ZeRORuntimeException: You are using ZeRO-Offload with a client provided optimizer

显卡配置：2张 V100 32G （共四张，有两张别人占用中，用完后可实现利用4卡V100）
按照默认accelerate配置报错：cuda out of memory，观察发现默认配置中 offload_optimizer_device 和 offload_param_device 参数均为none，后按照accelerate教程，将这两个参数均改成 cpu 报错：

accelerate 配置如下：

command_file: null
commands: null
compute_environment: LOCAL_MACHINE
deepspeed_config:
gradient_accumulation_steps: 1
gradient_clipping: 1.0
offload_optimizer_device: cpu
offload_param_device: cpu
zero3_save_16bit_model: true
zero_stage: 3
zero3_init_flag: true
zero_force_ds_cpu_optimizer: False
distributed_type: DEEPSPEED
downcast_bf16: 'no'
dynamo_backend: 'NO'
fsdp_config: {}
gpu_ids: null
machine_rank: 0
main_process_ip: null
main_process_port: null
main_training_function: main
megatron_lm_config: {}
mixed_precision: fp16
num_machines: 1
num_processes: 2
rdzv_backend: static
same_network: true
tpu_name: null
tpu_zone: null
use_cpu: false

和别人沟通后得知8*A100(40G)可按照官方的脚本直接训练，不用修改cpu offload参数，受限于机器资源，怎样能实现cpu offload 且不报错呢？需要配置deepspeed 配置文件（.json）吗？

Start by reviewing the provided Accelerate and DeepSpeed configuration, then reproduce the ZeRORuntimeException with CPU offload enabled on the stated GPU setup. Done means identifying a supported configuration or documenting that CPU offload cannot work with this training path without changing the optimizer setup.