MCPcopy Create free account
hub / github.com/deepspeedai/DeepSpeed / __init__

Method __init__

deepspeed/linear/quantization.py:136–144  ·  view source on GitHub ↗
(self,
                 input_dim: int,
                 output_dim: int,
                 bias: bool = False,
                 quantization_config: QuantizationConfig = None,
                 dtype=torch.bfloat16)

Source from the content-addressed store, hash-verified

134 """
135
136 def __init__(self,
137 input_dim: int,
138 output_dim: int,
139 bias: bool = False,
140 quantization_config: QuantizationConfig = None,
141 dtype=torch.bfloat16):
142 super().__init__(input_dim, output_dim, bias=bias, dtype=dtype)
143 assert dtype == torch.bfloat16, "currently only supports bfloat16 dtype"
144 self.weight = QuantizedParameter(self.weight.data, quantization_config=quantization_config)
145
146 def forward(self, input: torch.Tensor) -> torch.Tensor:
147 return F.linear(input, self.weight.dequantized(), self.bias)

Callers

nothing calls this directly

Calls 1

QuantizedParameterClass · 0.85

Tested by

no test coverage detected