diff --git a/.github/workflows/test-numpy-compatibility.yml b/.github/workflows/test-numpy-compatibility.yml
index 4a7bcec5c9..6e708f0f45 100644
--- a/.github/workflows/test-numpy-compatibility.yml
+++ b/.github/workflows/test-numpy-compatibility.yml
@@ -9,6 +9,9 @@ on:
       - "tests/test_cli.py"
       - "funasr/auto/auto_model.py"
       - "funasr/utils/load_utils.py"
+      - "funasr/bin/realtime_ws.py"
+      - "tests/test_realtime_ws_service.py"
+      - "tests/test_realtime_ws_hf_overrides.py"
       - "funasr/utils/timestamp_tools.py"
       - "funasr/models/paraformer/model.py"
       - "funasr/models/contextual_paraformer/model.py"
@@ -40,6 +43,9 @@ on:
       - "tests/test_cli.py"
       - "funasr/auto/auto_model.py"
       - "funasr/utils/load_utils.py"
+      - "funasr/bin/realtime_ws.py"
+      - "tests/test_realtime_ws_service.py"
+      - "tests/test_realtime_ws_hf_overrides.py"
       - "funasr/utils/timestamp_tools.py"
       - "funasr/models/paraformer/model.py"
       - "funasr/models/contextual_paraformer/model.py"
@@ -96,7 +102,7 @@ jobs:
       - name: Install CPU runtime and requested NumPy version
         run: |
           python -m pip install torch==2.10.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/cpu
-          python -m pip install "numpy==$NUMPY_VERSION" -e . pytest einops pytorch-wpe --report install-report.json
+          python -m pip install "numpy==$NUMPY_VERSION" -e . pytest websockets einops pytorch-wpe --report install-report.json
       - name: Record resolved environment and validate dependencies
         run: |
           git rev-parse HEAD | tee source-commit.txt
@@ -113,6 +119,8 @@ jobs:
             tests/test_pypi_metadata.py \
             tests/test_sensevoice_word_timestamps.py \
             tests/test_pcm_input_format.py \
+            tests/test_realtime_ws_service.py \
+            tests/test_realtime_ws_hf_overrides.py \
             tests/test_load_audio_bytes.py \
             tests/test_timestamp_tools.py \
             tests/test_punc_model_none.py \
diff --git a/funasr/bin/realtime_ws.py b/funasr/bin/realtime_ws.py
index bce661b762..b6f96b54de 100644
--- a/funasr/bin/realtime_ws.py
+++ b/funasr/bin/realtime_ws.py
@@ -1348,6 +1348,9 @@ def load_models(args):
             gpu_memory_utilization=getattr(args, 'gpu_memory_utilization', 0.8),
             max_model_len=getattr(args, 'max_model_len', 2048),
             enforce_eager=getattr(args, 'enforce_eager', False),
+            vllm_kwargs={"hf_overrides": getattr(args, "hf_overrides", None)}
+            if getattr(args, "hf_overrides", None)
+            else {},
         )
         _vllm_engine = RealtimeBatchingEngine(
             engine,
@@ -1614,12 +1617,25 @@ async def main(args):
         await asyncio.Future()
 
 
+def _parse_hf_overrides(value):
+    overrides = json.loads(value)
+    if not isinstance(overrides, dict):
+        raise argparse.ArgumentTypeError("--hf-overrides must be a JSON object")
+    return overrides
+
+
 def build_arg_parser():
     parser = argparse.ArgumentParser(description="Fun-ASR-Nano Streaming WebSocket Server")
     parser.add_argument("--port", type=int, default=10095)
     parser.add_argument("--model", type=str, default="FunAudioLLM/Fun-ASR-Nano-2512")
     parser.add_argument("--hub", type=str, default="ms", choices=["ms", "hf"])
     parser.add_argument("--device", type=str, default="cuda:0")
+    parser.add_argument(
+        "--hf-overrides",
+        type=_parse_hf_overrides,
+        default=None,
+        help=("JSON object forwarded to the vLLM model configuration."),
+    )
     parser.add_argument(
         "--vad-device",
         type=str,
diff --git a/tests/test_realtime_ws_hf_overrides.py b/tests/test_realtime_ws_hf_overrides.py
new file mode 100644
index 0000000000..f4ff6b7aa6
--- /dev/null
+++ b/tests/test_realtime_ws_hf_overrides.py
@@ -0,0 +1,41 @@
+import pytest
+
+from test_realtime_ws_service import load_service_module
+
+
+@pytest.mark.parametrize("value", ["[]", "[1]", '"model"', "1", "true", "false", "null"])
+def test_cli_rejects_non_object_hf_overrides(value, capsys):
+    parser = load_service_module().build_arg_parser()
+
+    with pytest.raises(SystemExit) as exc:
+        parser.parse_args(["--hf-overrides", value])
+
+    assert exc.value.code == 2
+    assert "JSON object" in capsys.readouterr().err
+
+
+@pytest.mark.parametrize("value", ["{", "{bad}", '{"architectures":}'])
+def test_cli_rejects_malformed_hf_overrides(value):
+    parser = load_service_module().build_arg_parser()
+
+    with pytest.raises(SystemExit) as exc:
+        parser.parse_args(["--hf-overrides", value])
+
+    assert exc.value.code == 2
+
+
+@pytest.mark.parametrize(
+    ("value", "expected"),
+    [
+        ("{}", {}),
+        ('{"architectures":["Qwen2ForCausalLM"]}', {"architectures": ["Qwen2ForCausalLM"]}),
+        ('{"rope_scaling":{"factor":2.0},"use_cache":false}',
+         {"rope_scaling": {"factor": 2.0}, "use_cache": False}),
+    ],
+)
+def test_cli_preserves_hf_override_objects(value, expected):
+    parser = load_service_module().build_arg_parser()
+
+    args = parser.parse_args(["--hf-overrides", value])
+
+    assert args.hf_overrides == expected
diff --git a/tests/test_realtime_ws_service.py b/tests/test_realtime_ws_service.py
index 82292ecd5d..21e5cd416f 100644
--- a/tests/test_realtime_ws_service.py
+++ b/tests/test_realtime_ws_service.py
@@ -119,6 +119,17 @@ def test_cli_defaults_disable_speaker_and_bound_partial_window():
     assert args.log_decode_profile is False
     assert args.vad_device == "cpu"
     assert args.vad_ncpu == 1
+    assert args.hf_overrides is None
+
+
+def test_cli_accepts_hf_overrides_json_object():
+    module = load_service_module()
+
+    args = module.build_arg_parser().parse_args(
+        ["--hf-overrides", '{"architectures":["FunASRForCausalLM"]}']
+    )
+
+    assert args.hf_overrides == {"architectures": ["FunASRForCausalLM"]}
 
 
 def test_cli_accepts_client_endpoint_mode():
@@ -1806,6 +1817,39 @@ def test_load_models_reads_postprocess_hotword_file(monkeypatch, tmp_path):
     assert auto_model_calls == []
 
 
+def test_load_models_forwards_hf_overrides_to_vllm(monkeypatch):
+    module = load_service_module()
+    module._vllm_engine = None
+    module._asr_kwargs = None
+    module._vad_model = None
+    module._spk_model = None
+    vllm_calls = []
+
+    import funasr
+
+    monkeypatch.setattr(funasr, "AutoModel", lambda *args, **kwargs: object())
+    vllm_stub = types.ModuleType("funasr.auto.auto_model_vllm")
+    vllm_stub.AutoModelVLLM = (
+        lambda **kwargs: vllm_calls.append(kwargs) or object()
+    )
+    monkeypatch.setitem(sys.modules, "funasr.auto.auto_model_vllm", vllm_stub)
+
+    args = module.build_arg_parser().parse_args(
+        [
+            "--endpoint-mode",
+            "client",
+            "--hf-overrides",
+            '{"architectures":["FunASRForCausalLM"]}',
+        ]
+    )
+
+    module.load_models(args)
+
+    assert vllm_calls[0]["vllm_kwargs"] == {
+        "hf_overrides": {"architectures": ["FunASRForCausalLM"]}
+    }
+
+
 def test_two_hour_session_keeps_audio_bounded_and_duration_absolute():
     module = load_service_module()
     sample_rate = 10