[Bugfix] Fix decode tokens w. CUDA graph

[comaniac]

Fixes #6703.
This is not related to speculative decoding particularly but not sure why no unit test failed. Will try to add a unit test in this or a follow-up PR.

cc @cadedaniel @alexm-neuralmagic

[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 consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your fast-check build on Buildkite UI.

Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).

To run full CI, you can do one of these:

• Comment /ready on the PR
• Add ready label to the PR
• Enable auto-merge.

🚀

[cadedaniel]

How do we know this fixes the issue?

[comaniac]

How do we know this fixes the issue?

I tested it locally and will add a unit test to catch this later. Just put the PR here to unblock benchmarking first.

[comaniac]

Turns out that num_decode_tokens is probably the only attribute missed in the unit test...
The new added assertion will fail without this PR.