arg['tensor_parallel_size'] is overrided by GPU's number

[hyeonu]

I want to set PIPELINE_PARALLEL_SIZE instead of TENSOR_PARALLEL_SIZE for multiple GPUs.

But tensor_parallel_size is always set to number of GPU.

It's needed to check whether the env was set or not in engine_args.py:

    num_gpus = device_count()
    if num_gpus > 1:
        args["tensor_parallel_size"] = num_gpus