fix typo in LmHeadLinearAllreduce initialization

Author: ftian1

deepspeed/inference/quantization/layers.py

@@ -135,7 +135,7 @@ def forward(self, input: Tensor) -> Tensor:
 class QuantizedLmHeadLinearAllreduce(nn.Linear):
 
     def __init__(self, config: Dict, pre_quant_layer: nn.Linear) -> None:
-        super(QuantizedLinearLayer, self).__init__(in_features=pre_quant_layer.weight.shape[1],
+        super(QuantizedLmHeadLinearAllreduce, self).__init__(in_features=pre_quant_layer.weight.shape[1],
                                                    out_features=pre_quant_layer.weight.shape[0],
                                                    bias=pre_quant_layer.bias is not None,
                                                    device=pre_quant_layer.weight.device,

deepspeed/inference/quantization/quantization.py

@@ -85,8 +85,6 @@ def _init_group_wise_weight_quantization(model: nn.Module, ds_config: Dict) -> n
             if is_zero3_enabled:
                 module.weight.all_gather()
 
-            assert module.weight.dtype == torch.float16, 'Model weight is expected in half.'
-
             new_module = QUANTIZATION_LAYER_MAPPINGS[type(module)](matched_quantization_config, module)
 
             if is_zero3_enabled: