[Bugfix] Avoid repeatedly creating dummy data during engine startup

[DarkLight1337]

This PR fixes an issue where the startup time of multimodal models is multipled because dummy data is created multiple times during profile run and graph capturing. Instead of disabling cache when dummy data is generated, the cache is now always enabled. To conserve memory, the cache is instead cleared at the end of the engine start process.

Cache reset code is taken from #16478

[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.

🚀

[ywang96]

I'm actually not seeing the speedup from this PR with the following test script

# python3 test.py --model Qwen/Qwen2.5-VL-3B-Instruct

# test.py
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.engine.async_llm_engine import AsyncEngineArgs
from vllm.engine.async_llm_engine import UsageContext
from vllm.utils import FlexibleArgumentParser
from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
import time

if __name__ == "__main__":
    parser = FlexibleArgumentParser()
    parser = AsyncEngineArgs.add_cli_args(parser)
    args = parser.parse_args()
    engine_args = AsyncEngineArgs.from_cli_args(args)
    vllm_config = engine_args.create_engine_config(UsageContext.ENGINE_CONTEXT)
    
    start_time = time.perf_counter()
    encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
        model_config=vllm_config.model_config,
        scheduler_config=vllm_config.scheduler_config,
        mm_registry=MULTIMODAL_REGISTRY,
    )
    print(f"Time taken: {time.perf_counter() - start_time}")

On main

Time taken: 10.416103872470558

This branch

Time taken: 16.340763847343624

[DarkLight1337]

The cache is only effective if you run compute_encoder_budget a second time.

[DarkLight1337]

In actual inference this means that the multi-modal data used in dummy run is cached (since it's called after compute_encoder_budget), so there should still be an overall speedup.

[DarkLight1337]

OK it seems that I do need to implement the async version of this method...