[Feature]: membind all NUMA nodes for all CPUs in list

[hpcpony]

🚀 The feature, motivation and pitch

When specifying VLLM_CPU_OMP_THREADS_BIND it might be better if vLLM did a membind to all or the nodes related to the CPU list rather than just doing it to the node associated with first CPU in the list.

Alternatives

Using numactl and OMP environment variables (i.e. OMP_PLACES, etc...) give better overall control of binding but isn't compatible with CPU tensor parallel partitioning.

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.