[torch.compile] allow tracking forward time

[youkaichao]

when benchmarking torch.compile performance, I always get this question: if the performance gain is not satisfactory, is it because torch.compile does not optimize the model well, or is it because of the scheduling overhead?

this pr adds the tracking for forward time, so that we can directly test the perf of torch.compile for certain sizes.

e.g.

$ VLLM_LOG_BATCHSIZE_INTERVAL=1.0 python3 benchmarks/benchmark_latency.py --model meta-llama/Meta-Llama-3-8B --batch-size 1 --load-format dummy
INFO 12-10 19:52:26 forward_context.py:88] Batchsize forward time stats (batchsize, count, median_time(ms)): [(1, 5054, 6.5), (32, 41, 7.51)]

$ VLLM_LOG_BATCHSIZE_INTERVAL=1.0 python3 benchmarks/benchmark_latency.py --model meta-llama/Meta-Llama-3-8B --batch-size 1 --load-format dummy -O "{'level': 3, 'candidate_compile_sizes': [1]}"
INFO 12-10 19:54:34 forward_context.py:88] Batchsize forward time stats (batchsize, count, median_time(ms)): [(1, 5049, 5.93), (32, 41, 7.35)]

then it is clear that the forward time improves from 6.5ms to 5.93ms, 8.8% improvement. And the end-to-end 7.7% improvement in latency in #11078 is shadowed a little bit in the end-to-end pipeline.

[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]

Batchsize forward time stats (batchsize, count, median_time(ms)): [(1, 5054, 6.5), (32, 41, 7.51)]

the 32 should be the prefill length. so it is interesting to see that torch.compile also accelerates prefill in this case.