[BUG]RuntimeError: output tensor must have the same type as input tensor

when i set reduce_scatter flag to true
zero++ config settings:

{
  "bf16": {
    "enabled": "auto"
  },
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": "auto",
      "betas": "auto",
      "eps": "auto",
      "weight_decay": "auto"
    }
  },
  "scheduler": {
    "type": "WarmupDecayLR",
    "params": {
      "total_num_steps": "auto",
      "warmup_min_lr": "auto",
      "warmup_max_lr": "auto",
      "warmup_num_steps": "auto"
    }
  },
  "zero_optimization": {
    "stage": 3,
    "offload_optimizer": {
      "device": "cpu",
      "pin_memory": true
    },
    "offload_param": {
      "device": "cpu",
      "pin_memory": true
    },
    "overlap_comm": true,
    "contiguous_gradients": true,
    "sub_group_size": 1e9,

    "stage3_prefetch_bucket_size": "auto",
    "stage3_param_persistence_threshold": "auto",
    "stage3_max_live_parameters": 1e9,
    "stage3_max_reuse_distance": 1e9,
    "stage3_gather_16bit_weights_on_model_save": false,

    "reduce_bucket_size": 10000000,
    "reduce_scatter": true,

    "zero_quantized_weights": true,
    "zero_hpz_partition_size": 16,
    "zero_quantized_gradients": true

  },
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "steps_per_print": 5,
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "wall_clock_breakdown": false
}

The report provides a DeepSpeed JSON configuration but no source file, test, traceback, environment, or model details. Start by reproducing the error with reduce_scatter enabled and the posted ZeRO-3 settings, then trace the operation producing the tensor type mismatch. Done means identifying the incompatible tensors and adding a regression test or confirmed fix.