[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

tests/model_executor/models/cosyvoice3/test_cosyvoice3_model_helpers.py

@@ -138,27 +138,6 @@ def _make_sampling_metadata(
     )
 
 
-def test_split_request_ids_uses_seq_token_counts():
-    CosyVoice3Model, _ = _cosyvoice3_model_and_runner()
-    ids = torch.tensor([10, 11, 12, 13, 14], dtype=torch.long)
-    chunks = CosyVoice3Model._split_request_ids(ids, [2, 2, 2])
-    assert [c.tolist() for c in chunks] == [[10, 11], [12, 13], [14]]
-
-
-def test_split_request_ids_honors_single_request_seq_token_counts():
-    CosyVoice3Model, _ = _cosyvoice3_model_and_runner()
-    ids = torch.tensor([10, 11, 12, 13, 14], dtype=torch.long)
-    chunks = CosyVoice3Model._split_request_ids(ids, [3])
-    assert [c.tolist() for c in chunks] == [[10, 11, 12]]
-
-
-def test_sanitize_codec_tokens_filters_out_of_range():
-    model = _make_code2wav_model()
-    raw = torch.tensor([-1, 0, 3, 4, 99], dtype=torch.long)
-    clean = model._sanitize_codec_tokens(raw)
-    assert clean.tolist() == [0, 3]
-
-
 def test_forward_prefers_token_offset_when_present():
     model = _make_code2wav_model()
 
@@ -265,6 +244,31 @@ def test_forward_uses_non_stream_decode_without_chunk_metadata():
     assert len(model.code2wav.forward_streaming_calls) == 0
     call = model.code2wav.forward_calls[0]
     assert call["token"].tolist() == [[0, 1, 2]]
+    assert call["token_offset_tokens"] == 0
+
+
+def test_forward_uses_non_stream_talker_prefill_offset():
+    model = _make_code2wav_model()
+
+    runtime_info = [
+        {
+            "embed": {
+                "speech_token": torch.tensor([[1, 2, 3]], dtype=torch.long),
+                "speech_feat": torch.tensor([[[0.1, 0.2], [0.3, 0.4]]], dtype=torch.float32),
+                "embedding": torch.tensor([[0.5, 0.6]], dtype=torch.float32),
+            },
+            "meta": {"talker_prefill_offset": 3},
+        }
+    ]
+
+    model.forward(
+        input_ids=torch.tensor([0, 1, 2], dtype=torch.long),
+        positions=torch.tensor([0, 1, 2], dtype=torch.long),
+        model_intermediate_buffer=runtime_info,
+        seq_token_counts=[3],
+    )
+
+    assert model.code2wav.forward_calls[0]["token_offset_tokens"] == 3
 
 
 def test_forward_reuses_streaming_cache_state_between_chunks():