[UX] Remove --hf-overrides necessity for Qwen3-ASR-ForceAligner (#63274)

Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

docs/models/pooling_models/token_classify.md

@@ -64,11 +64,7 @@ If your model is not in the above list, we will try to automatically convert the
 
 | Architecture                                  | Models              | Inputs            | Example HF Models                          | [LoRA](../../features/lora.md) | [PP](../../serving/parallelism_scaling.md) |
 | --------------------------------------------- | ------------------- | ----------------- | ------------------------------------------ | ------------------------------ | ------------------------------------------ |
-| `Qwen3ASRForcedAlignerForTokenClassification` | Qwen3-ForcedAligner | T + A<sup>+</sup> | `Qwen/Qwen3-ForcedAligner-0.6B` (see note) |                                | ✅︎                                         |
-
-!!! note
-    Forced alignment usage requires `--hf-overrides '{"architectures": ["Qwen3ASRForcedAlignerForTokenClassification"]}'`.
-    Please refer to [examples/pooling/token_classify/forced_alignment_offline.py](../../../examples/pooling/token_classify/forced_alignment_offline.py).
+| `Qwen3ASRForcedAlignerForTokenClassification` | Qwen3-ForcedAligner | T + A<sup>+</sup> | `Qwen/Qwen3-ForcedAligner-0.6B`            |                                | ✅︎                                         |
 
 ### Reward Models
 

examples/pooling/token_classify/forced_alignment_offline.py

@@ -30,7 +30,6 @@ def parse_args():
         model="Qwen/Qwen3-ForcedAligner-0.6B",
         runner="pooling",
         enforce_eager=True,
-        hf_overrides={"architectures": ["Qwen3ASRForcedAlignerForTokenClassification"]},
     )
     return parser.parse_args()
 

tests/models/multimodal/pooling/test_qwen3_asr_forced_aligner.py

@@ -35,11 +35,6 @@ def test_qwen3_forced_aligner(
         dtype=dtype,
         enforce_eager=True,
         max_model_len=512,
-        hf_overrides={
-            "architectures": [
-                "Qwen3ASRForcedAlignerForTokenClassification",
-            ],
-        },
     ) as vllm_model:
         outputs = vllm_model.llm.encode(
             [{"prompt": prompt, "multi_modal_data": {"audio": audio}}],

vllm/model_executor/models/qwen3_asr_forced_aligner.py

@@ -41,9 +41,6 @@ class Qwen3ASRForcedAlignerForTokenClassification(
         llm = LLM(
             model="Qwen/Qwen3-ForcedAligner-0.6B",
             runner="pooling",
-            hf_overrides={
-                "architectures": ["Qwen3ASRForcedAlignerForTokenClassification"]
-            },
         )
         outputs = llm.encode(
             [{"prompt": prompt, "multi_modal_data": {"audio": audio}}],

vllm/transformers_utils/model_arch_config_convertor.py

@@ -651,6 +651,19 @@ class Qwen3_5ModelArchConfigConvertor(ModelArchConfigConvertorBase):
         super().__init__(hf_config, hf_text_config, revision)
 
 
+class Qwen3ASRModelArchConfigConvertor(ModelArchConfigConvertorBase):
+    def __init__(
+        self,
+        hf_config: PreTrainedConfig,
+        hf_text_config: PreTrainedConfig,
+        revision: str | None = None,
+    ):
+        thinker_config = getattr(hf_config, "thinker_config", None)
+        if getattr(thinker_config, "model_type", None) == "qwen3_forced_aligner":
+            hf_config.architectures = ["Qwen3ASRForcedAlignerForTokenClassification"]
+        super().__init__(hf_config, hf_text_config, revision)
+
+
 class Qwen3OmniMoeModelArchConfigConvertor(ModelArchConfigConvertorBase):
     def __init__(
         self,
@@ -865,6 +878,7 @@ MODEL_ARCH_CONFIG_CONVERTORS = {
     "pangu_ultra_moe_mtp": PanguUltraMoeMTPModelArchConfigConvertor,
     "qwen3_5": Qwen3_5ModelArchConfigConvertor,
     "qwen3_5_mtp": Qwen3_5MTPModelArchConfigConvertor,
+    "qwen3_asr": Qwen3ASRModelArchConfigConvertor,
     "qwen3_omni_moe": Qwen3OmniMoeModelArchConfigConvertor,
     "qwen3_next_mtp": Qwen3NextMTPModelArchConfigConvertor,
     "qwen4_exp_mtp": Qwen4ExpMTPModelArchConfigConvertor,