GLM-130B/quantization/functional.py

import torch

from kernels import extract_weight_to_half


class W8A16Linear(torch.autograd.Function):
    @staticmethod
    def forward(ctx, inp: torch.Tensor, quant_w: torch.Tensor, scale_w: torch.Tensor, weight_bit_width):
        ctx.inp_shape = inp.size()
        ctx.weight_shape = quant_w.size()
        ctx.weight_bit_width = weight_bit_width
        out_features = quant_w.size(0)
        inp = inp.contiguous().view(-1, inp.size(-1))
        weight = extract_weight_to_half(quant_w, scale_w, weight_bit_width)
        output = inp.mm(weight.t())
        ctx.save_for_backward(inp, quant_w, scale_w)
        return output.view(*(ctx.inp_shape[:-1] + (out_features,)))

    @staticmethod
    def backward(ctx, grad_output: torch.Tensor):
        inp, quant_w, scale_w = ctx.saved_tensors
        weight = extract_weight_to_half(quant_w, scale_w, ctx.weight_bit_width)
        grad_output = grad_output.contiguous().view(-1, weight.size(0))
        grad_input = grad_output.mm(weight)
        grad_weight = grad_output.t().mm(inp)
        return grad_input.view(ctx.inp_shape), grad_weight.view(ctx.weight_shape), None