🐛 [Bug] WAN2.1 graph break with Dynamic shapes

Bug Description

While compiling WAN2.1 with dynamic shapes, I found that there is graph break. Such operators are not supported:

• torch.ops.aten.empty.memory_format + Operator Count: 2
• torch.ops.aten.scatter.src + Operator Count: 4

You can use this script to reproduce:

pip install git+https://github.com/huggingface/diffusers.git

import os
import numpy as np
import torch
import torch_tensorrt as torch_trt
from diffusers.utils import export_to_video
from diffusers import AutoencoderKLWan, WanPipeline
from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
from diffusers.models.transformers.transformer_wan import WanTransformer3DModel
import argparse
import sys


def main(args):
        

    model_name = f"wan21_frames{args.num_frames}"


    dtype = torch.float16

    settings = {
        # "use_python_runtime": True,
        "enabled_precisions": {torch.float16},
        "immutable_weights": False,
        "min_block_size": 1,
        "offload_module_to_cpu": True,
        "optimization_level": 3,
    }





    # Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers
    model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
    vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float16)
    flow_shift = 5.0 # 5.0 for 720P, 3.0 for 480P
    scheduler = UniPCMultistepScheduler(prediction_type='flow_prediction', use_flow_sigmas=True, num_train_timesteps=1000, flow_shift=flow_shift)
    pipe = WanPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.float16, transformer=WanTransformer3DModel(num_layers=1))
    pipe.scheduler = scheduler
    pipe.to(dtype).to("cuda")

    prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
    negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"

    model = pipe.transformer

    FRAMES = torch.export.Dim("num_frames", min=1, max=11)
    dynamic_shapes = {
        "hidden_states": {2: FRAMES},
        "timestep": {},
        "encoder_hidden_states": {},
        "return_dict": None,
    }


    with torch_trt.dynamo.Debugger(log_level="debug", 
                                    profile_format="trex", 
                                    save_engine_profile=True, 
                                    # capture_fx_graph_before=["remove_detach"], 
                                    # capture_fx_graph_after=["remove_num_users_is_0_nodes"],
                                    logging_dir="/home/profile/logging/torchtrt", engine_builder_monitor=False,):
        trt_mod = torch_trt.MutableTorchTensorRTModule(model, **settings)
        trt_mod.set_expected_dynamic_shape_range((), dynamic_shapes)
        pipe.transformer = trt_mod
        output = pipe(
            prompt=prompt,
            negative_prompt=negative_prompt,
            height=720,
            width=1280,
            num_frames=args.num_frames,
            guidance_scale=5.0,
        ).frames[0]