Cuda Error in executeMemcpy: 1 (invalid argument)

I am building a TensorRT network using Python API and getting the following error when I call execute_async (to inference network):

[TensorRT] ERROR: ../rtSafe/cuda/ (1246) - Cuda Error in executeMemcpy: 1 (invalid argument)[

There is no ANY debug information or smth that could help me to understand what am I doing wrong.

I have 2 GRU layers in my network, I construct each layer using Pytorch Weights (from GRU Cell) the following way:

module = ctx.method_args[0]
    input_trt = #Layer input
    hidden_trt = # Layer hidden

    layer_count = 1
    hidden_size = module.hidden_size
    max_seq_len = 1

    kernel_hh = module.weight_hh.detach().cpu().numpy()

    rh = kernel_hh[:512,:]
    zh = kernel_hh[512:1024, :]
    hh = kernel_hh[1024:1536,:]
    kernel_ih = module.weight_ih.detach().cpu().numpy()

    ri = kernel_ih[:512,:]
    zi = kernel_ih[512:1024, :]
    hi = kernel_ih[1024:1536,:]

    #bias = trt.Weights(torch_dtype_to_trt(module.weight.dtype))
    # if module.bias is not None:
    bias_hh = module.bias_hh.detach().cpu().numpy()
    brh = bias_hh[:512]
    bzh = bias_hh[512:1024]
    bhh = bias_hh[1024:1536]
    bias_ih = module.bias_hh.detach().cpu().numpy()

    bri = bias_ih[:512]
    bzi = bias_ih[512:1024]
    bhi = bias_ih[1024:1536]

    # reshape to 2D
    layer_hidden =
    layer_hidden.reshape_dims = (0, 1, -1)
    layer_input =
    layer_input.reshape_dims = (0, 1, -1)

    layer_rep =
        max_seq_length = 1,
    layer_rep.hidden_state = layer_hidden.get_output(0)


    layer_rep.set_weights_for_gate(0,trt.tensorrt.RNNGateType.UPDATE, True, zi)
    layer_rep.set_weights_for_gate(0,trt.tensorrt.RNNGateType.RESET, True, ri)
    layer_rep.set_weights_for_gate(0,trt.tensorrt.RNNGateType.HIDDEN, True, hi)

layer_rep.set_weights_for_gate(0,trt.tensorrt.RNNGateType.UPDATE, False, zh)
    layer_rep.set_weights_for_gate(0,trt.tensorrt.RNNGateType.RESET, False, rh)
    layer_rep.set_weights_for_gate(0,trt.tensorrt.RNNGateType.HIDDEN, False, hh)

layer_rep.set_bias_for_gate(0, trt.tensorrt.RNNGateType.UPDATE, True, bri)
    layer_rep.set_bias_for_gate(0, trt.tensorrt.RNNGateType.RESET, True, bzi)
    layer_rep.set_bias_for_gate(0, trt.tensorrt.RNNGateType.HIDDEN, True, bhi)

layer_rep.set_bias_for_gate(0, trt.tensorrt.RNNGateType.UPDATE, False, brh)
    layer_rep.set_bias_for_gate(0, trt.tensorrt.RNNGateType.RESET, False, bzh)
    layer_rep.set_bias_for_gate(0, trt.tensorrt.RNNGateType.HIDDEN, False, bhh)

# reshape back to 1D
    layer_res =
    layer_res.reshape_dims = (0, -1)

    output._trt = layer_res.get_output(0)

If I remove these layers, then error disappears.

How can I debug it and while there is no open-source Python API available?

Thank you!

Could you please let us know if you are still facing this issue?