vllm - 💡(How to fix) Fix [Bug]: Performance drop when using multiple VLLM on different numa nodes --> bugs in ompmultiprocessing.py

# Start both vLLM instances + nginx load balancer
./run_bench.sh start

# Run benchmark (48 prompts, 32 concurrent, through LB)
./run_bench.sh bench

[docker-compose.broken.yml](https://github.com/user-attachments/files/28071993/docker-compose.broken.yml)
[docker-compose.fixed.yml](https://github.com/user-attachments/files/28071992/docker-compose.fixed.yml)
[run_bench.sh](https://github.com/user-attachments/files/28071994/run_bench.sh)
[nginx_conf.txt](https://github.com/user-attachments/files/28072042/nginx_conf.txt)

Script is building and running 2 VLLM instances in docker with pinning taken from K8s setup via NRI_driver (checked it is pinned correctly) and next run benchmark.

Results
When using VLLM_CPU_OMP_THREADS_BIND=auto 
I instance is performing well (by accident)
(APIServer pid=1) INFO 05-20 19:14:31 [loggers.py:271] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: **203.2 tokens/s**, Running: 16 reqs
while second replica is generating 
(APIServer pid=1) INFO 05-20 19:16:10 [loggers.py:271] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 17.6 tokens/s, Running: 16 reqs

Why this is happening:
looking on binding during VLLM load:

Instance that is performing well:

(EngineCore pid=112) INFO 05-20 19:07:25 [ompmultiprocessing.py:187]    VLLM_CPU_OMP_THREADS_BIND='auto', auto_setup=True, skip_setup=False
(EngineCore pid=112) INFO 05-20 19:07:25 [ompmultiprocessing.py:187]    local_world_size=1, reserve_cpu_num=1
**(EngineCore pid=112) INFO 05-20 19:07:25 [ompmultiprocessing.py:187]    local_rank=0, core ids=[1, 10, 11, 12, 13, 122, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155
, 156, 157, 158]
(EngineCore pid=112) INFO 05-20 19:07:25 [ompmultiprocessing.py:187]    reserved_cpus=[159]**


Intance that performs poor:
(EngineCore pid=112) INFO 05-20 19:07:24 [ompmultiprocessing.py:187]    VLLM_CPU_OMP_THREADS_BIND='auto', auto_setup=True, skip_setup=False                                                                                         (EngineCore pid=112) INFO 05-20 19:07:24 [ompmultiprocessing.py:187]    local_world_size=1, reserve_cpu_num=1
**(EngineCore pid=112) INFO 05-20 19:07:24 [ompmultiprocessing.py:187]    local_rank=0, core ids=[]                                                                                                                                   (EngineCore pid=112) INFO 05-20 19:07:24 [ompmultiprocessing.py:187]    reserved_cpus=[]**

What happened with our second instance

Our 2-pod StatefulSet was allocated:
- **Pod 0** cpuset: `60,100-119,181-191` → all cores on **NUMA node 1**
- **Pod 1** cpuset: `1,10-13,122,134-159` → all cores on **NUMA node 0**

With auto-bind (`VLLM_CPU_OMP_THREADS_BIND` unset), both pods (rank 0) tried to bind OMP threads to NUMA node 0 cores (CPUs 0-59, 120-179).

- **Pod 1** (cpuset on node 0): worked correctly by coincidence — its cpuset overlapped with node 0.
- **Pod 0** (cpuset on node 1): vLLM attempted to bind to node 0 cores like `0-59`, but those cores are **not in the container's cpuset** (`60,100-119,181-191`)

How to fix it:

We can define VLLM_CPU_OMP_THREADS_BIND as a list of vcpus from cpuset, this way we always pass correct list of CPUs (ballon policy cares for proper vCPUs list we just can take it from /sys/fs/cgroup/cpuset.cpus.effective)
, but there is bug #2 as without passing VLLM_CPU_OMP_THREADS_BIND=auto  `VLLM_CPU_NUM_OF_RESERVED_CPU` is silently ignored and list of reserved cores is empty --> ends up with poor performance.(because the reserved-CPU logic only exists in the auto-bind path.)

How did we hack to get proper performance numbers on both VLLM instances:
We are setting VLLM_CPU_OMP_THREADS_BIND as a list from /sys/fs/cgroup/cpuset.cpus.effective (proper list defined by ballons policy) and removing last cores.
```bash
export VLLM_CPU_OMP_THREADS_BIND=$(python3 -c "
r = open('/sys/fs/cgroup/cpuset.cpus.effective').read().strip()
cores = [c for p in r.split(',')
         for c in (range(int(p.split('-')[0]), int(p.split('-')[1])+1) if '-' in p else [int(p)])]
print(','.join(map(str, cores[:-1])))
")