xinference 1.0.0__py3-none-any.whl → 1.1.0__py3-none-any.whl

xinference/model/audio/f5tts.py

@@ -0,0 +1,195 @@
+# 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
+import os
+import re
+from io import BytesIO
+from typing import TYPE_CHECKING, Optional
+
+if TYPE_CHECKING:
+    from .core import AudioModelFamilyV1
+
+logger = logging.getLogger(__name__)
+
+
+class F5TTSModel:
+    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._vocoder = None
+        self._kwargs = kwargs
+
+    @property
+    def model_ability(self):
+        return self._model_spec.model_ability
+
+    def load(self):
+        import os
+        import sys
+
+        # 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 f5_tts.infer.utils_infer import load_model, load_vocoder
+        from f5_tts.model import DiT
+
+        vocoder_name = self._kwargs.get("vocoder_name", "vocos")
+        vocoder_path = self._kwargs.get("vocoder_path")
+
+        if vocoder_name not in ["vocos", "bigvgan"]:
+            raise Exception(f"Unsupported vocoder name: {vocoder_name}")
+
+        if vocoder_path is not None:
+            self._vocoder = load_vocoder(
+                vocoder_name=vocoder_name, is_local=True, local_path=vocoder_path
+            )
+        else:
+            self._vocoder = load_vocoder(vocoder_name=vocoder_name, is_local=False)
+
+        model_cls = DiT
+        model_cfg = dict(
+            dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4
+        )
+        if vocoder_name == "vocos":
+            exp_name = "F5TTS_Base"
+            ckpt_step = 1200000
+        elif vocoder_name == "bigvgan":
+            exp_name = "F5TTS_Base_bigvgan"
+            ckpt_step = 1250000
+        else:
+            assert False
+        ckpt_file = os.path.join(
+            self._model_path, exp_name, f"model_{ckpt_step}.safetensors"
+        )
+        logger.info(f"Loading %s...", ckpt_file)
+        self._model = load_model(
+            model_cls, model_cfg, ckpt_file, mel_spec_type=vocoder_name
+        )
+
+    def _infer(self, ref_audio, ref_text, text_gen, model_obj, mel_spec_type, speed):
+        import numpy as np
+        from f5_tts.infer.utils_infer import infer_process, preprocess_ref_audio_text
+
+        config = {}
+        main_voice = {"ref_audio": ref_audio, "ref_text": ref_text}
+        if "voices" not in config:
+            voices = {"main": main_voice}
+        else:
+            voices = config["voices"]
+            voices["main"] = main_voice
+        for voice in voices:
+            (
+                voices[voice]["ref_audio"],
+                voices[voice]["ref_text"],
+            ) = preprocess_ref_audio_text(
+                voices[voice]["ref_audio"], voices[voice]["ref_text"]
+            )
+            print("Voice:", voice)
+            print("Ref_audio:", voices[voice]["ref_audio"])
+            print("Ref_text:", voices[voice]["ref_text"])
+
+        final_sample_rate = None
+        generated_audio_segments = []
+        reg1 = r"(?=\[\w+\])"
+        chunks = re.split(reg1, text_gen)
+        reg2 = r"\[(\w+)\]"
+        for text in chunks:
+            if not text.strip():
+                continue
+            match = re.match(reg2, text)
+            if match:
+                voice = match[1]
+            else:
+                print("No voice tag found, using main.")
+                voice = "main"
+            if voice not in voices:
+                print(f"Voice {voice} not found, using main.")
+                voice = "main"
+            text = re.sub(reg2, "", text)
+            gen_text = text.strip()
+            ref_audio = voices[voice]["ref_audio"]
+            ref_text = voices[voice]["ref_text"]
+            print(f"Voice: {voice}")
+            audio, final_sample_rate, spectragram = infer_process(
+                ref_audio,
+                ref_text,
+                gen_text,
+                model_obj,
+                self._vocoder,
+                mel_spec_type=mel_spec_type,
+                speed=speed,
+            )
+            generated_audio_segments.append(audio)
+
+        if generated_audio_segments:
+            final_wave = np.concatenate(generated_audio_segments)
+            return final_sample_rate, final_wave
+        return None, None
+
+    def speech(
+        self,
+        input: str,
+        voice: str,
+        response_format: str = "mp3",
+        speed: float = 1.0,
+        stream: bool = False,
+        **kwargs,
+    ):
+        import f5_tts
+        import soundfile
+        import tomli
+
+        if stream:
+            raise Exception("F5-TTS does not support stream generation.")
+
+        prompt_speech: Optional[bytes] = kwargs.pop("prompt_speech", None)
+        prompt_text: Optional[str] = kwargs.pop("prompt_text", None)
+
+        if prompt_speech is None:
+            base = os.path.dirname(f5_tts.__file__)
+            config = os.path.join(base, "infer/examples/basic/basic.toml")
+            with open(config, "rb") as f:
+                config_dict = tomli.load(f)
+            prompt_speech = os.path.join(base, config_dict["ref_audio"])
+            prompt_text = config_dict["ref_text"]
+
+        assert self._model is not None
+        vocoder_name = self._kwargs.get("vocoder_name", "vocos")
+        sample_rate, wav = self._infer(
+            ref_audio=prompt_speech,
+            ref_text=prompt_text,
+            text_gen=input,
+            model_obj=self._model,
+            mel_spec_type=vocoder_name,
+            speed=speed,
+        )
+
+        # Save the generated audio
+        with BytesIO() as out:
+            with soundfile.SoundFile(
+                out, "w", sample_rate, 1, format=response_format.upper()
+            ) as f:
+                f.write(wav)
+            return out.getvalue()

xinference/model/audio/fish_speech.py

@@ -81,12 +81,14 @@ class FishSpeechModel:
             if not is_device_available(self._device):
                 raise ValueError(f"Device {self._device} is not available!")
 
-        logger.info("Loading Llama model...")
+        enable_compile = self._kwargs.get("compile", False)
+        precision = self._kwargs.get("precision", torch.bfloat16)
+        logger.info("Loading Llama model, compile=%s...", enable_compile)
         self._llama_queue = launch_thread_safe_queue(
             checkpoint_path=self._model_path,
             device=self._device,
-            precision=torch.bfloat16,
-            compile=False,
+            precision=precision,
+            compile=enable_compile,
         )
         logger.info("Llama model loaded, loading VQ-GAN model...")
 
@@ -112,9 +114,10 @@ class FishSpeechModel:
         top_p,
         repetition_penalty,
         temperature,
+        seed="0",
         streaming=False,
     ):
-        from fish_speech.utils import autocast_exclude_mps
+        from fish_speech.utils import autocast_exclude_mps, set_seed
         from tools.api import decode_vq_tokens, encode_reference
         from tools.llama.generate import (
             GenerateRequest,
@@ -122,6 +125,11 @@ class FishSpeechModel:
             WrappedGenerateResponse,
         )
 
+        seed = int(seed)
+        if seed != 0:
+            set_seed(seed)
+            logger.warning(f"set seed: {seed}")
+
         # Parse reference audio aka prompt
         prompt_tokens = encode_reference(
             decoder_model=self._model,
@@ -137,7 +145,7 @@ class FishSpeechModel:
             top_p=top_p,
             repetition_penalty=repetition_penalty,
             temperature=temperature,
-            compile=False,
+            compile=self._kwargs.get("compile", False),
             iterative_prompt=chunk_length > 0,
             chunk_length=chunk_length,
             max_length=2048,
@@ -153,22 +161,20 @@ class FishSpeechModel:
             )
         )
 
-        if streaming:
-            yield wav_chunk_header(), None, None
-
         segments = []
 
         while True:
-            result: WrappedGenerateResponse = response_queue.get()  # type: ignore
+            result: WrappedGenerateResponse = response_queue.get()
             if result.status == "error":
-                raise Exception(str(result.response))
+                raise result.response
 
-            result: GenerateResponse = result.response  # type: ignore
+            result: GenerateResponse = result.response
             if result.action == "next":
                 break
 
             with autocast_exclude_mps(
-                device_type=self._model.device.type, dtype=torch.bfloat16
+                device_type=self._model.device.type,
+                dtype=self._kwargs.get("precision", torch.bfloat16),
             ):
                 fake_audios = decode_vq_tokens(
                     decoder_model=self._model,
@@ -179,7 +185,7 @@ class FishSpeechModel:
             segments.append(fake_audios)
 
             if streaming:
-                yield (fake_audios * 32768).astype(np.int16).tobytes(), None, None
+                yield fake_audios, None, None
 
         if len(segments) == 0:
             raise Exception("No audio generated, please check the input text.")
@@ -204,29 +210,59 @@ class FishSpeechModel:
         logger.warning("Fish speech does not support setting voice: %s.", voice)
         if speed != 1.0:
             logger.warning("Fish speech does not support setting speed: %s.", speed)
-        if stream is True:
-            logger.warning("stream mode is not implemented.")
         import torchaudio
 
-        result = list(
-            self._inference(
-                text=input,
-                enable_reference_audio=False,
-                reference_audio=None,
-                reference_text=kwargs.get("reference_text", ""),
-                max_new_tokens=kwargs.get("max_new_tokens", 1024),
-                chunk_length=kwargs.get("chunk_length", 200),
-                top_p=kwargs.get("top_p", 0.7),
-                repetition_penalty=kwargs.get("repetition_penalty", 1.2),
-                temperature=kwargs.get("temperature", 0.7),
-            )
+        prompt_speech = kwargs.get("prompt_speech")
+        prompt_text = kwargs.get("prompt_text", kwargs.get("reference_text", ""))
+        result = self._inference(
+            text=input,
+            enable_reference_audio=kwargs.get(
+                "enable_reference_audio", prompt_speech is not None
+            ),
+            reference_audio=prompt_speech,
+            reference_text=prompt_text,
+            max_new_tokens=kwargs.get("max_new_tokens", 1024),
+            chunk_length=kwargs.get("chunk_length", 200),
+            top_p=kwargs.get("top_p", 0.7),
+            repetition_penalty=kwargs.get("repetition_penalty", 1.2),
+            temperature=kwargs.get("temperature", 0.7),
+            streaming=stream,
         )
-        sample_rate, audio = result[0][1]
-        audio = np.array([audio])
 
-        # Save the generated audio
-        with BytesIO() as out:
-            torchaudio.save(
-                out, torch.from_numpy(audio), sample_rate, format=response_format
-            )
-            return out.getvalue()
+        if stream:
+
+            def _stream_generator():
+                with BytesIO() as out:
+                    writer = torchaudio.io.StreamWriter(out, format=response_format)
+                    writer.add_audio_stream(
+                        sample_rate=self._model.spec_transform.sample_rate,
+                        num_channels=1,
+                    )
+                    i = 0
+                    last_pos = 0
+                    with writer.open():
+                        for chunk in result:
+                            chunk = chunk[0]
+                            if chunk is not None:
+                                chunk = chunk.reshape((chunk.shape[0], 1))
+                                trans_chunk = torch.from_numpy(chunk)
+                                writer.write_audio_chunk(i, trans_chunk)
+                                new_last_pos = out.tell()
+                                if new_last_pos != last_pos:
+                                    out.seek(last_pos)
+                                    encoded_bytes = out.read()
+                                    yield encoded_bytes
+                                    last_pos = new_last_pos
+
+            return _stream_generator()
+        else:
+            result = list(result)
+            sample_rate, audio = result[0][1]
+            audio = np.array([audio])
+
+            # Save the generated audio
+            with BytesIO() as out:
+                torchaudio.save(
+                    out, torch.from_numpy(audio), sample_rate, format=response_format
+                )
+                return out.getvalue()

xinference/model/audio/model_spec.json

@@ -103,6 +103,86 @@
     "model_ability": "audio-to-text",
     "multilingual": false
   },
+  {
+    "model_name": "whisper-tiny-mlx",
+    "model_family": "whisper",
+    "model_id": "mlx-community/whisper-tiny",
+    "model_ability": "audio-to-text",
+    "multilingual": true,
+    "engine": "mlx"
+  },
+  {
+    "model_name": "whisper-tiny.en-mlx",
+    "model_family": "whisper",
+    "model_id": "mlx-community/whisper-tiny.en-mlx",
+    "model_ability": "audio-to-text",
+    "multilingual": false,
+    "engine": "mlx"
+  },
+  {
+    "model_name": "whisper-base-mlx",
+    "model_family": "whisper",
+    "model_id": "mlx-community/whisper-base-mlx",
+    "model_ability": "audio-to-text",
+    "multilingual": true,
+    "engine": "mlx"
+  },
+  {
+    "model_name": "whisper-base.en-mlx",
+    "model_family": "whisper",
+    "model_id": "mlx-community/whisper-base.en-mlx",
+    "model_ability": "audio-to-text",
+    "multilingual": false,
+    "engine": "mlx"
+  },
+  {
+    "model_name": "whisper-small-mlx",
+    "model_family": "whisper",
+    "model_id": "mlx-community/whisper-small-mlx",
+    "model_ability": "audio-to-text",
+    "multilingual": true,
+    "engine": "mlx"
+  },
+  {
+    "model_name": "whisper-small.en-mlx",
+    "model_family": "whisper",
+    "model_id": "mlx-community/whisper-small.en-mlx",
+    "model_ability": "audio-to-text",
+    "multilingual": false,
+    "engine": "mlx"
+  },
+  {
+    "model_name": "whisper-medium-mlx",
+    "model_family": "whisper",
+    "model_id": "mlx-community/whisper-medium-mlx",
+    "model_ability": "audio-to-text",
+    "multilingual": true,
+    "engine": "mlx"
+  },
+  {
+    "model_name": "whisper-medium.en-mlx",
+    "model_family": "whisper",
+    "model_id": "mlx-community/whisper-medium.en-mlx",
+    "model_ability": "audio-to-text",
+    "multilingual": false,
+    "engine": "mlx"
+  },
+  {
+    "model_name": "whisper-large-v3-mlx",
+    "model_family": "whisper",
+    "model_id": "mlx-community/whisper-large-v3-mlx",
+    "model_ability": "audio-to-text",
+    "multilingual": true,
+    "engine": "mlx"
+  },
+  {
+    "model_name": "whisper-large-v3-turbo-mlx",
+    "model_family": "whisper",
+    "model_id": "mlx-community/whisper-large-v3-turbo",
+    "model_ability": "audio-to-text",
+    "multilingual": true,
+    "engine": "mlx"
+  },
   {
     "model_name": "SenseVoiceSmall",
     "model_family": "funasr",
@@ -162,5 +242,13 @@
     "model_revision": "069c573759936b35191d3380deb89183c0656f59",
     "model_ability": "text-to-audio",
     "multilingual": true
+  },
+  {
+    "model_name": "F5-TTS",
+    "model_family": "F5-TTS",
+    "model_id": "SWivid/F5-TTS",
+    "model_revision": "4dcc16f297f2ff98a17b3726b16f5de5a5e45672",
+    "model_ability": "text-to-audio",
+    "multilingual": true
   }
 ]

xinference/model/audio/model_spec_modelscope.json

@@ -73,5 +73,14 @@
     "model_revision": "master",
     "model_ability": "text-to-audio",
     "multilingual": true
+  },
+  {
+    "model_name": "F5-TTS",
+    "model_family": "F5-TTS",
+    "model_hub": "modelscope",
+    "model_id": "SWivid/F5-TTS_Emilia-ZH-EN",
+    "model_revision": "master",
+    "model_ability": "text-to-audio",
+    "multilingual": true
   }
 ]

xinference/model/audio/whisper_mlx.py

@@ -0,0 +1,208 @@
+# 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 functools
+import itertools
+import logging
+import tempfile
+from typing import TYPE_CHECKING, List, Optional
+
+if TYPE_CHECKING:
+    from .core import AudioModelFamilyV1
+
+logger = logging.getLogger(__name__)
+
+
+class WhisperMLXModel:
+    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
+        self._use_lighting = False
+
+    @property
+    def model_ability(self):
+        return self._model_spec.model_ability
+
+    def load(self):
+        use_lightning = self._kwargs.get("use_lightning", "auto")
+        if use_lightning not in ("auto", True, False, None):
+            raise ValueError("use_lightning can only be True, False, None or auto")
+
+        if use_lightning == "auto" or use_lightning is True:
+            try:
+                import mlx.core as mx
+                from lightning_whisper_mlx.transcribe import ModelHolder
+            except ImportError:
+                if use_lightning == "auto":
+                    use_lightning = False
+                else:
+                    error_message = "Failed to import module 'lightning_whisper_mlx'"
+                    installation_guide = [
+                        "Please make sure 'lightning_whisper_mlx' is installed.\n",
+                    ]
+
+                    raise ImportError(
+                        f"{error_message}\n\n{''.join(installation_guide)}"
+                    )
+            else:
+                use_lightning = True
+        if not use_lightning:
+            try:
+                import mlx.core as mx  # noqa: F811
+                from mlx_whisper.transcribe import ModelHolder  # noqa: F811
+            except ImportError:
+                error_message = "Failed to import module 'mlx_whisper'"
+                installation_guide = [
+                    "Please make sure 'mlx_whisper' is installed.\n",
+                ]
+
+                raise ImportError(f"{error_message}\n\n{''.join(installation_guide)}")
+            else:
+                use_lightning = False
+
+        logger.info(
+            "Loading MLX whisper from %s, use lightning: %s",
+            self._model_path,
+            use_lightning,
+        )
+        self._use_lighting = use_lightning
+        self._model = ModelHolder.get_model(self._model_path, mx.float16)
+
+    def transcriptions(
+        self,
+        audio: bytes,
+        language: Optional[str] = None,
+        prompt: Optional[str] = None,
+        response_format: str = "json",
+        temperature: float = 0,
+        timestamp_granularities: Optional[List[str]] = None,
+    ):
+        return self._call(
+            audio,
+            language=language,
+            prompt=prompt,
+            response_format=response_format,
+            temperature=temperature,
+            timestamp_granularities=timestamp_granularities,
+            task="transcribe",
+        )
+
+    def translations(
+        self,
+        audio: bytes,
+        language: Optional[str] = None,
+        prompt: Optional[str] = None,
+        response_format: str = "json",
+        temperature: float = 0,
+        timestamp_granularities: Optional[List[str]] = None,
+    ):
+        if not self._model_spec.multilingual:
+            raise RuntimeError(
+                f"Model {self._model_spec.model_name} is not suitable for translations."
+            )
+        return self._call(
+            audio,
+            language=language,
+            prompt=prompt,
+            response_format=response_format,
+            temperature=temperature,
+            timestamp_granularities=timestamp_granularities,
+            task="translate",
+        )
+
+    def _call(
+        self,
+        audio: bytes,
+        language: Optional[str] = None,
+        prompt: Optional[str] = None,
+        response_format: str = "json",
+        temperature: float = 0,
+        timestamp_granularities: Optional[List[str]] = None,
+        task: str = "transcribe",
+    ):
+        if self._use_lighting:
+            from lightning_whisper_mlx.transcribe import transcribe_audio
+
+            transcribe = functools.partial(
+                transcribe_audio, batch_size=self._kwargs.get("batch_size", 12)
+            )
+        else:
+            from mlx_whisper import transcribe  # type: ignore
+
+        with tempfile.NamedTemporaryFile(delete=True) as f:
+            f.write(audio)
+
+            kwargs = {"task": task}
+            if response_format == "verbose_json":
+                if timestamp_granularities == ["word"]:
+                    kwargs["word_timestamps"] = True  # type: ignore
+
+            result = transcribe(
+                f.name,
+                path_or_hf_repo=self._model_path,
+                language=language,
+                temperature=temperature,
+                initial_prompt=prompt,
+                **kwargs,
+            )
+            text = result["text"]
+            segments = result["segments"]
+            language = result["language"]
+
+            if response_format == "json":
+                return {"text": text}
+            elif response_format == "verbose_json":
+                if not timestamp_granularities or timestamp_granularities == [
+                    "segment"
+                ]:
+                    return {
+                        "task": task,
+                        "language": language,
+                        "duration": segments[-1]["end"] if segments else 0,
+                        "text": text,
+                        "segments": segments,
+                    }
+                else:
+                    assert timestamp_granularities == ["word"]
+
+                    def _extract_word(word: dict) -> dict:
+                        return {
+                            "start": word["start"].item(),
+                            "end": word["end"].item(),
+                            "word": word["word"],
+                        }
+
+                    words = [
+                        _extract_word(w)
+                        for w in itertools.chain(*[s["words"] for s in segments])
+                    ]
+                    return {
+                        "task": task,
+                        "language": language,
+                        "duration": words[-1]["end"] if words else 0,
+                        "text": text,
+                        "words": words,
+                    }
+            else:
+                raise ValueError(f"Unsupported response format: {response_format}")