Trim trailing frames in offline encode path

tests/test_bugfix_encode_trim_optional.py

@@ -0,0 +1,53 @@
+import importlib.util
+import os
+import unittest
+
+
+class EncodeTrimBugfixOptionalTests(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        enabled = os.getenv("VOXMLX_ENABLE_MLX_RUNTIME_TESTS", "").strip().lower() in {
+            "1",
+            "true",
+            "yes",
+            "on",
+        }
+        if not enabled:
+            raise unittest.SkipTest(
+                "Set VOXMLX_ENABLE_MLX_RUNTIME_TESTS=1 to run MLX runtime optional tests"
+            )
+        if importlib.util.find_spec("mlx") is None:
+            raise unittest.SkipTest("mlx is required for runtime optional tests")
+
+    def test_encode_trims_trailing_frames_not_leading_frames(self):
+        import mlx.core as mx
+        import numpy as np
+
+        from voxmlx.model import VoxtralRealtime
+
+        class _FakeEncoder:
+            def __call__(self, mel):
+                length = int(mel.shape[1])
+                return mx.arange(length, dtype=mx.float32).reshape(1, length, 1)
+
+        class _FakeAdapter:
+            def __call__(self, x):
+                return x
+
+        class _FakeModel:
+            def __init__(self):
+                self.encoder = _FakeEncoder()
+                self.adapter = _FakeAdapter()
+                self.downsample_factor = 3
+
+        fake = _FakeModel()
+        mel = mx.zeros((128, 9), dtype=mx.float32)
+        out = VoxtralRealtime.encode(fake, mel)
+
+        # Expected: keep earliest 6 encoded frames (0..5), grouped by 3.
+        expected = np.array([[0.0, 1.0, 2.0], [3.0, 4.0, 5.0]], dtype=np.float32)
+        np.testing.assert_allclose(np.array(out), expected, atol=0.0, rtol=0.0)
+
+
+if __name__ == "__main__":
+    unittest.main()

voxmlx/model.py

@@ -77,7 +77,7 @@ def encode(self, mel: mx.array) -> mx.array:
         # Truncate T to be even (for conv stride 2)
         T = mel.shape[1]
         if T % 2 != 0:
-            mel = mel[:, 1:]
+            mel = mel[:, :-1]
 
         x = self.encoder(mel)  # [1, T/2, encoder_dim]
         x = x[0]  # [T/2, encoder_dim]
@@ -86,7 +86,7 @@ def encode(self, mel: mx.array) -> mx.array:
         L = x.shape[0]
         remainder = L % self.downsample_factor
         if remainder != 0:
-            x = x[remainder:]
+            x = x[:-remainder]
             L = x.shape[0]
 
         # Reshape: [T/2, 1280] -> [T/8, 5120]