[bugfix] Shim PreTrainedModel._tp_plan to fix MOSS-TTS-Nano load crash (#4398)

Signed-off-by: akshatvishu <akshatnayak197@gmail.com>

tests/e2e/offline_inference/test_moss_tts_nano_expansion.py

@@ -32,7 +32,6 @@ _OMNI_RUNNER_PARAM = (
 )
 
 pytestmark = [
-    pytest.mark.skip(reason="issue#4361"),
     pytest.mark.full_model,
     pytest.mark.tts,
     pytest.mark.parametrize("omni_runner", [_OMNI_RUNNER_PARAM], indirect=True),

tests/e2e/online_serving/test_moss_tts_nano_expansion.py

@@ -26,7 +26,6 @@ from tests.helpers.stage_config import get_deploy_config_path
 
 # ``omni`` for all tests; 001/002 also get ``full_model`` via ``TestMossTtsNanoFull`` (003 is core+advanced only, no full_model).
 pytestmark = [
-    pytest.mark.skip(reason="issue#4361"),
     pytest.mark.full_model,
     pytest.mark.tts,
 ]

vllm_omni/model_executor/models/moss_tts_nano/modeling_moss_tts_nano.py

@@ -194,7 +194,11 @@ class MossTTSNanoForGeneration(nn.Module):
             tts_dtype = torch.float32
 
         logger.info("Loading MOSS-TTS-Nano LM from %s (dtype=%s)", self.model_path, tts_dtype)
-        from transformers import AutoModelForCausalLM
+        from transformers import AutoModelForCausalLM, PreTrainedModel
+
+        # Shim _tp_plan to prevent TypeError in transformers when loading remote-code models
+        if getattr(PreTrainedModel, "_tp_plan", None) is None:
+            PreTrainedModel._tp_plan = {}
 
         # Two orthogonal trust_remote_code transformers fixes:
         #  - _hf_load_without_tp_warmup: tp_plan=None warmup crash (transformers 5.8.x)