fix bug when using FSDP

[kaixuanliu]

No description provided.

[kaixuanliu]

@jiqing-feng @casper-hansen pls help review.

[kaixuanliu]

Cases to fix:
run following cmd line in peft:
accelerate launch examples/olora_finetuning/olora_finetuning.py --base_model hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4 --init_lora_weights gaussian
with accelerate config file:

compute_environment: LOCAL_MACHINE
debug: false
distributed_type: FSDP
downcast_bf16: 'no'
enable_cpu_affinity: false
fsdp_config:
  fsdp_auto_wrap_policy: NO_WRAP
  fsdp_backward_prefetch: NO_PREFETCH
  fsdp_cpu_ram_efficient_loading: true
  fsdp_forward_prefetch: false
  fsdp_offload_params: false
  fsdp_sharding_strategy: NO_SHARD
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sync_module_states: true
  fsdp_use_orig_params: false
ipex_config:
  use_xpu: true
machine_rank: 0
main_training_function: main
mixed_precision: 'no'
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
use_cpu: false