline.load_lora_weights(
+        #     "Kijai/WanVideo_comfy", 
+        #     weight_name="Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors", 
+        #     adapter_name="lightning"
+        # )
+        # kwargs_lora = {}
+        # kwargs_lora["load_into_transformer_2"] = True
+        # pipeline.load_lora_weights(
+        #     "Kijai/WanVideo_comfy", 
+        #      weight_name="Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors", 
+        #     #weight_name="Wan22-Lightning/Wan2.2-Lightning_T2V-A14B-4steps-lora_LOW_fp16.safetensors", 
+        #     adapter_name="lightning_2", **kwargs_lora
+        # )
+        # pipeline.set_adapters(["lightning", "lightning_2"], adapter_weights=[1., 1.])
+
+
+        # pipeline.fuse_lora(adapter_names=["lightning"], lora_scale=3., components=["transformer"])
+        # pipeline.fuse_lora(adapter_names=["lightning_2"], lora_scale=1., components=["transformer_2"])
+        # pipeline.unload_lora_weights()
+
+
+
+        pipeline.load_lora_weights(
+            "Kijai/WanVideo_comfy", 
+            weight_name="Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors", 
+            adapter_name="lightning"
+        )
+        kwargs_lora = {}
+        kwargs_lora["load_into_transformer_2"] = True
+        # pipeline.load_lora_weights(
+        #     #"drozbay/Wan2.2_A14B_lora_extract", 
+        #      "Kijai/WanVideo_comfy", 
+        #     #weight_name="MTVCrafter/Wan2_1_MTV-Crafter_motion_adapter_bf16.safetensors", 
+        #      #weight_name="Skyreels/Wan2_1_Skyreels-v2-T2V-720P_LoRA_rank_64_fp16.safetensors",
+        #     #weight_name="Pusa/Wan21_PusaV1_LoRA_14B_rank512_bf16.safetensors", 
+        #     weight_name="Wan22-Lightning/Wan2.2-Lightning_T2V-A14B-4steps-lora_LOW_fp16.safetensors", 
+        #     adapter_name="lightning_2", **kwargs_lora
+        # )
+
+
+
+        pipeline.load_lora_weights(
+            #"drozbay/Wan2.2_A14B_lora_extract", 
+             "deadman44/Wan2.2_T2i_T2v_LoRA", 
+            #weight_name="MTVCrafter/Wan2_1_MTV-Crafter_motion_adapter_bf16.safetensors", 
+             #weight_name="Skyreels/Wan2_1_Skyreels-v2-T2V-720P_LoRA_rank_64_fp16.safetensors",
+            weight_name="lora_wan2.2_myjd_Low_v01.safetensors", 
+            #weight_name="Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1.1/low_noise_model.safetensors", 
+            adapter_name="lightning_2", **kwargs_lora
+        )
+
+        # pipeline.load_lora_weights(
+        #     #"drozbay/Wan2.2_A14B_lora_extract", 
+        #      "ostris/wan22_i2v_14b_orbit_shot_lora", 
+        #     #weight_name="MTVCrafter/Wan2_1_MTV-Crafter_motion_adapter_bf16.safetensors", 
+        #      #weight_name="Skyreels/Wan2_1_Skyreels-v2-T2V-720P_LoRA_rank_64_fp16.safetensors",
+        #     weight_name="wan22_14b_i2v_orbit_low_noise.safetensors", 
+        #     #weight_name="Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1.1/low_noise_model.safetensors", 
+        #     adapter_name="lightning_200", **kwargs_lora
+        # )
+
+
+
+
+
+
+        pipeline.set_adapters(["lightning", "lightning_2"], adapter_weights=[1., 1.])
+
+
+        pipeline.fuse_lora(adapter_names=["lightning"], lora_scale=3., components=["transformer"])
+        pipeline.fuse_lora(adapter_names=["lightning_2"], lora_scale=1., components=["transformer_2"])
+       
+        pipeline.unload_lora_weights()
+        
+        with capture_component_call(pipeline, 'transformer') as call:
+            pipeline(*args, **kwargs)
+        
+        dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
+        dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
+
+        quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
+        quantize_(pipeline.transformer_2, Float8DynamicActivationFloat8WeightConfig())
+        
+        hidden_states: torch.Tensor = call.kwargs['hidden_states']
+        hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
+        if hidden_states.shape[-1] > hidden_states.shape[-2]:
+            hidden_states_landscape = hidden_states
+            hidden_states_portrait = hidden_states_transposed
+        else:
+            hidden_states_landscape = hidden_states_transposed
+            hidden_states_portrait = hidden_states
+
+        exported_landscape_1 = torch.export.export(
+            mod=pipeline.transformer,
+            args=call.args,
+            kwargs=call.kwargs | {'hidden_states': hidden_states_landscape},
+            dynamic_shapes=dynamic_shapes,
+        )
+        
+        exported_portrait_2 = torch.export.export(
+            mod=pipeline.transformer_2,
+            args=call.args,
+            kwargs=call.kwargs | {'hidden_states': hidden_states_portrait},
+            dynamic_shapes=dynamic_shapes,
+        )
+
+        compiled_landscape_1 = aoti_compile(exported_landscape_1, INDUCTOR_CONFIGS)
+        compiled_portrait_2 = aoti_compile(exported_portrait_2, INDUCTOR_CONFIGS)
+
+        compiled_landscape_2 = ZeroGPUCompiledModel(compiled_landscape_1.archive_file, compiled_portrait_2.weights)
+        compiled_portrait_1 = ZeroGPUCompiledModel(compiled_portrait_2.archive_file, compiled_landscape_1.weights)
+
+        return (
+            compiled_landscape_1,
+            compiled_landscape_2,
+            compiled_portrait_1,
+            compiled_portrait_2,
+        )
+
     quantize_(pipeline.text_encoder, Int8WeightOnlyConfig())
     cl1, cl2, cp1, cp2 = compile_transformer()