Forward-merge release/26.10 into main

python/cuml/cuml/feature_extraction/_vectorizers.py

@@ -214,11 +214,17 @@ def get_char_ngrams(self, ngram_size, str_series, doc_id_sr):
             tokens = str_series.str.tokenize(self.delimiter)
             del str_series
 
-            padding = Series(self.delimiter).repeat(len(tokens))
+            # tokens keeps the original per-document index (repeated per
+            # token); reset both to a plain range first so the two str.cat()
+            # calls below align positionally instead of by that index.
+            tokens = tokens.reset_index(drop=True)
+            padding = (
+                Series(self.delimiter)
+                .repeat(len(tokens))
+                .reset_index(drop=True)
+            )
             tokens = tokens.str.cat(padding)
-            padding = padding.reset_index(drop=True)
             tokens = padding.str.cat(tokens)
-            tokens = tokens.reset_index(drop=True)
 
             ngram_sr = tokens.str.character_ngrams(n=ngram_size)
 
@@ -235,6 +241,13 @@ def get_char_ngrams(self, ngram_size, str_series, doc_id_sr):
             ngram_count = doc_id_df.groupby("doc_id", sort=True).sum()[
                 "ngram_count"
             ]
+            # A document that tokenizes to zero tokens (e.g. an empty string)
+            # never appears in the groupby above, so its doc_id is silently
+            # missing from ngram_count's index instead of being present with
+            # a count of 0. Reindex onto token_count's full per-document
+            # index so the two stay aligned for get_ngrams' later
+            # ngram_count[not_empty_docs] boolean filter.
+            ngram_count = ngram_count.reindex(token_count.index, fill_value=0)
             return ngram_sr, ngram_count, token_count
 
         if ngram_size == 1:

python/cuml/tests/test_text_feature_extraction.py

@@ -432,6 +432,23 @@ def test_vectorizer_get_feature_names_deprecated(cls):
     np.testing.assert_array_equal(res, model.get_feature_names_out())
 
 
+def test_tfidf_vectorizer_char_wb_ngrams():
+    # Regression test for #8416: get_char_ngrams misaligned padded tokens
+    # across documents once index alignment relied on the original
+    # per-document index instead of a reset range index.
+    vectorizer = TfidfVectorizer(analyzer="char_wb", ngram_range=(2, 6))
+    tfidf_mat = vectorizer.fit_transform(DOCS_GPU)
+
+    ref_vectorizer = SkTfidfVect(analyzer="char_wb", ngram_range=(2, 6))
+    ref = ref_vectorizer.fit_transform(DOCS)
+
+    cp.testing.assert_array_almost_equal(tfidf_mat.todense(), ref.toarray())
+    assert_array_equal(
+        vectorizer.get_feature_names_out(),
+        ref_vectorizer.get_feature_names_out(),
+    )
+
+
 # ----------------------------------------------------------------
 # HashingVectorizer tests
 # ----------------------------------------------------------------