[3/5][core]refactor communication layer: PR 3 of 5, all other models in non async mode (#3719)

Signed-off-by: natureofnature <wzliu@connect.hku.hk>

tests/model_executor/models/cosyvoice3/test_cosyvoice3_components.py

@@ -293,7 +293,13 @@ def test_code2wav_forward_finalizes_hift_tail():
     model = object.__new__(CosyVoice3Code2Wav)
     nn.Module.__init__(model)
     model.hift = DummyHiFT()
-    model._forward_mel = lambda **_: torch.ones((1, 80, 8), dtype=torch.float32)
+    forward_mel_calls = []
+
+    def fake_forward_mel(**kwargs):
+        forward_mel_calls.append(kwargs)
+        return torch.ones((1, 80, 8), dtype=torch.float32)
+
+    model._forward_mel = fake_forward_mel
 
     out = model.forward(
         token=torch.tensor([[1, 2, 3]], dtype=torch.int32),
@@ -304,3 +310,4 @@ def test_code2wav_forward_finalizes_hift_tail():
 
     assert out.shape == (1, 1, 8)
     assert model.hift.finalize_calls == [True]
+    assert forward_mel_calls[0]["token_offset_tokens"] == 0