def backward()

in bitsandbytes/autograd/_functions.py [0:0]
42 lines of code
13 McCabe index (conditional complexity)

    def backward(ctx, grad_output):
        if ctx.is_empty:
            bias_grad = None if ctx.bias is None else torch.zeros_like(ctx.bias)
            return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, bias_grad, None
        req_gradA, req_gradB, _, req_gradBias, _ = ctx.needs_input_grad
        CAt, subA, A = ctx.tensors
        SCAt, idx = ctx.tensor_states
        formatB = ctx.formatB
        state = ctx.state
        grad_A = grad_B = grad_bias = None

        if req_gradBias:
            # compute grad_bias first before changing grad_output dtype
            grad_bias = grad_output.sum(0, dtype=ctx.dtype_bias)

        # Cast grad_output to fp16
        if len(grad_output.shape) == 3:
            grad_output = grad_output.reshape(-1, grad_output.shape[-1]).contiguous()

        Cgrad, Cgradt, SCgrad, SCgradt, coo_tensor = F.double_quant(grad_output.to(torch.float16))
        if req_gradB:
            CxAt, SAt = F.transform(CAt, formatB, transpose=True)
            C32grad, Sgrad = F.transform(Cgradt, "col32", transpose=True)
            gradB32, SgradB32 = F.igemmlt(C32grad, CxAt, Sgrad, SAt)
            grad_B = F.mm_dequant(gradB32, SgradB32, SCgradt, SCAt)
            if state.threshold > 0.0 and subA is not None:
                grad_B[:, idx] += torch.matmul(grad_output.t(), subA)

        if req_gradA:
            if state.CBt is not None:
                C32grad, Sgrad = F.transform(Cgrad, "col32")
                if state.CxBt is None:
                    state.CxBt, state.SBt = F.transform(state.CBt, to_order=formatB, transpose=True)
                gradA32, SgradA32 = F.igemmlt(C32grad, state.CxBt, Sgrad, state.SBt)
                grad_A = F.mm_dequant(gradA32, SgradA32, SCgrad, state.SCBt).view(ctx.grad_shape).to(ctx.dtype_A)

            elif state.CB is not None:
                CB = state.CB.to(ctx.dtype_A, copy=True).mul_(state.SCB.unsqueeze(1).mul(1.0 / 127.0))
                grad_A = torch.matmul(grad_output, CB).view(ctx.grad_shape).to(ctx.dtype_A)
            elif state.CxB is not None:
                CB = (
                    undo_layout(state.CxB, state.tile_indices)
                    .to(ctx.dtype_A)
                    .mul_(state.SCB.unsqueeze(1).mul(1.0 / 127.0))
                )
                grad_A = torch.matmul(grad_output, CB).view(ctx.grad_shape).to(ctx.dtype_A)
            else:
                raise Exception("State must contain either CBt or CB or CxB matrix for backward")

        return grad_A, grad_B, None, grad_bias, None