使用 Altas300 对 GroundingDino 进行模型部署，导出的 onnx 模型为动态模型，在转换为 om 模型时，固定了其尺寸。

• 模型转换指令为

  atc --model=groundingdino_swint_ogc.onnx --framework=5 --output=groundingdino --soc_version=Ascend310 --input_shape="img:1,3,800,1200;input_ids:1,6;attention_mask:1,6;position_ids:1,6;token_type_ids:1,6;text_token_mask:1,6,6"

• 在模型转换过程中的输出为：

  ATC start working now, please wait for a moment.
  ...
  ATC run failed, Please check the detail log, Try 'atc --help' for more information
  E10042: 2025-04-26-11:53:07.780.080 GenerateOfflineModel execute failed.
          TraceBack (most recent call last):
          op[/Concat_3], the input shape dims should be equal except merge axis,shapes:[[1, ], [1, 1, ], ]axis:0[FUNC:ConcatInferShapeCommon][FILE:split_combination_ops.cc][LINE:1213]
          Call InferShapeAndType for node:/Concat_3(ConcatD) failed[FUNC:Infer][FILE:infershape_pass.cc][LINE:120]
          process pass InferShapePass on node:/Concat_3 failed, ret:4294967295[FUNC:RunPassesOnNode][FILE:base_pass.cc][LINE:570]
          build graph failed, graph id:0, ret:1343242270[FUNC:BuildModelWithGraphId][FILE:ge_generator.cc][LINE:1615]
          GenerateOfflineModel execute failed.

我在论坛上查找了相似问题，最相似的问题是 “华为atlas 200idka2和atlas 200dk部署groundingdino出现的问题” 以及 “[GroundingDINO]ONNX模型转OM模型过程中遇到input shape dim未对齐问题” 等。前者似乎成功解决了这个问题，但是显示帖子被删除，后者版主回复了两个参考思路。

• 针对第一个思路，使用 onnxsim 进行优化后转换，爆同样的错误。
• 针对第二个思路，我使用下方代码删除了 concat 3 4 5 6 四个算子的 unsqueeze 节点，在进行模型转换时，没有报 concat 算子问题，但是爆了几百个算子不支持问题，报错信息类似于 No supported Ops kernel and engine are found for [/encoder/blocks.1/attn/Einsum], optype [Einsum]

  import onnx
  
  # concat 3 (3 4 5)
  # concat 4 (6 7 8)
  # concat 6 (9 10)
  # concat 7 (11 12)
  # concat 9 (13)
  # concat 10 (14)
  # concat 11 (15 16)
  # concat 13 (17 18 19 20 21 22)
  # concat 14 (23 24 25 26 27 28)
  
  def main(node_name_del, node_name_main):
      model = onnx.load("groundingdino_swint_ogc_modified.onnx")
      graph = model.graph
  
      # 创建映射：unsqueeze_output → 原始 input
      unsqueeze_map = {}
  
      # 记录要删除的节点
      nodes_to_remove = []
  
      for node in graph.node:
          if node.name in node_name_del:
              unsqueeze_map[node.output[0]] = node.input[0]
              nodes_to_remove.append(node)
  
      # 替换 Concat_3 中的输入
      for node in graph.node:
          if node.name == node_name_main:
              node.input[:] = [unsqueeze_map.get(inp, inp) for inp in node.input]
  
      # 移除 unsqueeze 节点
      for node in nodes_to_remove:
          graph.node.remove(node)
  
      onnx.save(model, "groundingdino_swint_ogc_modified.onnx")
  
  node_name_main_list = ["/Concat_3", "/Concat_4", "/Concat_6", "/Concat_7"] # "/Concat_9", "/Concat_10", "/Concat_11", "/Concat_13", "/Concat_14"
  node_name_del_dict = {
      "/Concat_3": ["/Unsqueeze_3", "/Unsqueeze_4", "/Unsqueeze_5"],
      "/Concat_4": ["/Unsqueeze_6", "/Unsqueeze_7", "/Unsqueeze_8"],
      "/Concat_6": ["/Unsqueeze_9", "/Unsqueeze_10"],
      "/Concat_7": ["/Unsqueeze_11", "/Unsqueeze_12"],
      "/Concat_9": ["/Unsqueeze_13"],
      "/Concat_10": ["/Unsqueeze_14"],
      "/Concat_11": ["/Unsqueeze_15", "/Unsqueeze_16"],
      "/Concat_13": ["/Unsqueeze_17", "/Unsqueeze_18", "/Unsqueeze_19", "/Unsqueeze_20", "/Unsqueeze_21", "/Unsqueeze_22"],
      "/Concat_14": ["/Unsqueeze_23", "/Unsqueeze_24", "/Unsqueeze_25", "/Unsqueeze_26", "/Unsqueeze_27", "/Unsqueeze_28"],
  }
  for node_name_main in node_name_main_list:
      print(f"Processing {node_name_main}...")
      node_name_del = node_name_del_dict[node_name_main]
      main(node_name_del, node_name_main)