[BUG] multi-node inference initialization fails when trying not to use replace_with_kernel_inject

Describe the bug
I was following HuggingFace's script for deepspeed inference and found it doesn't work when kernel_inject is False

To Reproduce
Script: https://github.com/huggingface/transformers-bloom-inference/blob/main/bloom-inference-scripts/bloom-ds-inference.py
Change kernel_inject=True (line 121) to kernel_inject=False (when kernel_inject=True it works)
Run: deepspeed --num_gpus=4 bloom-ds-inference.py --name="bigscience/bloom-7b1"

I am hoping to test some model with auto tensor parallelism, which deepspeed hasn't supported kernel yet, but it suffers from the same loading issue below. Can you please advice how to address this?

Expected behavior
Looks like a bug in load_checkpoint during the initialization of deepspeed engine:

File "/home/xx/venv/lib/python3.9/site-packages/deepspeed/__init__.py", line 324, in init_inference
            engine = InferenceEngine(model, config=ds_inference_config)engine = InferenceEngine(model, config=ds_inference_config)engine = InferenceEngine(model, config=ds_inference_config)


  File "/home/xx/venv/lib/python3.9/site-packages/deepspeed/inference/engine.py", line 155, in __init__
  File "/home/xx/venv/lib/python3.9/site-packages/deepspeed/inference/engine.py", line 155, in __init__
  File "/home/xx/venv/lib/python3.9/site-packages/deepspeed/inference/engine.py", line 155, in __init__
    engine = InferenceEngine(model, config=ds_inference_config)
  File "/home/xx/venv/lib/python3.9/site-packages/deepspeed/inference/engine.py", line 155, in __init__
        self._load_checkpoint(config.checkpoint)    self._load_checkpoint(config.checkpoint)

self._load_checkpoint(config.checkpoint)  File "/home/xx/venv/lib/python3.9/site-packages/deepspeed/inference/engine.py", line 450, in _load_checkpoint
  File "/home/xx/venv/lib/python3.9/site-packages/deepspeed/inference/engine.py", line 450, in _load_checkpoint

File "/home/xx/venv/lib/python3.9/site-packages/deepspeed/inference/engine.py", line 450, in _load_checkpoint
    self._load_checkpoint(config.checkpoint)
  File "/home/xx/venv/lib/python3.9/site-packages/deepspeed/inference/engine.py", line 450, in _load_checkpoint
        load_path, checkpoint, quantize_config = sd_loader.load(self._config.tensor_parallel.tp_size,load_path, checkpoint, quantize_config = sd_loader.load(self._config.tensor_parallel.tp_size,

    load_path, checkpoint, quantize_config = sd_loader.load(self._config.tensor_parallel.tp_size,AttributeErrorAttributeError
: load_path, checkpoint, quantize_config = sd_loader.load(self._config.tensor_parallel.tp_size,'dict' object has no attribute 'load': AttributeError

Start with bloom-ds-inference.py at the kernel_inject setting on line 121 and reproduce with the provided deepspeed command using kernel_inject=False. Then trace inference/engine.py through InferenceEngine initialization and _load_checkpoint, focusing on the reported sd_loader.load call; done means multi-node initialization no longer raises the shown AttributeError.