bug: CRD VllmRuntime : increase shm size

[dotmobo]

Describe the bug

Hi

Currently we cannot increase the shm size like the helm chart for tensor parallelism

To Reproduce

Find shm size option in crd api

Expected behavior

Add shmSize: 24g

Additional context

No response

[ruizhang0101]

Thanks for bringing this up! We will take a look at it and feel free to contribute to it as well.

[Tobi-Adesoye]

When you run distributed tensor-parallelism or multi-node setups via Ray, standard container shared memory limits (64MB default) will choke and crash almost instantly. The processes need massive IPC pipelines to pass model weights and communication tensors.

Setting the CRD's shared memory limit to match your system RAM (or mounting a Memory type emptyDir to /dev/shm) is the immediate infrastructure patch needed here. That said, the cleaner the tensor layout alignment is before it hits the shared memory space, the lower the actual cross-worker communication overhead becomes.