scikit-learn
diff --git a/‎sklearn/metrics/_pairwise_distances_reduction/_argkmin_classmode.pyx.tp
Lines changed: 9 additions & 2 deletions b/‎sklearn/metrics/_pairwise_distances_reduction/_argkmin_classmode.pyx.tp
Lines changed: 9 additions & 2 deletions
diff --git a/‎sklearn/metrics/_pairwise_distances_reduction/_base.pyx.tp
Lines changed: 1 addition & 1 deletion b/‎sklearn/metrics/_pairwise_distances_reduction/_base.pyx.tp
Lines changed: 1 addition & 1 deletion
diff --git a/‎sklearn/metrics/_pairwise_distances_reduction/_dispatcher.py
Lines changed: 13 additions & 17 deletions b/‎sklearn/metrics/_pairwise_distances_reduction/_dispatcher.py
Lines changed: 13 additions & 17 deletions
@@ -39,6 +39,7 @@ cdef class ArgKminClassMode{{name_suffix}}(ArgKmin{{name_suffix}}):
         chunk_size=None,
         dict metric_kwargs=None,
         str strategy=None,
+        precomputed_matrix = None
     ):
         """Compute the argkmin reduction with Y_labels.
 
@@ -55,8 +56,14 @@ cdef class ArgKminClassMode{{name_suffix}}(ArgKmin{{name_suffix}}):
         """
         # Use a generic implementation that handles most scipy
         # metrics by computing the distances between 2 vectors at a time.
-        if metric == 'precomputed':
-            return PrecomputedDistanceMatrix.dist(i,j)
+
+        # If a precomputed matrix is provided at the class level, skip the 
+        # rest of the computation and just return the variable stored in the 
+        # datasets_pair file. TODO: check syntax
+
+        if precomputed_matrix:
+            return PrecomputedDistanceMatrix.precomputed
+           
         pda = ArgKminClassMode{{name_suffix}}(
             datasets_pair=DatasetsPair{{name_suffix}}.get_for(X, Y, metric, metric_kwargs),
             k=k,
 
@@ -129,7 +129,7 @@ cdef class BaseDistancesReduction{{name_suffix}}:
     minimal redundant code.
 
     If metric is 'precomputed' and the precomputed matrix is provided, 
-    able to return it.
+    a subclass must be able to access it through the compute method.
     """
 
 
 
@@ -12,7 +12,6 @@
     BOOL_METRICS,
     METRIC_MAPPING64,
     DistanceMetric,
-
 )
 from ._argkmin import (
     ArgKmin32,
@@ -32,6 +31,7 @@
     RadiusNeighborsClassMode64,
 )
 
+
 def sqeuclidean_row_norms(X, num_threads):
     """Compute the squared euclidean norm of the rows of X in parallel.
 
@@ -81,7 +81,9 @@ def valid_metrics(cls) -> List[str]:
             "hamming",
             *BOOL_METRICS,
         }
-        return sorted(({"sqeuclidean", "precomputed"} | set(METRIC_MAPPING64.keys())) - excluded)
+        return sorted(
+            ({"sqeuclidean", "precomputed"} | set(METRIC_MAPPING64.keys())) - excluded
+        )
 
     @classmethod
     def is_usable_for(cls, X, Y, metric) -> bool:
@@ -105,12 +107,8 @@ def is_usable_for(cls, X, Y, metric) -> bool:
         -------
         True if the dispatcher can be used, else False.
         """
-        
-        #TODO: have the option to use the dispatcher but pass on the compute function
-        # instead call the precomputed distance for (i,j)
-        #question: for the rest of the indices should the compute method be defined?
-        # look at  EuclideanRadiusNeighbors{32,64}
-        if metric == 'precomputed':
+
+        if metric == "precomputed":
             return True
 
         # FIXME: the current Cython implementation is too slow for a large number of
@@ -176,8 +174,7 @@ def compute(
         This method is an abstract class method: it has to be implemented
         for all subclasses.
         """
-      
-      
+        
 
 
 class ArgKmin(BaseDistancesReductionDispatcher):
@@ -201,6 +198,7 @@ def compute(
         Y,
         k,
         metric="euclidean",
+        precomputed_matrix = None,
         chunk_size=None,
         metric_kwargs=None,
         strategy=None,
@@ -430,10 +428,10 @@ def compute(
         for the concrete implementation are therefore freed when this classmethod
         returns.
         """
-        #TODO: to maintain RAII, look at the implementation of the compute method
-        if metric == 'precomputed':
-            return PrecomputedDistanceMatrix.precomputed_distance()
-        
+        # TODO: to maintain RAII, look at the implementation of the compute method
+        # if metric == 'precomputed':
+        #   return PrecomputedDistanceMatrix.precomputed_distance()
+
         if X.dtype == Y.dtype == np.float64:
             return RadiusNeighbors64.compute(
                 X=X,
@@ -591,9 +589,7 @@ def compute(
         for the concrete implementation are therefore freed when this classmethod
         returns.
         """
-        if metric == "precomputed":
-            return PrecomputedDistanceMatrix.precomputed_distance()
-        
+
         if weights not in {"uniform", "distance"}:
             raise ValueError(
                 "Only the 'uniform' or 'distance' weights options are supported"