The paper presents new differencing algorithms for post-processing GPS data, using double or triple carrier phase differences and multiple baseline sessions. The characteristic feature of the new algorithms is, that they use full sets of Schreiber's type observation differences with theoretically proved diagonal weight matrices. The proposed estimation models are equivalent to the least squares estimation applied to the original system of un-differenced observation equations. The theoretical ground of the algorithms are the theorems on the properties of differencing equations of Schreiber's type. The theorems become practically useful mainly in case of functional models with triple-differences. In a classical approach, this task was simplified for the sake of necessity of inverting non diagonal covariance matrix, usually of a large dimension. Diagonal weight matrix is also obtained in case of multiple point observation session where correlation of the GPS vectors forces in practice the use of the simplified stochastic models. The proposed method eliminates also the problem of selection of a reference satellite. It is very important especially in case of long observation sessions. The algorithms are applied in professional software for GPS relative positioning.