[REQUEST] How to achieve DeepNVMe (ZeRO 3 offloading) via GPU Direct Storage?

Is your feature request related to a problem? Please describe.
I'm new to DeepSpeed and am interested in using DeepNVMe to address GPU out-of-memory issues, the results in the blog post really impressed me: https://github.com/deepspeedai/DeepSpeed/tree/master/blogs/deepnvme/08-2024.
But I'm do not know how to use it. For example, if I want to fine-tune a large model, do I need to create DeepNVMe handle by myself to offload the optimizer state? As described in the tutorial: https://www.deepspeed.ai/tutorials/deepnvme/

Or I just need to configure the ds_config:
ds_config = {
"train_batch_size": 16,
"gradient_accumulation_steps": 1,
"fp16": {"enabled": False},
"optimizer": {
"type": "AdamW",
"params": {"lr": 2e-5, "weight_decay": 0.01}
},
"zero_optimization": {
"stage": 3,
"offload_optimizer": {"device": "nvme", “nvme_path”： “/local_nvme”, "pin_memory": True},
"reduce_bucket_size": 1e7,
"stage3_prefetch_bucket_size": 2e8,
}
}
Is this necessary?
“aio”： {
“block_size”： 262144，
“queue_depth”： 32，
“thread_count”： 1，
“single_submit”： false，
“overlap_events”： true
}

Start with the DeepNVMe tutorial and the DeepNVMe blog post linked in the issue, then compare their usage with the supplied ZeRO-3 configuration and AIO settings. Done means clearly documenting whether users need to create a DeepNVMe handle or can rely on configuration for optimizer offloading, including the role of the shown settings.