[Frontend] speed up import time of vllm.config

[davidxia]

by changing some modules in vllm/multimodal to lazily import expensive modules like transformers or only importing them for type checkers when not used during runtime.

contributes to #14924

I ran on main branch python -X importtime -c 'import vllm' 2> import.log && tuna import.log. The visualized call tree shows vllm.config accounts for the majority of the total import time at 55.5%.

On this branch, vllm.config's share decreased to 52.5%.

python -c 'import vllm'

on a Google Compute Engine a2-highgpu-1g (12 vCPUs, 85 GB Memory) instance with 1 A100 GPU

~3% decrease in mean time

before (main branch commit 94d8ec8)

$ hyperfine 'python -c "import vllm"' --warmup 3 --runs 100
Benchmark 1: python -c "import vllm"
  Time (mean ± σ):      9.643 s ±  0.140 s    [User: 10.393 s, System: 1.913 s]
  Range (min … max):    9.485 s … 10.084 s    100 runs

after (my PR commit 054f562)

$ hyperfine 'python -c "import vllm"' --warmup 3 --runs 100
Benchmark 1: python -c "import vllm"
  Time (mean ± σ):      9.355 s ±  0.096 s    [User: 10.284 s, System: 1.740 s]
  Range (min … max):    9.205 s …  9.711 s    100 runs

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

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 either: Add ready label to the PR or enable auto-merge.

🚀

[mergify]

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

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

[davidxia]

@aarnphm this is ready for review, thanks! cc @simon-mo @Chen-0210

[mergify]

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

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

[aarnphm]

I'm a bit hesitant to optimize this file lazily, given that this touches a lot of components within vLLM.

Also let's try to reduce some hint change to minimum.

This PR will requires running the whole suite to make sure it won't introduce any regression.

[mergify]

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

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

[mergify]

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

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

[mergify]

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

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

[davidxia]

@aarnphm thanks for reviewing again. I rebased away the conflict and fix the pre-commit Python formatting check. All checks pass now and ready for another review. 🙏

[davidxia]

On my M1 Mac with 64GB memory, py312, editable install of vllm following these docs

before with master commit 3443aaf

$ hyperfine 'python -c "import vllm.config"' --warmup 10 --runs 100
Benchmark 1: python -c "import vllm.config"
  Time (mean ± σ):      4.332 s ±  0.088 s    [User: 5.246 s, System: 2.792 s]
  Range (min … max):    4.151 s …  4.748 s    100 runs

after with master commit 7108934

$ hyperfine 'python -c "import vllm.config"' --warmup 10 --runs 100
Benchmark 1: python -c "import vllm.config"
  Time (mean ± σ):      4.239 s ±  0.040 s    [User: 5.104 s, System: 2.981 s]
  Range (min … max):    4.145 s …  4.389 s    100 runs

~2.15% speed up in the average import times ((4.332-4.239)÷4.332)