fix json serialization issue

[maxdebayser]

The compile/test_full_graph.py tests were failing because a CompilationConfig object containing non serializable attributes was being dumped to json to be parsed from json again.

Here is an example of a failed run: https://buildkite.com/vllm/ci-aws/builds/11643#_

Traceback (most recent call last):
  File "tests/utils.py", line 658, in wrapper
    f(*args, **kwargs)
  File "tests/compile/test_full_graph.py", line 17, in test_full_graph
    check_full_graph_support(model,
  File "tests/compile/utils.py", line 84, in check_full_graph_support
    llm = LLM(model=model,
          ^^^^^^^^^^^^^^^^
  File "/vllm/vllm/utils.py", line 1034, in inner
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "/vllm/vllm/entrypoints/llm.py", line 189, in __init__
    json.dumps(compilation_config))
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".pyenv/versions/3.11.6/lib/python3.11/json/__init__.py", line 231, in dumps
    return _default_encoder.encode(obj)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".pyenv/versions/3.11.6/lib/python3.11/json/encoder.py", line 200, in encode
    chunks = self.iterencode(o, _one_shot=True)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".pyenv/versions/3.11.6/lib/python3.11/json/encoder.py", line 258, in iterencode
    return _iterencode(o, 0)
           ^^^^^^^^^^^^^^^^^
  File ".pyenv/versions/3.11.6/lib/python3.11/json/encoder.py", line 180, in default
    raise TypeError(f'Object of type {o.__class__.__name__} '
TypeError: Object of type CompilationConfig is not JSON serializable

Printing the model dump of the CompilationConfig object show a non serializable PosixPath object as well as two functions:

{'level': 2, 'backend': '', 'custom_ops': [], 'splitting_ops': ['vllm.unified_flash_attention', 'vllm.unified_flash_infer', 'vllm.unified_v1_flash_attention'], 'use_inductor': True, 'inductor_specialize_for_cudagraph_no_more_than': None, 'inductor_compile_sizes': {}, 'inductor_compile_config': {}, 'inductor_passes': {}, 'use_cudagraph': False, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': None, 'cudagraph_copy_inputs': False, 'pass_config': {'dump_graph_stages': [], 'dump_graph_dir': PosixPath('.'), 'enable_fusion': True, 'enable_reshape': True}, 'compile_sizes': <function PrivateAttr at 0x7f5906143ce0>, 'capture_sizes': <function PrivateAttr at 0x7f5906143ce0>, 'enabled_custom_ops': {}, 'disabled_custom_ops': {}}```

cc: @youkaichao 

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[youkaichao]

sorry for the trouble, I have fixed it in #10581