Smthri

template <typename scalar_t, typename idx_t>
void softmaxLUT_kernel(scalar_t* input,  scalar_t* output, idx_t n, idx_t c, idx_t sizeC,
                  idx_t input_sN, idx_t input_sC,  idx_t output_sN, idx_t output_sC,
                  idx_t multiplier, idx_t zero_point, idx_t* exp_table) {
    /*
     * Perform quantized softmax for one element
     *
     * output_i = multiplier / sum (table[A_j - A_i + 255] for all j)
     */

Smthri

import os
import sys
import argparse
import numpy as np
import torch
import torch.quantization as tq
from pathlib import Path
from torch import nn
from _collections_abc import Iterable
from collections import OrderedDict

Smthri

import numpy as np
import torch
import torch.quantization as tq
from torch import nn
from copy import deepcopy

## Manual quantization

# weight, bias, input
kernel_size = 3