[Bugfix] Defer VoxCPM2 NPU patch to avoid platform import cycle (#8330)

Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>

tests/model_executor/models/minicpmo_4_5/test_cuda_graph_wrapper.py

@@ -231,8 +231,12 @@ def _cfm_inputs(
     time_emb = torch.randn(batch_size, 1, hidden, device=device)
     cnn_cache = torch.randn(depth, batch_size, hidden, 2, device=device)
     att_cache = torch.randn(depth, batch_size, 1, old_att_len, hidden, device=device)
-    cnn_out = torch.empty(depth, batch_size, hidden, 2, device=device)
-    att_out = torch.empty(depth, batch_size, 1, old_att_len + chunk_size, hidden, device=device)
+    # The fake estimator only fills ``att_out[:, :, :, :chunk_size]``. The tail
+    # is the previous cache slot and is not written. ``empty`` leaves it
+    # uninitialized, so eager and replay compare different garbage and CI
+    # fails when that garbage is NaN.
+    cnn_out = torch.zeros(depth, batch_size, hidden, 2, device=device)
+    att_out = torch.zeros(depth, batch_size, 1, old_att_len + chunk_size, hidden, device=device)
     return estimator_input, time_emb, cnn_cache, att_cache, cnn_out, att_out
 
 

tests/platforms/npu/test_voxcpm2_loc_dit_npugraph.py

@@ -76,12 +76,13 @@ def test_model_patch_wraps_init_and_is_idempotent(monkeypatch) -> None:
     assert model.prefix == "talker"
 
 
-def test_npu_platform_registers_patch_and_keeps_worker_cls(monkeypatch) -> None:
+def test_npu_platform_defers_patch_until_device_setup(monkeypatch) -> None:
     pytest.importorskip(
         "vllm_ascend",
         reason="NPU platform registration requires the optional vllm-ascend plugin",
     )
     from vllm_ascend import utils as ascend_utils
+    from vllm_ascend.platform import NPUPlatform
 
     from vllm_omni.platforms.npu import _310p
     from vllm_omni.platforms.npu.models import (
@@ -115,7 +116,21 @@ def test_npu_platform_registers_patch_and_keeps_worker_cls(monkeypatch) -> None:
     worker_cls = NPUOmniPlatform.get_omni_ar_worker_cls()
 
     assert worker_cls == "vllm_omni.platforms.npu.worker.npu_ar_worker.NPUARWorker"
-    assert calls == ["voxcpm2"]
+    assert calls == []
+
+    device = torch.device("npu:0")
+    monkeypatch.setattr(
+        NPUPlatform,
+        "set_device",
+        classmethod(lambda cls, device: calls.append(("set_device", device))),
+    )
+    monkeypatch.setattr(ascend_utils, "enable_custom_op", lambda: calls.append("custom_ops"))
+    monkeypatch.setattr(torch.npu, "config", SimpleNamespace(allow_internal_format=False))
+
+    NPUOmniPlatform.set_device(device)
+
+    assert calls == [("set_device", device), "voxcpm2", "custom_ops"]
+    assert torch.npu.config.allow_internal_format
 
 
 def test_loc_dit_npugraph_supports_wrapped_subclass_and_wraps_once(monkeypatch) -> None:

vllm_omni/platforms/npu/models/voxcpm2_talker.py

@@ -38,8 +38,8 @@ def apply_voxcpm2_talker_patch() -> None:
     if _PATCHED:
         return
 
-    # The model defers resolving ``current_omni_platform`` until construction,
-    # so importing it while NPUOmniPlatform is initialized is cycle-free.
+    # Call this only after platform initialization: the model's transitive
+    # imports resolve ``current_omni_platform``.
     from vllm_omni.model_executor.models.voxcpm2.voxcpm2_talker import (
         VoxCPM2TalkerForConditionalGeneration,
     )

vllm_omni/platforms/npu/platform.py

@@ -68,21 +68,26 @@ class NPUOmniPlatform(OmniPlatform, NPUPlatform):
         from vllm_omni.platforms.npu.models.qwen3_tts_tokenizer_v2 import (
             apply_qwen3_tts_tokenizer_v2_patch,
         )
-        from vllm_omni.platforms.npu.models.voxcpm2_talker import (
-            apply_voxcpm2_talker_patch,
-        )
 
         adapt_patch(is_global_patch=True)
         apply_minicpmo_4_5_code2wav_patch()
         apply_qwen3_tts_patches()
         apply_qwen3_tts_tokenizer_v2_patch()
-        apply_voxcpm2_talker_patch()
         apply_310p_patches()
 
     @classmethod
     def set_device(cls, device: torch.device) -> None:
         super().set_device(device)
 
+        # The model's dependencies resolve current_omni_platform, so importing
+        # it in __init__ would re-enter platform initialization. Install the
+        # patch in the worker, before it constructs the model instead.
+        from vllm_omni.platforms.npu.models.voxcpm2_talker import (
+            apply_voxcpm2_talker_patch,
+        )
+
+        apply_voxcpm2_talker_patch()
+
         # Register vllm_ascend custom ops (torch.ops._C_ascend.*).
         from vllm_ascend.utils import enable_custom_op