[7/N] torch.compile, reduce compilation time

[youkaichao]

Normal run:

$ vllm serve meta-llama/Meta-Llama-3-8B
Memory profiling took 1.16 seconds
# GPU blocks: 27846, # CPU blocks: 2048
Graph capturing finished in 15 secs, took 0.32 GiB

Run with inductor compiling the full graph:

$ vllm serve meta-llama/Meta-Llama-3-8B -O 3
Memory profiling took 28.21 seconds
# GPU blocks: 27825, # CPU blocks: 2048
Graph capturing finished in 32 secs, took 0.33 GiB

Run with inductor compiling the piecewise graph:

$ vllm serve meta-llama/Meta-Llama-3-8B -O 3
Memory profiling took 21.31 seconds
# GPU blocks: 27846, # CPU blocks: 2048
Graph capturing finished in 21 secs, took 0.33 GiB

Results show we can reduce both the inductor compilation time and the cudagraph capture time. Here I only compile one graph for symbolic shape. I don't know why it helps cudagraph capture time, it's just running the model for 2 warmup runs. Might be related to triton?

Note: need to get #10273 in so that inductor can cache the compilation. currently i just disable post grad custom pass to test it.

I think the compilation result should be the same, but in the full graph level, inductor needs to run the full passes at the full graph level, and can generate every triton kernels for every layer.

[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.

🚀

[WoosukKwon]

What are these default values for?

[youkaichao]

they are the default splitting ops we know inductor cannot do any optimization, and they are good candidates to split the model.

[mgoin]

LGTM thank you

[youkaichao]

test errors are unrelated. merging.