Repair MiniCPM-o and InternVLA scoped E2E paths

tests/model_executor/models/minicpmo_4_5/test_preprocess_embeddings.py

@@ -0,0 +1,43 @@
+"""Regression for text requests sharing the Omni preprocess embedding buffer."""
+
+from types import SimpleNamespace
+
+import torch
+
+from vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni import (
+    MiniCPMO45OmniForConditionalGeneration,
+)
+
+
+def test_text_preprocess_replaces_uninitialized_embedding_buffer() -> None:
+    ids = torch.tensor([4, 7])
+    scratch = torch.zeros((2, 3))
+    expected = torch.tensor([[4.0, 4.0, 4.0], [7.0, 7.0, 7.0]])
+    model = SimpleNamespace(model_stage="llm", get_input_embeddings=lambda input_ids: input_ids[:, None].expand(-1, 3).float())
+
+    _, embeds, _ = MiniCPMO45OmniForConditionalGeneration.preprocess(
+        model,
+        ids,
+        input_embeds=scratch,
+        _omni_input_embeds_precomputed=False,
+    )
+
+    torch.testing.assert_close(embeds, expected)
+
+
+def test_text_preprocess_preserves_precomputed_multimodal_embeddings() -> None:
+    ids = torch.tensor([4, 7])
+    prepared = torch.tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
+    model = SimpleNamespace(
+        model_stage="llm",
+        get_input_embeddings=lambda _input_ids: (_ for _ in ()).throw(AssertionError("must reuse embeddings")),
+    )
+
+    _, embeds, _ = MiniCPMO45OmniForConditionalGeneration.preprocess(
+        model,
+        ids,
+        input_embeds=prepared,
+        _omni_input_embeds_precomputed=True,
+    )
+
+    assert embeds is prepared