How GLM built its own inference infrastructure