vllm

[ROCm] Enable the cuMem CUDA-graph pool offload on ROCm

#59523

Open