Fix issue: cuda index incorrectly -1 when tp_size>1

[ZX-ModelCloud]

vllm now needs to call the init_distributed_environment method correctly so that _LOCAL_RANK will not be -1 when tp_size>1.

vllm source code: https://github.com/vllm-project/vllm/blob/main/vllm/distributed/parallel_state.py#L193

Someone has reproduced this issue: #487 (comment)

[Qubitium]

Tested with vllm 0.4.3. No issues.

[merrymercy]

@Qubitium @ZhouXingg Thanks for the contributions. They are merged in #511.