xinference 1.2.0__py3-none-any.whl → 1.2.2__py3-none-any.whl

xinference/model/audio/kokoro.py

@@ -0,0 +1,139 @@
+# Copyright 2022-2023 XProbe Inc.
+#
+# 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 logging
+from io import BytesIO
+from typing import TYPE_CHECKING, Optional
+
+import numpy as np
+
+from ...device_utils import get_available_device, is_device_available
+
+if TYPE_CHECKING:
+    from .core import AudioModelFamilyV1
+
+logger = logging.getLogger(__name__)
+
+
+class KokoroModel:
+    # The available voices, should keep sync with https://huggingface.co/hexgrad/Kokoro-82M/tree/main/voices
+    VOICES = [
+        "af_alloy",
+        "af_aoede",
+        "af_bella",
+        "af_jessica",
+        "af_kore",
+        "af_nicole",
+        "af_nova",
+        "af_river",
+        "af_sarah",
+        "af_sky",
+        "am_adam",
+        "am_echo",
+        "am_eric",
+        "am_fenrir",
+        "am_liam",
+        "am_michael",
+        "am_onyx",
+        "am_puck",
+        "bf_alice",
+        "bf_emma",
+        "bf_isabella",
+        "bf_lily",
+        "bm_daniel",
+        "bm_fable",
+        "bm_george",
+        "bm_lewis",
+    ]
+
+    def __init__(
+        self,
+        model_uid: str,
+        model_path: str,
+        model_spec: "AudioModelFamilyV1",
+        device: Optional[str] = None,
+        **kwargs,
+    ):
+        self._model_uid = model_uid
+        self._model_path = model_path
+        self._model_spec = model_spec
+        self._device = device
+        self._model = None
+        self._kwargs = kwargs
+
+    @property
+    def model_ability(self):
+        return self._model_spec.model_ability
+
+    def load(self):
+        if self._device is None:
+            self._device = get_available_device()
+        else:
+            if not is_device_available(self._device):
+                raise ValueError(f"Device {self._device} is not available!")
+
+        import os
+
+        from kokoro import KModel, KPipeline
+
+        config_path = os.path.join(self._model_path, "config.json")
+        model_path = os.path.join(self._model_path, "kokoro-v1_0.pth")
+        # LANG_CODES = dict(
+        #     a='American English',
+        #     b='British English',
+        # )
+        lang_code = self._kwargs.get("lang_code", "a")
+        self._model = KPipeline(
+            lang_code=lang_code,
+            model=KModel(config=config_path, model=model_path),
+            device=self._device,
+        )
+
+    def speech(
+        self,
+        input: str,
+        voice: str,
+        response_format: str = "mp3",
+        speed: float = 1.0,
+        stream: bool = False,
+        **kwargs,
+    ):
+        import soundfile
+
+        if stream:
+            raise Exception("Kokoro does not support stream mode.")
+        assert self._model is not None
+        if not voice:
+            voice = next(iter(self.VOICES))
+            logger.info("Auto select speaker: %s", voice)
+        elif not voice.endswith(".pt") and voice not in self.VOICES:
+            raise ValueError(
+                f"Invalid voice: {voice}, available speakers: {self.VOICES}"
+            )
+        else:
+            logger.info("Using custom voice pt: %s", voice)
+        logger.info("Speech kwargs: %s", kwargs)
+        generator = self._model(text=input, voice=voice, speed=speed, **kwargs)
+        results = list(generator)
+        audio = np.concatenate([r[2] for r in results])
+        # Save the generated audio
+        with BytesIO() as out:
+            with soundfile.SoundFile(
+                out,
+                "w",
+                24000,
+                1,
+                format=response_format.upper(),
+            ) as f:
+                f.write(audio)
+            return out.getvalue()

xinference/model/audio/melotts.py

@@ -0,0 +1,110 @@
+# Copyright 2022-2023 XProbe Inc.
+#
+# 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 logging
+from io import BytesIO
+from typing import TYPE_CHECKING, Optional
+
+from ...device_utils import get_available_device, is_device_available
+
+if TYPE_CHECKING:
+    from .core import AudioModelFamilyV1
+
+logger = logging.getLogger(__name__)
+
+
+class MeloTTSModel:
+    def __init__(
+        self,
+        model_uid: str,
+        model_path: str,
+        model_spec: "AudioModelFamilyV1",
+        device: Optional[str] = None,
+        **kwargs,
+    ):
+        self._model_uid = model_uid
+        self._model_path = model_path
+        self._model_spec = model_spec
+        self._device = device
+        self._model = None
+        self._kwargs = kwargs
+
+    @property
+    def model_ability(self):
+        return self._model_spec.model_ability
+
+    def load(self):
+        if self._device is None:
+            self._device = get_available_device()
+        else:
+            if not is_device_available(self._device):
+                raise ValueError(f"Device {self._device} is not available!")
+
+        import os
+        import sys
+
+        import nltk
+
+        # English language requires download averaged_perceptron_tagger_eng
+        nltk.download("averaged_perceptron_tagger_eng")
+
+        # The yaml config loaded from model has hard-coded the import paths. please refer to: load_hyperpyyaml
+        sys.path.insert(0, os.path.join(os.path.dirname(__file__), "../../thirdparty"))
+
+        from melo.api import TTS
+
+        config_path = os.path.join(self._model_path, "config.json")
+        ckpt_path = os.path.join(self._model_path, "checkpoint.pth")
+        self._model = TTS(
+            language=self._model_spec.language,
+            device=self._device,
+            config_path=config_path,
+            ckpt_path=ckpt_path,
+        )
+
+    def speech(
+        self,
+        input: str,
+        voice: str,
+        response_format: str = "mp3",
+        speed: float = 1.0,
+        stream: bool = False,
+        **kwargs,
+    ):
+        import soundfile
+
+        if stream:
+            raise Exception("MeloTTS does not support stream mode.")
+        assert self._model is not None
+        speaker_ids = self._model.hps.data.spk2id
+        if not voice:
+            voice = next(iter(speaker_ids.keys()))
+            logger.info("Auto select speaker: %s", voice)
+        elif voice not in speaker_ids:
+            raise ValueError(
+                f"Invalid voice: {voice}, available speakers: {speaker_ids}"
+            )
+        audio = self._model.tts_to_file(
+            text=input, speaker_id=speaker_ids[voice], speed=speed, **kwargs
+        )
+        # Save the generated audio
+        with BytesIO() as out:
+            with soundfile.SoundFile(
+                out,
+                "w",
+                self._model.hps.data.sampling_rate,
+                1,
+                format=response_format.upper(),
+            ) as f:
+                f.write(audio)
+            return out.getvalue()

xinference/model/audio/model_spec.json

@@ -266,5 +266,85 @@
     "model_revision": "7642bb232e3fcacf92c51c786edebb8624da6b93",
     "model_ability": "text-to-audio",
     "multilingual": true
+  },
+  {
+    "model_name": "MeloTTS-English",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-English",
+    "model_revision": "bb4fb7346d566d277ba8c8c7dbfdf6786139b8ef",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "EN"
+  },
+  {
+    "model_name": "MeloTTS-English-v2",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-English-v2",
+    "model_revision": "a53e3509c4ee4ff16d79272feb2474ff864e18f3",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "EN"
+  },
+  {
+    "model_name": "MeloTTS-English-v3",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-English-v3",
+    "model_revision": "f7c4a35392c0e9be24a755f1edb4c3f63040f759",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "EN"
+  },
+  {
+    "model_name": "MeloTTS-French",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-French",
+    "model_revision": "1e9bf590262392d8bffb679b0a3b0c16b0f9fdaf",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "FR"
+  },
+  {
+    "model_name": "MeloTTS-Japanese",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-Japanese",
+    "model_revision": "367f8795464b531b4e97c1515bddfc1243e60891",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "JP"
+  },
+  {
+    "model_name": "MeloTTS-Spanish",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-Spanish",
+    "model_revision": "dbb5496df39d11a66c1d5f5a9ca357c3c9fb95fb",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "ES"
+  },
+  {
+    "model_name": "MeloTTS-Chinese",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-Chinese",
+    "model_revision": "af5d207a364ea4208c6f589c89f57f88414bdd16",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "ZH"
+  },
+  {
+    "model_name": "MeloTTS-Korean",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-Korean",
+    "model_revision": "0207e5adfc90129a51b6b03d89be6d84360ed323",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "KR"
+  },
+  {
+    "model_name": "Kokoro-82M",
+    "model_family": "Kokoro",
+    "model_id": "hexgrad/Kokoro-82M",
+    "model_revision": "7a29fcdf8e997bac6d6f5f6f0c2f0b92912f6102",
+    "model_ability": "text-to-audio",
+    "multilingual": true
   }
 ]

xinference/model/audio/model_spec_modelscope.json

@@ -17,6 +17,15 @@
     "model_ability": "audio-to-text",
     "multilingual": true
   },
+  {
+    "model_name": "Belle-whisper-large-v3-zh",
+    "model_family": "whisper",
+    "model_hub": "modelscope",
+    "model_id": "Xorbits/Belle-whisper-large-v3-zh",
+    "model_revision": "master",
+    "model_ability": "audio-to-text",
+    "multilingual": false
+  },
   {
     "model_name": "SenseVoiceSmall",
     "model_family": "funasr",
@@ -91,5 +100,14 @@
     "model_revision": "master",
     "model_ability": "text-to-audio",
     "multilingual": true
+  },
+  {
+    "model_name": "Kokoro-82M",
+    "model_family": "Kokoro",
+    "model_hub": "modelscope",
+    "model_id": "AI-ModelScope/Kokoro-82M",
+    "model_revision": "master",
+    "model_ability": "text-to-audio",
+    "multilingual": true
   }
 ]

xinference/model/audio/whisper.py

@@ -13,9 +13,12 @@
 # limitations under the License.
 import logging
 import os
+import typing
 from glob import glob
 from typing import TYPE_CHECKING, Dict, List, Optional, Union
 
+from typing_extensions import TypedDict
+
 from ...device_utils import (
     get_available_device,
     get_device_preferred_dtype,
@@ -28,6 +31,13 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 
+class WhisperModelConfig(TypedDict, total=False):
+    chunk_length_s: Optional[float]
+    stride_length_s: Optional[float]
+    return_timestamps: Optional[bool]
+    batch_size: Optional[int]
+
+
 class WhisperModel:
     def __init__(
         self,
@@ -35,6 +45,7 @@ class WhisperModel:
         model_path: str,
         model_spec: "AudioModelFamilyV1",
         device: Optional[str] = None,
+        max_new_tokens: Optional[int] = 128,
         **kwargs,
     ):
         self._model_uid = model_uid
@@ -42,7 +53,21 @@ class WhisperModel:
         self._model_spec = model_spec
         self._device = device
         self._model = None
-        self._kwargs = kwargs
+        self._max_new_tokens = max_new_tokens
+        self._model_config: WhisperModelConfig = self._sanitize_model_config(
+            typing.cast(WhisperModelConfig, kwargs)
+        )
+
+    def _sanitize_model_config(
+        self, model_config: Optional[WhisperModelConfig]
+    ) -> WhisperModelConfig:
+        if model_config is None:
+            model_config = WhisperModelConfig()
+        model_config.setdefault("chunk_length_s", 30)
+        model_config.setdefault("stride_length_s", None)
+        model_config.setdefault("return_timestamps", False)
+        model_config.setdefault("batch_size", 16)
+        return model_config
 
     @property
     def model_ability(self):
@@ -75,10 +100,10 @@ class WhisperModel:
             model=model,
             tokenizer=processor.tokenizer,
             feature_extractor=processor.feature_extractor,
-            max_new_tokens=128,
-            chunk_length_s=30,
-            batch_size=16,
-            return_timestamps=False,
+            chunk_length_s=self._model_config.get("chunk_length_s"),
+            stride_length_s=self._model_config.get("stride_length_s"),
+            return_timestamps=self._model_config.get("return_timestamps"),
+            batch_size=self._model_config.get("batch_size"),
             torch_dtype=torch_dtype,
             device=self._device,
         )
@@ -185,13 +210,13 @@ class WhisperModel:
             logger.warning(
                 "Prompt for whisper transcriptions will be ignored: %s", prompt
             )
+        generate_kwargs = {"max_new_tokens": self._max_new_tokens, "task": "transcribe"}
+        if language is not None:
+            generate_kwargs["language"] = language
+
         return self._call_model(
             audio=audio,
-            generate_kwargs=(
-                {"language": language, "task": "transcribe"}
-                if language is not None
-                else {"task": "transcribe"}
-            ),
+            generate_kwargs=generate_kwargs,
             response_format=response_format,
             temperature=temperature,
             timestamp_granularities=timestamp_granularities,

xinference/model/llm/llama_cpp/core.py

@@ -28,7 +28,7 @@ from ....types import (
 )
 from ..core import LLM
 from ..llm_family import LLMFamilyV1, LLMSpecV1
-from ..utils import QWEN_TOOL_CALL_FAMILY, ChatModelMixin
+from ..utils import DEEPSEEK_TOOL_CALL_FAMILY, QWEN_TOOL_CALL_FAMILY, ChatModelMixin
 
 logger = logging.getLogger(__name__)
 
@@ -123,18 +123,22 @@ class LlamaCppModel(LLM):
 
             raise ImportError(f"{error_message}\n\n{''.join(installation_guide)}")
 
-        # handle legacy cache.
-        model_path = os.path.realpath(
-            os.path.join(
-                self.model_path,
-                self.model_spec.model_file_name_template.format(
-                    quantization=self.quantization
-                ),
+        if os.path.isfile(self.model_path):
+            # mostly passed from --model_path
+            model_path = os.path.realpath(self.model_path)
+        else:
+            # handle legacy cache.
+            model_path = os.path.realpath(
+                os.path.join(
+                    self.model_path,
+                    self.model_spec.model_file_name_template.format(
+                        quantization=self.quantization
+                    ),
+                )
             )
-        )
-        legacy_model_file_path = os.path.join(self.model_path, "model.bin")
-        if os.path.exists(legacy_model_file_path):
-            model_path = legacy_model_file_path
+            legacy_model_file_path = os.path.join(self.model_path, "model.bin")
+            if os.path.exists(legacy_model_file_path):
+                model_path = legacy_model_file_path
 
         try:
             self._llm = Llama(
@@ -272,8 +276,11 @@ class LlamaCppChatModel(LlamaCppModel, ChatModelMixin):
         model_family = self.model_family.model_family or self.model_family.model_name
         tools = generate_config.pop("tools", []) if generate_config else None
         full_context_kwargs = {}
-        if tools and model_family in QWEN_TOOL_CALL_FAMILY:
-            full_context_kwargs["tools"] = tools
+        if tools:
+            if model_family in QWEN_TOOL_CALL_FAMILY:
+                full_context_kwargs["tools"] = tools
+            elif model_family in DEEPSEEK_TOOL_CALL_FAMILY:
+                self._tools_to_messages_for_deepseek(messages, tools)
         assert self.model_family.chat_template is not None
         full_prompt = self.get_full_context(
             messages, self.model_family.chat_template, **full_context_kwargs