🐛 [Bug] RuntimeError: [Error thrown at core/conversion/converters/impl/select.cpp:236] Expected const_layer to be true but got false

[pupumao]

Bug Description

when i convert jit model, got the error
this is my forward code:
input x shape is (batch, 6, height, width), first step is to split x into two tensors, but failed

    def forward(self, x):
        fg = x[:,0:3,:,:]   ## this line got error
        bg = x[:,3:,:,:]
        
        fg = self.backbone(fg)
        bg = self.backbone(bg)
        out = self.heads(fg, bg)
        return out

complete traceback:

ERROR: [Torch-TensorRT TorchScript Conversion Context] - 3: [network.cpp::addConstant::1052] Error Code 3: Internal Error (Parameter check failed at: optimizer/api/network.cpp::addConstant::1052, condition: !weights.values == !weights.count
)
Traceback (most recent call last):
  File "model_converter.py", line 263, in <module>
    engine = get_engine(model_info.trt_engine_path, calib, int8_mode=int8_mode, optimize_params=optimize_params)
  File "model_converter.py", line 173, in get_engine
    return build_engine(max_batch_size)
  File "model_converter.py", line 95, in build_engine
    return build_engine_from_jit(max_batch_size)
  File "model_converter.py", line 80, in build_engine_from_jit
    tensorrt_engine_model = torch_tensorrt.ts.convert_method_to_trt_engine(traced_model, "forward", **compile_settings)
  File "/usr/local/lib/python3.6/dist-packages/torch_tensorrt/ts/_compiler.py", line 211, in convert_method_to_trt_engine
    return _C.convert_graph_to_trt_engine(module._c, method_name, _parse_compile_spec(compile_spec))
RuntimeError: [Error thrown at core/conversion/converters/impl/select.cpp:236] Expected const_layer to be true but got false
Unable to create constant layer from node: %575 : Tensor = aten::slice(%570, %13, %12, %14, %13) # /data/small_detection/centernet_pytorch_small_detection/models/low_freeze_comb_net.py:455:0

Environment

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

• PyTorch Version (e.g., 1.0): 1.4.0
• CPU Architecture: arm (nx)
• OS (e.g., Linux):
• How you installed PyTorch: docker of nvidia l4t
• Python version: 3.6.9
• CUDA version: 10.2.300
• Tensorrt version: 8.0.1.6

[p517332051]

I tried to reproduce your bug with your code, but failed.Maybe your torch version is too low，try to use higher than torch 1.8.0.

[pupumao]

I use 1.10.1 to trace jit model, but still got same error
my steps:
1、trace jit model on training environment (GPU nvidia tesla T4, torch 1.10.1+cu102 (at first I use 1.4.0))
2、use torch_tensorrt to convert jit model to tensorrt engine (nvidia xavier, torch 1.10.0)

I tested in embed arm device, not on gpu, may be difference ?

[narendasan]

Can you provide the code to reproduce this?

[inocsin]

I use 1.10.1 to trace jit model, but still got same error my steps: 1、trace jit model on training environment (GPU nvidia tesla T4, torch 1.10.1+cu102 (at first I use 1.4.0)) 2、use torch_tensorrt to convert jit model to tensorrt engine (nvidia xavier, torch 1.10.0)

I tested in embed arm device, not on gpu, may be difference ?

@pupumao This may due to dynamic shape issue, can you try this pr #1110

[pupumao]

already solved

it's not a bug, just because wrong input setting for torch-tensorrt api(error kept require_grad=True)