diff --git a/cornac/models/bpr/recom_bpr.pyx b/cornac/models/bpr/recom_bpr.pyx
index 3e6ba9574..f99818f7f 100644
--- a/cornac/models/bpr/recom_bpr.pyx
+++ b/cornac/models/bpr/recom_bpr.pyx
@@ -216,7 +216,6 @@ class BPR(Recommender, ANNMixin):
         """
         cdef:
             long num_samples = len(user_ids), s, i_index, j_index, correct = 0, skipped = 0
-            long num_items = self.num_items
             integral f, i_id, j_id, thread_id
             floating z, score, temp
             bool use_bias = self.use_bias
diff --git a/cornac/models/mf/backend_cpu.pyx b/cornac/models/mf/backend_cpu.pyx
index 4dd4174a3..78622be6c 100644
--- a/cornac/models/mf/backend_cpu.pyx
+++ b/cornac/models/mf/backend_cpu.pyx
@@ -19,7 +19,6 @@ import multiprocessing
 
 cimport cython
 from cython.parallel import prange
-from cython cimport floating, integral
 from libcpp cimport bool
 from libc.math cimport abs
 
@@ -28,27 +27,32 @@ cimport numpy as np
 from tqdm.auto import trange
 
 
+ctypedef np.int64_t INT64_t
+
+
 @cython.boundscheck(False)
 @cython.wraparound(False)
-def fit_sgd(integral[:] rid, integral[:] cid, floating[:] val,
-            floating[:, :] U, floating[:, :] V, 
-            floating[:] Bu, floating[:] Bi,
-            integral num_users, integral num_items,
-            floating lr, floating reg, floating mu,
+def fit_sgd(INT64_t[:] rid, INT64_t[:] cid, float[:] val,
+            float[:, :] U, float[:, :] V, 
+            float[:] Bu, float[:] Bi,
+            float lr, float reg, float mu,
             int max_iter, int num_threads,
             bool use_bias, bool early_stop, bool verbose):
     """Fit the model parameters (U, V, Bu, Bi) with SGD"""
     cdef:
-        integral num_ratings = val.shape[0]
-        integral num_factors = U.shape[1]
+        INT64_t num_ratings = val.shape[0]
+        INT64_t u, i, j
+
+        int num_factors = U.shape[1]
+        int f
 
-        floating loss = 0
-        floating last_loss = 0
-        floating r, r_pred, error, u_f, i_f, delta_loss
-        integral u, i, f, j
+        float loss = 0
+        float last_loss = 0
+        float r, r_pred, error, u_f, i_f, delta_loss
+        
 
-        floating * user
-        floating * item
+        float * user
+        float * item
 
     progress = trange(max_iter, disable=not verbose)
     for epoch in progress:
diff --git a/cornac/models/mf/recom_mf.py b/cornac/models/mf/recom_mf.py
index e948d2684..cfd04012a 100644
--- a/cornac/models/mf/recom_mf.py
+++ b/cornac/models/mf/recom_mf.py
@@ -198,8 +198,6 @@ def _fit_cpu(self, train_set, val_set):
             self.i_factors,
             self.u_biases,
             self.i_biases,
-            self.num_users,
-            self.num_items,
             self.learning_rate,
             self.lambda_reg,
             self.global_mean,