Why cast_lean() before conv op ?

[MirkoCalvi]

Description

In all the codegenerated convolutions calls (like in Conv, Conv_Clip, Conv_Relu..) before the conv operation cast_lean() is called on the Kernel of the convolution.

//--KERNEL CAST--
        if (need_kernel_cast) {
            // Generate cast for kernel
            const kernel_name = try IR_utils.getSanitizedName(self.op_Conv.input_W.name);
            _ = try writer.print(
                \\
                \\    // Cast kernel from {s} to {s}
                \\    var tensor_{s}_casted = Tensor({s}).fromShape(&allocator, @constCast(param_lib.tensor_{s}.shape)) catch return -2;
                \\    defer tensor_{s}_casted.deinit();
                \\    tensMath.cast_lean({s}, {s}, @constCast(&param_lib.tensor_{s}), &tensor_{s}_casted, zant.onnx.DataType.FLOAT) catch return -1;
                \\
            , .{
                self.op_Conv.input_W.ty.toString(),
                target_type,
                kernel_name,
                target_type,
                kernel_name,
                kernel_name,
                self.op_Conv.input_W.ty.toString(),
                target_type,
                kernel_name,
                kernel_name,
            });
            final_kernel_string = try std.mem.concat(allocator, u8, &[_][]const u8{ "@constCast(&tensor_", kernel_name, "_casted)" });
            need_free_kernel = true;
        } else {
            final_kernel_string = tensor_W_string;
        }

Why? Isn't it an unnecessary step? Shouldn't we have the kernels already in the correct form? if not, why aren't we doing the casting when needed instead of adding another element-wise operation before?

Expected Behavior

Correct functioning without the cast_lean() on the kernel.