feat(audio): batch Qwen3-ASR transcriptions (#5172)

Author: qinxuye

xinference/model/audio/qwen3_asr.py

@@ -12,24 +12,32 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import asyncio
 import logging
 import os
 import tempfile
+from collections import defaultdict
 from typing import TYPE_CHECKING, List, Optional, Tuple
 
+from xoscar import extensible
+
 from ...device_utils import (
     get_available_device,
     get_device_preferred_dtype,
     is_device_available,
 )
+from ...utils import make_hashable
+from ..batch import BatchMixin
 
 if TYPE_CHECKING:
     from .core import AudioModelFamilyV2
 
 logger = logging.getLogger(__name__)
 
+QWEN3_ASR_DEFAULT_BATCH_INTERVAL = 0.1
+
 
-class Qwen3ASRModel:
+class Qwen3ASRModel(BatchMixin):
     def __init__(
         self,
         model_uid: str,
@@ -44,8 +52,22 @@ def __init__(
         self._model_spec = model_spec
         self._device = device
         self._model = None
+        batch_size = kwargs.pop("batch_size", None)
+        batch_interval = kwargs.pop("batch_interval", QWEN3_ASR_DEFAULT_BATCH_INTERVAL)
         self._kwargs = kwargs
 
+        batching_kwargs = {}
+        max_inference_batch_size = kwargs.get("max_inference_batch_size")
+        if batch_size is not None:
+            batching_kwargs["batch_size"] = batch_size
+        elif max_inference_batch_size is not None and int(max_inference_batch_size) > 0:
+            # qwen_asr consumes this as a native from_pretrained argument;
+            # reuse it as Xinference's request-batch limit when no override is set.
+            batching_kwargs["batch_size"] = max_inference_batch_size
+        if batch_interval is not None:
+            batching_kwargs["batch_interval"] = batch_interval
+        BatchMixin.__init__(self, self._batch_transcribe, **batching_kwargs)
+
     @property
     def model_ability(self):
         return self._model_spec.model_ability
@@ -139,7 +161,108 @@ def _extract_text_and_language(self, result) -> Tuple[str, Optional[str]]:
 
         return str(result), None
 
-    def transcriptions(
+    def _format_transcription_result(self, result, response_format: str):
+        text, detected_language = self._extract_text_and_language(result)
+        if response_format == "json":
+            return {"text": text}
+        if response_format == "verbose_json":
+            return {
+                "task": "transcribe",
+                "language": detected_language,
+                "text": text,
+            }
+        raise ValueError(f"Unsupported response format: {response_format}")
+
+    def _run_transcription_batch(
+        self,
+        audios: List[bytes],
+        languages: List[Optional[str]],
+        response_formats: List[str],
+        transcription_kwargs: dict,
+    ) -> List[dict]:
+        assert self._model is not None
+        with tempfile.TemporaryDirectory() as temp_dir:
+            audio_paths = []
+            for i, audio in enumerate(audios):
+                audio_path = os.path.join(temp_dir, f"audio-{i}")
+                with open(audio_path, "wb") as f:
+                    f.write(audio)
+                audio_paths.append(audio_path)
+
+            results = self._model.transcribe(
+                audio=audio_paths,
+                language=languages,
+                **transcription_kwargs,
+            )
+
+        if not isinstance(results, list):
+            raise RuntimeError(
+                f"Qwen3-ASR returned an invalid batch result: {type(results)}"
+            )
+        if len(results) != len(audios):
+            raise RuntimeError(
+                "Qwen3-ASR returned a different number of results than inputs: "
+                f"{len(results)} != {len(audios)}"
+            )
+        return [
+            self._format_transcription_result(result, response_format)
+            for result, response_format in zip(results, response_formats)
+        ]
+
+    @extensible
+    def _batch_transcribe(
+        self,
+        audio: bytes,
+        language: Optional[str],
+        response_format: str,
+        transcription_kwargs: dict,
+    ) -> dict:
+        return self._run_transcription_batch(
+            [audio], [language], [response_format], transcription_kwargs
+        )[0]
+
+    @_batch_transcribe.batch  # type: ignore
+    async def _batch_transcribe(self, args_list, kwargs_list):
+        grouped = defaultdict(
+            lambda: {
+                "audios": [],
+                "languages": [],
+                "response_formats": [],
+                "transcription_kwargs": None,
+                "indices": [],
+            }
+        )
+
+        for index, (args, kwargs) in enumerate(zip(args_list, kwargs_list)):
+            if kwargs:
+                raise TypeError("Qwen3-ASR internal batch call expects positional args")
+            audio, language, response_format, transcription_kwargs = args
+            key = make_hashable(transcription_kwargs)
+            group = grouped[key]
+            group["audios"].append(audio)
+            group["languages"].append(language)
+            group["response_formats"].append(response_format)
+            group["transcription_kwargs"] = transcription_kwargs
+            group["indices"].append(index)
+
+        results_with_indices = []
+        for group in grouped.values():
+            results = await asyncio.to_thread(
+                self._run_transcription_batch,
+                group["audios"],
+                group["languages"],
+                group["response_formats"],
+                group["transcription_kwargs"],
+            )
+            results_with_indices.extend(zip(group["indices"], results))
+
+        results_with_indices.sort(key=lambda item: item[0])
+        return [result for _, result in results_with_indices]
+
+    def _get_batch_size(self, *args, **kwargs) -> int:
+        return 1
+
+    async def transcriptions(
         self,
         audio: bytes,
         language: Optional[str] = None,
@@ -163,25 +286,12 @@ def transcriptions(
             logger.warning(
                 "Prompt for Qwen3-ASR transcriptions will be ignored: %s", prompt
             )
+        if response_format not in ("json", "verbose_json"):
+            raise ValueError(f"Unsupported response format: {response_format}")
 
         kw = dict(getattr(self._model_spec, "default_transcription_config", None) or {})
         kw.update(kwargs)
-
-        with tempfile.NamedTemporaryFile(buffering=0) as f:
-            f.write(audio)
-            assert self._model is not None
-            result = self._model.transcribe(audio=f.name, language=language, **kw)
-            text, detected_language = self._extract_text_and_language(result)
-
-        if response_format == "json":
-            return {"text": text}
-        if response_format == "verbose_json":
-            return {
-                "task": "transcribe",
-                "language": detected_language,
-                "text": text,
-            }
-        raise ValueError(f"Unsupported response format: {response_format}")
+        return await self._batch_transcribe(audio, language, response_format, kw)
 
     def translations(
         self,

xinference/model/audio/tests/test_qwen3_asr.py

@@ -0,0 +1,169 @@
+# Copyright 2022-2026 Xinference Holdings Pte. Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import asyncio
+from contextlib import suppress
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+from ..qwen3_asr import Qwen3ASRModel
+
+
+class _FakeQwenASR:
+    def __init__(self):
+        self.calls = []
+
+    def transcribe(self, audio, language, **kwargs):
+        payloads = [Path(path).read_bytes().decode() for path in audio]
+        self.calls.append(
+            {"payloads": payloads, "language": list(language), "kwargs": kwargs}
+        )
+        return [
+            SimpleNamespace(text=payload, language=lang or "Detected")
+            for payload, lang in zip(payloads, language)
+        ]
+
+
+def _new_model(batch_size=8, batch_interval=0.01, **kwargs):
+    model_spec = SimpleNamespace(
+        model_name="Qwen3-ASR-0.6B",
+        model_ability=["audio2text"],
+        default_transcription_config={},
+    )
+    model = Qwen3ASRModel(
+        "qwen3-asr",
+        "/unused",
+        model_spec,
+        batch_size=batch_size,
+        batch_interval=batch_interval,
+        **kwargs,
+    )
+    model._model = _FakeQwenASR()
+    return model
+
+
+def test_qwen3_asr_uses_audio_batch_interval_default():
+    model_spec = SimpleNamespace(
+        model_name="Qwen3-ASR-0.6B",
+        model_ability=["audio2text"],
+        default_transcription_config={},
+    )
+
+    model = Qwen3ASRModel("default", "/unused", model_spec)
+    overridden = Qwen3ASRModel("overridden", "/unused", model_spec, batch_interval=0.02)
+
+    assert model.batch_interval == pytest.approx(0.1)
+    assert overridden.batch_interval == pytest.approx(0.02)
+
+
+async def _shutdown_batch_processor(model):
+    task = model._process_batch_task
+    if task is not None:
+        task.cancel()
+        with suppress(asyncio.CancelledError):
+            await task
+
+
+@pytest.mark.asyncio
+async def test_qwen3_asr_batches_requests_and_preserves_order():
+    model = _new_model()
+    try:
+        results = await asyncio.gather(
+            model.transcriptions(b"first", language="Chinese"),
+            model.transcriptions(
+                b"second", language="English", response_format="verbose_json"
+            ),
+            model.transcriptions(b"third"),
+        )
+
+        assert results == [
+            {"text": "first"},
+            {"task": "transcribe", "language": "English", "text": "second"},
+            {"text": "third"},
+        ]
+        assert model._model.calls == [
+            {
+                "payloads": ["first", "second", "third"],
+                "language": ["Chinese", "English", None],
+                "kwargs": {},
+            }
+        ]
+    finally:
+        await _shutdown_batch_processor(model)
+
+
+@pytest.mark.asyncio
+async def test_qwen3_asr_respects_batch_size():
+    model = _new_model(batch_size=2)
+    try:
+        results = await asyncio.gather(
+            model.transcriptions(b"first"),
+            model.transcriptions(b"second"),
+            model.transcriptions(b"third"),
+        )
+
+        assert results == [
+            {"text": "first"},
+            {"text": "second"},
+            {"text": "third"},
+        ]
+        assert [call["payloads"] for call in model._model.calls] == [
+            ["first", "second"],
+            ["third"],
+        ]
+    finally:
+        await _shutdown_batch_processor(model)
+
+
+@pytest.mark.asyncio
+async def test_qwen3_asr_groups_model_kwargs_and_survives_cancellation():
+    model = _new_model()
+    try:
+        cancelled = asyncio.create_task(
+            model.transcriptions(b"cancelled", language="Chinese", beam_size=1)
+        )
+        kept = asyncio.create_task(
+            model.transcriptions(b"kept", language="English", beam_size=1)
+        )
+        other_group = asyncio.create_task(
+            model.transcriptions(b"other", language="Japanese", beam_size=2)
+        )
+        await asyncio.sleep(0)
+        cancelled.cancel()
+
+        with pytest.raises(asyncio.CancelledError):
+            await cancelled
+        assert await kept == {"text": "kept"}
+        assert await other_group == {"text": "other"}
+
+        # A cancelled waiter must not stop the long-lived batch processor.
+        assert await model.transcriptions(b"after") == {"text": "after"}
+
+        assert model._model.calls[:2] == [
+            {
+                "payloads": ["kept"],
+                "language": ["English"],
+                "kwargs": {"beam_size": 1},
+            },
+            {
+                "payloads": ["other"],
+                "language": ["Japanese"],
+                "kwargs": {"beam_size": 2},
+            },
+        ]
+        assert model._model.calls[2]["payloads"] == ["after"]
+    finally:
+        await _shutdown_batch_processor(model)