vllm

[Bugfix] Fix cuda graph sizes when running with speculative decoding

#30330

Merged