infrastructure-nccl-timeout-nccl-hangs-and-causes-timeout-in-multi-gpu-vllm-de-0ad0773c

THEN Disable custom all-reduce by adding the `--disable-custom-all-reduce` and `--enforce-eager` flags when starting vLLM. Additionally, verify NCCL communication with a small test script (e.g., a simple all_reduce using PyTorch distributed) before running inference.

# Save as test.py, run with: NCCL_DEBUG=TRACE torchrun --nproc-per-node=8 test.py
import torch
import torch.distributed as dist
dist.init_process_group(backend="nccl")
data = torch.FloatTensor([1,] * 128).to(f"cuda:{dist.get_rank()}")
dist.all_reduce(data, op=dist.ReduceOp.SUM)
torch.cuda.synchronize()
value = data.mean().item()
assert value == dist.get_world_size()