[torch.compile] use depyf to dump torch.compile internals

[youkaichao]

depyf https://github.com/thuml/depyf is a library ( I maintain ) that is dedicated for dumping all sorts of internals from torch.compile .

with this pr, run vllm serve meta-llama/Meta-Llama-3-8B -O '{"level": 3, "inductor_specialize_for_cudagraph_no_more_than": 1, "debug_dump_path": "depyf_dump"}' , and we can get:

tree depyf_dump/rank_0 
depyf_dump/rank_0
├── __compiled_fn_1.AFTER_POST_GRAD.0.py
├── __compiled_fn_1.AFTER_POST_GRAD.1.py
├── __compiled_fn_1.AFTER_POST_GRAD.2.py
├── __compiled_fn_1.after_split.0.py
├── __compiled_fn_1.before_split.0.py
├── __compiled_fn_1.Captured_Graph.0.py
├── __compiled_fn_1.Forward_graph.0.py
├── __compiled_fn_1.Forward_graph.1.py
├── __compiled_fn_1.Forward_graph.2.py
├── __compiled_fn_1.Forward_graph.3.py
├── __compiled_fn_1.Forward_graph.4.py
├── __compiled_fn_1.Forward_graph.5.py
├── __compiled_fn_1.kernel_0.py
├── __compiled_fn_1.kernel_10.py
├── __compiled_fn_1.kernel_11.py
├── __compiled_fn_1.kernel_12.py
├── __compiled_fn_1.kernel_13.best_config
├── __compiled_fn_1.kernel_13.py
├── __compiled_fn_1.kernel_14.best_config
├── __compiled_fn_1.kernel_14.py
├── __compiled_fn_1.kernel_15.best_config
├── __compiled_fn_1.kernel_15.py
├── __compiled_fn_1.kernel_16.best_config
├── __compiled_fn_1.kernel_16.py
├── __compiled_fn_1.kernel_17.py
├── __compiled_fn_1.kernel_18.best_config
├── __compiled_fn_1.kernel_18.py
├── __compiled_fn_1.kernel_19.best_config
├── __compiled_fn_1.kernel_19.py
├── __compiled_fn_1.kernel_1.py
├── __compiled_fn_1.kernel_20.best_config
├── __compiled_fn_1.kernel_20.py
├── __compiled_fn_1.kernel_21.best_config
├── __compiled_fn_1.kernel_21.py
├── __compiled_fn_1.kernel_22.best_config
├── __compiled_fn_1.kernel_22.py
├── __compiled_fn_1.kernel_23.best_config
├── __compiled_fn_1.kernel_23.py
├── __compiled_fn_1.kernel_24.best_config
├── __compiled_fn_1.kernel_24.py
├── __compiled_fn_1.kernel_25.best_config
├── __compiled_fn_1.kernel_25.py
├── __compiled_fn_1.kernel_26.py
├── __compiled_fn_1.kernel_27.py
├── __compiled_fn_1.kernel_28.py
├── __compiled_fn_1.kernel_29.py
├── __compiled_fn_1.kernel_2.best_config
├── __compiled_fn_1.kernel_2.py
├── __compiled_fn_1.kernel_30.py
├── __compiled_fn_1.kernel_31.py
├── __compiled_fn_1.kernel_32.py
├── __compiled_fn_1.kernel_33.py
├── __compiled_fn_1.kernel_34.py
├── __compiled_fn_1.kernel_35.py
├── __compiled_fn_1.kernel_36.py
├── __compiled_fn_1.kernel_37.py
├── __compiled_fn_1.kernel_38.py
├── __compiled_fn_1.kernel_39.py
├── __compiled_fn_1.kernel_3.best_config
├── __compiled_fn_1.kernel_3.py
├── __compiled_fn_1.kernel_40.py
├── __compiled_fn_1.kernel_41.py
├── __compiled_fn_1.kernel_42.py
├── __compiled_fn_1.kernel_43.py
├── __compiled_fn_1.kernel_44.py
├── __compiled_fn_1.kernel_45.py
├── __compiled_fn_1.kernel_46.py
├── __compiled_fn_1.kernel_47.py
├── __compiled_fn_1.kernel_48.py
├── __compiled_fn_1.kernel_49.py
├── __compiled_fn_1.kernel_4.py
├── __compiled_fn_1.kernel_50.py
├── __compiled_fn_1.kernel_51.py
├── __compiled_fn_1.kernel_52.py
├── __compiled_fn_1.kernel_53.py
├── __compiled_fn_1.kernel_54.py
├── __compiled_fn_1.kernel_55.py
├── __compiled_fn_1.kernel_56.py
├── __compiled_fn_1.kernel_57.best_config
├── __compiled_fn_1.kernel_57.py
├── __compiled_fn_1.kernel_5.py
├── __compiled_fn_1.kernel_6.best_config
├── __compiled_fn_1.kernel_6.py
├── __compiled_fn_1.kernel_7.py
├── __compiled_fn_1.kernel_8.best_config
├── __compiled_fn_1.kernel_8.py
├── __compiled_fn_1.kernel_9.best_config
├── __compiled_fn_1.kernel_9.py
├── __compiled_fn_1.post_split_module.0.py
├── __compiled_fn_1.pre_insert_deferred_runtime_asserts___compiled_fn_1.0.py
├── __compiled_fn_1.pre_split_module.0.py
├── full_code_for_forward_0.py
└── __transformed_code_0_for_forward.py

it includes the dynamo compiled bytecode, inductor generated kernels, max-autotune configs, graph module during all transformations, etc.

[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.

🚀

[ProExpertProg]

How does depyf know to output after every pass? If it's not automatic, maybe we could add an annotation that can be added to the __call__ function so the graph is printed. And still provide a way to print mid-pass (for fusion, there's a two-step process so printing in the middle helps).

Finally, could you post an example output file from depyf?

[youkaichao]

How does depyf know to output after every pass?

it hooks the lazy_format_graph_code function. so we just need to call it after every pass. NOTE: just lazy_format_graph_code, not logger.debug("%s", lazy_format_graph_code("before split", self.graph)).

could you post an example output file from depyf?

see https://drive.google.com/drive/folders/1gBVguoCbCGKpKb8SveKrbZkIhiZe9jeS?usp=sharing

[ProExpertProg]

This looks good! If we could reduce the verbosity of the output that would be nice but happy to deal with that later

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @youkaichao.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[youkaichao]

yeah we can further organize the output in the future, to debug and improve the performance of inductor.