issues during exporting VAD ONNX

[kevinteng525]

Hello NV team,

Thanks a lot for your project!
I tried to export ONNX of VAD module, and followed guide to setup env, download datasets and checkpoints.
However during exporting ONNX

python export_no_prev.py /workspace/VAD/projects/configs/VAD/VAD_tiny_stage_2.py /workspace/VAD/ckpts/VAD_tiny.pth --launcher none --eval bbox --tmpdir tmp

I met issue

TypeError: z(): incompatible function arguments. The following argument types are supported:
1. (self: torch._C.Node, arg0: str, arg1: torch.Tensor) -> torch._C.Node
Invoked with: %1510 : Tensor = onnx::Constant(), scope:
bev_deploy.patch.inspect.AutoInspectModule::/projects.mmdet3d_plugin.VAD.VAD_transformer.VADPerceptionTransformer::transformer/projects.mmdet3d_plugin.VAD.modules.encoder.BEVFormerEncoder::encoder/projects.mmdet3d_plugin.VAD.modules.encoder.BEVFormerLayer::layers.0/projects.mmdet3d_plugin.VAD.modules.temporal_self_attention.TemporalSelfAttention::attentions.0
, 'value', 64
(Occurred when translating PythonOp).

After updated AV-Solutions/vad-trt/export_eval/bev_deploy/patch/bevformer/ms_deform_attn.py to:

# Fix: Ensure all args are proper tensor nodes for ONNX
        processed_args = []
        for arg in args:
            if isinstance(arg, (int, float)):  # If it's a scalar
                # Convert scalar to ONNX Constant tensor
                processed_args.append(g.op("Constant", value_t=torch.tensor(arg)))
            else:
                processed_args.append(arg)
        return g.op("custom_op::MultiScaleDeformableAttentionPlugin", *processed_args)

The issue solved.

However the exported ONNX cannot be inference:

onnx.onnx_cpp2py_export.checker.ValidationError: Nodes in a graph must be topologically sorted, however input '/positional_encoding/Unsqueeze_output_0' of node:
name: /positional_encoding/Expand OpType: Expand

Pls. provide some guidances.

Thanks,
Kevin