Address latest Qwen ASR review feedback

Signed-off-by: CodexIntegrationCheck <codex-integration-check@invalid.example>

Author: CodexIntegrationCheck

nemo_curator/models/asr/qwen_asr.py

@@ -23,14 +23,12 @@
 from __future__ import annotations
 
 import gc
-import inspect
 from copy import deepcopy
 from dataclasses import dataclass, field
 from typing import Any
 
 import numpy as np
 import torch
-import transformers
 from huggingface_hub import snapshot_download
 from loguru import logger
 
@@ -54,30 +52,6 @@ def _qwen_asr_model_cls() -> Any:  # noqa: ANN401
     return Qwen3ASRModel
 
 
-def _patch_transformers_compat() -> None:
-    """Accept qwen-asr's decorator-factory syntax on newer Transformers.
-
-    This matches the compatibility patch used by the nkoluguri Qwen-ASR
-    reference. Newer Transformers exposes ``check_model_inputs`` as a plain
-    decorator, while qwen-asr 0.0.6 still invokes it with parentheses.
-    """
-    try:
-        original = getattr(transformers, "check_model_inputs", None)
-        if original is None:
-            return
-        parameters = list(inspect.signature(original).parameters.values())
-        if parameters and parameters[0].name == "func":
-
-            def compat_check_model_inputs(*args: Any) -> Any:  # noqa: ANN401
-                if args and callable(args[0]):
-                    return original(args[0])
-                return original
-
-            transformers.check_model_inputs = compat_check_model_inputs
-    except Exception:  # noqa: BLE001, S110
-        pass
-
-
 @dataclass
 class QwenASRAdapter:
     """Run vLLM-backed Qwen3-ASR over Curator waveform items.
@@ -171,7 +145,6 @@ def load_model(self, *, num_gpus: int) -> None:
         )
         if model_kwargs["revision"] is None:
             del model_kwargs["revision"]
-        _patch_transformers_compat()
         try:
             self._model = _qwen_asr_model_cls().LLM(**model_kwargs)
         except Exception:

tests/models/asr/test_qwen_asr.py

@@ -27,7 +27,7 @@
 import torch
 
 from nemo_curator.models.asr.base import ASRAdapter
-from nemo_curator.models.asr.qwen_asr import _MIN_SAMPLES, QwenASRAdapter, _patch_transformers_compat
+from nemo_curator.models.asr.qwen_asr import _MIN_SAMPLES, QwenASRAdapter
 from nemo_curator.stages.audio.inference.asr.stage import ASRStage
 from nemo_curator.tasks import AudioTask
 
@@ -107,17 +107,7 @@ def test_qwen_adapter_copies_nested_vllm_kwargs() -> None:
 
 @pytest.mark.parametrize(
     "reserved_key",
-    [
-        "model",
-        "revision",
-        "gpu_memory_utilization",
-        "max_inference_batch_size",
-        "max_new_tokens",
-        "trust_remote_code",
-        "enforce_eager",
-        "enable_prefix_caching",
-        "prefix_caching_hash_algo",
-    ],
+    QwenASRAdapter()._model_owned_vllm_kwargs(),
 )
 def test_qwen_adapter_rejects_adapter_owned_vllm_kwargs(reserved_key: str) -> None:
     adapter = QwenASRAdapter(vllm_kwargs={reserved_key: object()})
@@ -190,34 +180,6 @@ def test_load_model_forwards_additional_vllm_kwargs() -> None:
     assert model_cls.LLM.call_args.kwargs["max_model_len"] == 8192
 
 
-def test_load_model_applies_nkoluguri_transformers_compat_patch() -> None:
-    model_cls = MagicMock()
-    adapter = QwenASRAdapter()
-
-    with (
-        patch("nemo_curator.models.asr.qwen_asr._patch_transformers_compat") as compat,
-        patch("nemo_curator.models.asr.qwen_asr._qwen_asr_model_cls", return_value=model_cls),
-    ):
-        adapter.load_model(num_gpus=1)
-
-    compat.assert_called_once_with()
-
-
-def test_transformers_compat_wraps_plain_decorator_for_factory_syntax() -> None:
-    def plain_decorator(func):  # noqa: ANN001, ANN202
-        return func
-
-    transformers_stub = SimpleNamespace(check_model_inputs=plain_decorator)
-    with patch("nemo_curator.models.asr.qwen_asr.transformers", transformers_stub):
-        _patch_transformers_compat()
-
-    def decorated() -> str:
-        return "ok"
-
-    assert transformers_stub.check_model_inputs()(decorated)() == "ok"
-    assert transformers_stub.check_model_inputs(decorated)() == "ok"
-
-
 def test_load_model_without_qwen_asr_names_required_extras() -> None:
     adapter = QwenASRAdapter()
     with (

tests/utils/test_vllm_utils.py

@@ -32,7 +32,7 @@
 )
 
 
-class TestValidateVllmKwargs:
+class TestVllmKwargs:
     def test_accepts_non_conflicting_kwargs(self) -> None:
         validate_vllm_kwargs(
             {"max_model_len": 8192},
@@ -51,8 +51,6 @@ def test_rejects_conflicts_in_sorted_order(self) -> None:
                 owner_description="adapter-owned arguments",
             )
 
-
-class TestMergeVllmKwargs:
     def test_merges_owned_kwargs_without_mutating_user_kwargs(self) -> None:
         user_kwargs = {"max_model_len": 8192, "compilation_config": {"cudagraph_mode": "NONE"}}