🐛 [Bug] BUGS encountered when using Dynamic shapes

Bug Description

I am trying to speed up inference on huggingface Chinese BERT (https://huggingface.co/bert-base-chinese), and I need to do dynamic dimension inference acceleration.But errors in the dynamic dimension confuse me

To Reproduce

When I use the Tensorrt API, I can achieve dynamic dimension acceleration normally. The key code is as follows

profile.set_shape(network.get_input(0).name, (1, 1), (16, 64), (32, 128))
profile.set_shape(network.get_input(1).name, (1, 1), (16, 64), (32, 128))

But when I use torch_ tensorrt. compile, i got an error message. the key code is as follows

trt_model = torch_tensorrt.compile(
    traced_mlm_model, 
    inputs= (torch_tensorrt.Input(min_shape=[1, 1],opt_shape=[16, 64],max_shape=[32, 128],dtype=torch.int32),  # input_ids
             torch_tensorrt.Input(min_shape=[1, 1],opt_shape=[16, 64],max_shape=[32, 128],dtype=torch.int32)),  # attention_mask
    enabled_precisions= {torch.float32}, # Run with 32-bit precision
    workspace_size=2000000000,
    truncate_long_and_double=True,
    debug=True
)

the error message is as follows

ERROR: [Torch-TensorRT TorchScript Conversion Context] - 4: [shapeContext.cpp::operator()::3602] Error Code 4: Shape Error (reshape dimension with more than one -1 wildcard. Reshaping [(+ (MAX 0 (+ (# 0 (SHAPE input_1)) -1)) 1),(+ (MAX 0 (+ (# 1 (SHAPE input_1)) -1)) 1)] to [-1,1,-1].)
ERROR: [Torch-TensorRT TorchScript Conversion Context] - 4: [graphShapeAnalyzer.cpp::needTypeAndDimensions::2212] Error Code 4: Internal Error ((Unnamed Layer* 31) [Shuffle]: output shape can not be computed)
ERROR: [Torch-TensorRT TorchScript Conversion Context] - 4: [graphShapeAnalyzer.cpp::needTypeAndDimensions::2212] Error Code 4: Internal Error ((Unnamed Layer* 31) [Shuffle]: output shape can not be computed)

Traceback (most recent call last):
File "xxx/recall_tensorrt/speedup.py", line 61, in
trt_model = torch_tensorrt.compile(
File "xxx/.conda/envs/trtinfer/lib/python3.8/site-packages/torch_tensorrt/_compile.py", line 133, in compile
return torch_tensorrt.ts.compile(
File "xxx/.conda/envs/trtinfer/lib/python3.8/site-packages/torch_tensorrt/ts/_compiler.py", line 139, in compile
compiled_cpp_mod = _C.compile_graph(module._c, _parse_compile_spec(spec))
RuntimeError: [Error thrown at core/conversion/converters/impl/unsqueeze.cpp:25] Expected dim <= nbDims && dim >= -(nbDims + 1) to be true but got false
Dimension out of range (expected to be in range of [-1, 0], but got 2)

Only when I adjust one of the dimensions to be consistent, will it not report an error. like this

trt_model = torch_tensorrt.compile(
    traced_mlm_model, 
    inputs= (torch_tensorrt.Input(min_shape=[1, 128],opt_shape=[16, 128],max_shape=[32, 128],dtype=torch.int32),  # input_ids
             torch_tensorrt.Input(min_shape=[1, 128],opt_shape=[16, 128],max_shape=[32, 128],dtype=torch.int32)),  # attention_mask
    enabled_precisions= {torch.float32}, # Run with 32-bit precision
    workspace_size=2000000000,
    truncate_long_and_double=True,
    debug=True
)

But this is not what I want :(

Expected behavior

Environment

Build information about Torch-TensorRT can be found by turning on debug messages

• Torch-TensorRT Version (e.g. 1.4.0):
• PyTorch Version (e.g. 2.0.1):
• CPU Architecture: x64
• OS (e.g., Linux): Linux
• How you installed PyTorch (conda, pip, libtorch, source): pip
• Build command you used (if compiling from source): .whl
• Are you using local sources or building from archives:
• Python version: 3.8
• CUDA version: 11.7
• GPU models and configuration: 4090
• Any other relevant information:

Additional context