[Bug Fix]Enabling a specific torch.dynamo config

[fangfangssj]

PR Category

Bug Fix

Description

运行extractor的时候，可能会因为dynamo的原因（dynamo遇到了不支持的算子，模型包含复杂的控制流，模型的输出包含标量等），捕捉到的整张图被打断，分成了很多很多的子图。
在extractor的时候，log信息多次打印出successfully的时，代表着extractor被拆分成了对应的子图，这时候，只保存了最后一个子图在文件里。
在validate的时候，也会导致对应的报错，甚至还会因为第一次被打断的地方比较靠前，导致只有打断前的一小部分用来计算hash值，导致hash值重复。

本PR增加了torch.dynamo config的来避免模型的输出包含标量时，图被打断。

Testing model: yolov8n on cuda:0
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
Graph and tensors for 'yolov8n' extracted successfully to: /root/graphnet_workspace/yolov8n
[OK] yolov8n

Check extractability ...
model_path='/root/graphnet_workspace/yolov8n/'
tensor(201762, device='cuda:0') tensor(14720, device='cuda:0')
torch.Size([1, 84, 8400])
model_path='/root/graphnet_workspace/yolov8n/'
Graph and tensors for 'temp' extracted successfully to: /tmp/tmp8k5ab9ef/temp
Graph and tensors for 'temp' extracted successfully to: /tmp/tmp8k5ab9ef/temp
Graph and tensors for 'temp' extracted successfully to: /tmp/tmp8k5ab9ef/temp
Graph and tensors for 'temp' extracted successfully to: /tmp/tmp8k5ab9ef/temp
tensor(526962, device='cuda:0') tensor(239, device='cuda:0')
torch.Size([1, 84, 8400])
Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/root/GraphNet/graph_net/torch/single_device_runner.py", line 103, in <module>
    main(args=args)
  File "/root/GraphNet/graph_net/torch/single_device_runner.py", line 51, in main
    with _dump_graph_hash_key_ctx(args) as dump_graph_options:
  File "/root/miniconda3/lib/python3.12/contextlib.py", line 144, in __exit__
    next(self.gen)
  File "/root/GraphNet/graph_net/torch/single_device_runner.py", line 46, in _dump_graph_hash_key_ctx
    assert len(mut_graph_codes) == 1, f"{len(mut_graph_codes)=}"
           ^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: len(mut_graph_codes)=4

[paddle-bot]

Thanks for your contribution!