[Bugfix]disable cuda graph when max_decode_seq_len is close to max_seq_len_to_capture

[Ximingwang-09]

when enable spec decoding（num_lookahead_slots = 7）and cuda graph. We meet the problem：ValueError: could not broadcast input array from shape (513,) into shape (512,)
When starting spec decode, in order to ensure sufficient space is allocated, new tokens + num_lookahead_slots slots are allocated by default. Therefore, when the input + output is 8186, 8186 + 7 just triggers the boundary of 8192, requiring an additional block to be allocated, resulting in the block_table length exceeding the input_block_tables[i] range.

[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.

🚀

[Ximingwang-09]

related issue：#5563

[youkaichao]

thanks for the great pr! I think lookahead slots should be the root cause.

however, instead of disabling cudagraph, I think we can actually use cudagraph in this case. we just need to remove the lookahead slots when we copy the block table.

[youkaichao]

close in favor of #8340