diff --git a/camembert_ner_manager.py b/camembert_ner_manager.py
index 627c990..155fd95 100644
--- a/camembert_ner_manager.py
+++ b/camembert_ner_manager.py
@@ -183,8 +183,16 @@ class CamembertNerManager:
         )
         offsets = encoding.pop("offset_mapping")[0]  # (seq_len, 2)
 
-        # Inférence
-        inputs = {k: v for k, v in encoding.items() if k in ("input_ids", "attention_mask")}
+        # Inférence. Certains tokenizers renvoient des tableaux int32 sous
+        # Windows, alors que le graphe CamemBERT ONNX attend des int64.
+        inputs = {}
+        for key, value in encoding.items():
+            if key not in ("input_ids", "attention_mask"):
+                continue
+            array = np.asarray(value)
+            if array.dtype != np.int64:
+                array = array.astype(np.int64)
+            inputs[key] = array
         outputs = self._session.run(None, inputs)
         logits = outputs[0][0]  # (seq_len, num_labels)
 
diff --git a/tests/unit/test_camembert_manager_cache.py b/tests/unit/test_camembert_manager_cache.py
index 1035db1..3e9264a 100644
--- a/tests/unit/test_camembert_manager_cache.py
+++ b/tests/unit/test_camembert_manager_cache.py
@@ -1,5 +1,7 @@
 import json
 
+import numpy as np
+
 
 def test_camembert_load_is_idempotent_and_reuses_process_session(tmp_path, monkeypatch):
     import camembert_ner_manager as module
@@ -53,3 +55,44 @@ def test_camembert_load_is_idempotent_and_reuses_process_session(tmp_path, monke
     assert first.is_loaded()
     assert second.is_loaded()
     assert first._session is second._session
+
+
+def test_camembert_predict_casts_tokenizer_inputs_to_int64():
+    import camembert_ner_manager as module
+
+    captured_inputs = {}
+
+    class FakeTokenizer:
+        def __call__(self, text, **kwargs):
+            return {
+                "input_ids": np.array([[5, 42, 6]], dtype=np.int32),
+                "attention_mask": np.array([[1, 1, 1]], dtype=np.int32),
+                "offset_mapping": np.array([[[0, 0], [0, 5], [0, 0]]], dtype=np.int64),
+            }
+
+    class FakeSession:
+        def run(self, output_names, inputs):
+            captured_inputs.update(inputs)
+            logits = np.array(
+                [[[8.0, 0.0], [0.0, 8.0], [8.0, 0.0]]],
+                dtype=np.float32,
+            )
+            return [logits]
+
+    manager = module.CamembertNerManager()
+    manager._loaded = True
+    manager._tokenizer = FakeTokenizer()
+    manager._session = FakeSession()
+    manager._id2label = {0: "O", 1: "B-PER"}
+
+    entities = manager.predict("Alice")
+
+    assert captured_inputs["input_ids"].dtype == np.int64
+    assert captured_inputs["attention_mask"].dtype == np.int64
+    assert len(entities) == 1
+    assert entities[0]["word"] == "Alice"
+    assert entities[0]["label"] == "PER"
+    assert entities[0]["bio_label"] == "B-PER"
+    assert entities[0]["start"] == 0
+    assert entities[0]["end"] == 5
+    assert entities[0]["score"] > 0.99