Method get_quant_params

backends/xnnpack/operators/node_visitor.py:277–335

(
        self, quant_params: QuantParams, xnn_graph: XNNGraph, external_tag: str = None
    )

        return dtype

    def get_quant_params(
        self, quant_params: QuantParams, xnn_graph: XNNGraph, external_tag: str = None
    ) -> XNNQuantParams:
        if quant_params.per_channel:
            scale = cast(torch.Tensor, quant_params.scale)
            buffer_idx = len(xnn_graph.constant_data)
            num_scales = scale.numel()

            if quant_params.per_channel_group:
                scale = scale.to(torch.bfloat16)

            num_bytes = scale.untyped_storage().nbytes()
            scale_array = ctypes.cast(
                scale.untyped_storage().data_ptr(),
                ctypes.POINTER(ctypes.c_char * num_bytes),
            ).contents
            scale_name = hashlib.sha256(bytes(scale_array)).hexdigest()
            scale_name = "scale_" + scale_name
            xnn_graph.constant_data.append(
                ConstantDataOffset(
                    offset=UINT64_MAX, size=num_bytes, named_key=scale_name
                )
            )
            if external_tag is not None:
                logging.info(
                    f"Adding constant data with name, key {scale_name} and external_tag {external_tag} to named_data_store"
                )
            self._named_data_store.add_named_data(
                scale_name, bytes(scale_array), CONSTANT_TENSOR_ALIGNMENT, external_tag
            )

            if quant_params.per_channel_group:
                return PerChannelGroupQuant(
                    scale=[],
                    channel_dim=quant_params.axis,
                    group_size=quant_params.group_size,
                    scale_buffer_idx=buffer_idx,
                    num_scales=num_scales,
                )
            else:
                return PerChannelQuant(
                    scale=[],
                    channel_dim=quant_params.axis,
                    scale_buffer_idx=buffer_idx,
                    num_scales=num_scales,
                )
        elif quant_params.is_dynamic:
            # NB:
            # We use per_token quantization for per_tensor quantization
            # Beacuase that's the only option in XNNPACK in absance of per_tensor dynamic quantization
            # TODO: Upstream support for per_tensor dynamic quantization or broadcasting same scale value internally
            return PerTokenDynamicQuant(
                num_nonbatch_dims=quant_params.num_nonbatch_dims,
            )

        return PerTensorQuant(
            scale=cast(float, quant_params.scale),
            zero_point=cast(int, quant_params.zp),