Method __init__

tensorrt_llm/quantization/layers.py:573–607

(self,
                 in_features,
                 out_features,
                 bias=True,
                 dtype=None,
                 tp_group=None,
                 tp_size=1,
                 gather_output=True,
                 quant_mode=QuantMode(0))

class Fp8RowwiseLinear(Linear):

    def __init__(self,
                 in_features,
                 out_features,
                 bias=True,
                 dtype=None,
                 tp_group=None,
                 tp_size=1,
                 gather_output=True,
                 quant_mode=QuantMode(0)):
        super().__init__(in_features,
                         out_features,
                         bias=bias,
                         dtype=dtype,
                         tp_group=tp_group,
                         tp_size=tp_size,
                         gather_output=gather_output)

        if not quant_mode.has_fp8_rowwise():
            raise ValueError(
                "Fp8 Rowwise Linear has to have act+weight quantization mode set"
            )

        weights_dtype = dtype
        if quant_mode.has_fp8_rowwise():
            weights_dtype = "fp8"

        self.weight = Parameter(shape=(self.out_features, self.in_features),
                                dtype=weights_dtype)

        if quant_mode.has_fp8_rowwise():
            self.per_channel_scale = Parameter(shape=(self.out_features, ),
                                               dtype="float32")

        self.quant_mode = quant_mode
        self.tllm_to_externel_key_dict = {"weight": ["weight", "weight_scale"]}

    def forward(self, x, lora_runtime_params=None):
        assert lora_runtime_params is None, "lora is not supported on SmoothQuantLinear now"