xinference 1.0.1__py3-none-any.whl → 1.1.1__py3-none-any.whl

xinference/model/audio/f5tts_mlx.py

@@ -0,0 +1,260 @@
+# 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 datetime
+import io
+import logging
+import os
+from io import BytesIO
+from pathlib import Path
+from typing import TYPE_CHECKING, Literal, Optional, Union
+
+import numpy as np
+from tqdm import tqdm
+
+if TYPE_CHECKING:
+    from .core import AudioModelFamilyV1
+
+logger = logging.getLogger(__name__)
+
+
+class F5TTSMLXModel:
+    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._model = None
+
+    @property
+    def model_ability(self):
+        return self._model_spec.model_ability
+
+    def load(self):
+        try:
+            import mlx.core as mx
+            from f5_tts_mlx.cfm import F5TTS
+            from f5_tts_mlx.dit import DiT
+            from f5_tts_mlx.duration import DurationPredictor, DurationTransformer
+            from vocos_mlx import Vocos
+        except ImportError:
+            error_message = "Failed to import module 'f5_tts_mlx'"
+            installation_guide = [
+                "Please make sure 'f5_tts_mlx' is installed.\n",
+            ]
+
+            raise ImportError(f"{error_message}\n\n{''.join(installation_guide)}")
+
+        path = Path(self._model_path)
+        # vocab
+
+        vocab_path = path / "vocab.txt"
+        vocab = {v: i for i, v in enumerate(Path(vocab_path).read_text().split("\n"))}
+        if len(vocab) == 0:
+            raise ValueError(f"Could not load vocab from {vocab_path}")
+
+        # duration predictor
+
+        duration_model_path = path / "duration_v2.safetensors"
+        duration_predictor = None
+
+        if duration_model_path.exists():
+            duration_predictor = DurationPredictor(
+                transformer=DurationTransformer(
+                    dim=512,
+                    depth=8,
+                    heads=8,
+                    text_dim=512,
+                    ff_mult=2,
+                    conv_layers=2,
+                    text_num_embeds=len(vocab) - 1,
+                ),
+                vocab_char_map=vocab,
+            )
+            weights = mx.load(duration_model_path.as_posix(), format="safetensors")
+            duration_predictor.load_weights(list(weights.items()))
+
+        # vocoder
+
+        vocos = Vocos.from_pretrained("lucasnewman/vocos-mel-24khz")
+
+        # model
+
+        model_path = path / "model.safetensors"
+
+        f5tts = F5TTS(
+            transformer=DiT(
+                dim=1024,
+                depth=22,
+                heads=16,
+                ff_mult=2,
+                text_dim=512,
+                conv_layers=4,
+                text_num_embeds=len(vocab) - 1,
+            ),
+            vocab_char_map=vocab,
+            vocoder=vocos.decode,
+            duration_predictor=duration_predictor,
+        )
+
+        weights = mx.load(model_path.as_posix(), format="safetensors")
+        f5tts.load_weights(list(weights.items()))
+        mx.eval(f5tts.parameters())
+
+        self._model = f5tts
+
+    def speech(
+        self,
+        input: str,
+        voice: str,
+        response_format: str = "mp3",
+        speed: float = 1.0,
+        stream: bool = False,
+        **kwargs,
+    ):
+        import mlx.core as mx
+        import soundfile as sf
+        import tomli
+        from f5_tts_mlx.generate import (
+            FRAMES_PER_SEC,
+            SAMPLE_RATE,
+            TARGET_RMS,
+            convert_char_to_pinyin,
+            split_sentences,
+        )
+
+        from .utils import ensure_sample_rate
+
+        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)
+        duration: Optional[float] = kwargs.pop("duration", None)
+        steps: Optional[int] = kwargs.pop("steps", 8)
+        cfg_strength: Optional[float] = kwargs.pop("cfg_strength", 2.0)
+        method: Literal["euler", "midpoint"] = kwargs.pop("method", "rk4")
+        sway_sampling_coef: float = kwargs.pop("sway_sampling_coef", -1.0)
+        seed: Optional[int] = kwargs.pop("seed", None)
+
+        prompt_speech_path: Union[str, io.BytesIO]
+        if prompt_speech is None:
+            base = os.path.join(os.path.dirname(__file__), "../../thirdparty/f5_tts")
+            config = os.path.join(base, "infer/examples/basic/basic.toml")
+            with open(config, "rb") as f:
+                config_dict = tomli.load(f)
+            prompt_speech_path = os.path.join(base, config_dict["ref_audio"])
+            prompt_text = config_dict["ref_text"]
+        else:
+            prompt_speech_path = io.BytesIO(prompt_speech)
+
+            if prompt_text is None:
+                raise ValueError("`prompt_text` cannot be empty")
+
+        audio, sr = sf.read(prompt_speech_path)
+        audio = ensure_sample_rate(audio, sr, SAMPLE_RATE)
+
+        audio = mx.array(audio)
+        ref_audio_duration = audio.shape[0] / SAMPLE_RATE
+        logger.debug(
+            f"Got reference audio with duration: {ref_audio_duration:.2f} seconds"
+        )
+
+        rms = mx.sqrt(mx.mean(mx.square(audio)))
+        if rms < TARGET_RMS:
+            audio = audio * TARGET_RMS / rms
+
+        sentences = split_sentences(input)
+        is_single_generation = len(sentences) <= 1 or duration is not None
+
+        if is_single_generation:
+            generation_text = convert_char_to_pinyin([prompt_text + " " + input])  # type: ignore
+
+            if duration is not None:
+                duration = int(duration * FRAMES_PER_SEC)
+
+            start_date = datetime.datetime.now()
+
+            wave, _ = self._model.sample(  # type: ignore
+                mx.expand_dims(audio, axis=0),
+                text=generation_text,
+                duration=duration,
+                steps=steps,
+                method=method,
+                speed=speed,
+                cfg_strength=cfg_strength,
+                sway_sampling_coef=sway_sampling_coef,
+                seed=seed,
+            )
+
+            wave = wave[audio.shape[0] :]
+            mx.eval(wave)
+
+            generated_duration = wave.shape[0] / SAMPLE_RATE
+            print(
+                f"Generated {generated_duration:.2f}s of audio in {datetime.datetime.now() - start_date}."
+            )
+
+        else:
+            start_date = datetime.datetime.now()
+
+            output = []
+
+            for sentence_text in tqdm(split_sentences(input)):
+                text = convert_char_to_pinyin([prompt_text + " " + sentence_text])  # type: ignore
+
+                if duration is not None:
+                    duration = int(duration * FRAMES_PER_SEC)
+
+                wave, _ = self._model.sample(  # type: ignore
+                    mx.expand_dims(audio, axis=0),
+                    text=text,
+                    duration=duration,
+                    steps=steps,
+                    method=method,
+                    speed=speed,
+                    cfg_strength=cfg_strength,
+                    sway_sampling_coef=sway_sampling_coef,
+                    seed=seed,
+                )
+
+                # trim the reference audio
+                wave = wave[audio.shape[0] :]
+                mx.eval(wave)
+
+                output.append(wave)
+
+            wave = mx.concatenate(output, axis=0)
+
+            generated_duration = wave.shape[0] / SAMPLE_RATE
+            logger.debug(
+                f"Generated {generated_duration:.2f}s of audio in {datetime.datetime.now() - start_date}."
+            )
+
+        # Save the generated audio
+        with BytesIO() as out:
+            with sf.SoundFile(
+                out, "w", SAMPLE_RATE, 1, format=response_format.upper()
+            ) as f:
+                f.write(np.array(wave))
+                return out.getvalue()

xinference/model/audio/fish_speech.py

@@ -11,10 +11,8 @@
 # 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 gc
 import logging
 import os.path
-import queue
 import sys
 from io import BytesIO
 from typing import TYPE_CHECKING, Optional
@@ -60,6 +58,7 @@ class FishSpeechModel:
         self._device = device
         self._llama_queue = None
         self._model = None
+        self._engine = None
         self._kwargs = kwargs
 
     @property
@@ -72,6 +71,7 @@ class FishSpeechModel:
             0, os.path.join(os.path.dirname(__file__), "../../thirdparty/fish_speech")
         )
 
+        from tools.inference_engine import TTSInferenceEngine
         from tools.llama.generate import launch_thread_safe_queue
         from tools.vqgan.inference import load_model as load_decoder_model
 
@@ -81,6 +81,11 @@ class FishSpeechModel:
             if not is_device_available(self._device):
                 raise ValueError(f"Device {self._device} is not available!")
 
+        # https://github.com/pytorch/pytorch/issues/129207
+        if self._device == "mps":
+            logger.warning("The Conv1d has bugs on MPS backend, fallback to CPU.")
+            self._device = "cpu"
+
         enable_compile = self._kwargs.get("compile", False)
         precision = self._kwargs.get("precision", torch.bfloat16)
         logger.info("Loading Llama model, compile=%s...", enable_compile)
@@ -102,102 +107,10 @@ class FishSpeechModel:
             device=self._device,
         )
 
-    @torch.inference_mode()
-    def _inference(
-        self,
-        text,
-        enable_reference_audio,
-        reference_audio,
-        reference_text,
-        max_new_tokens,
-        chunk_length,
-        top_p,
-        repetition_penalty,
-        temperature,
-        seed="0",
-        streaming=False,
-    ):
-        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,
-            GenerateResponse,
-            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,
-            reference_audio=reference_audio,
-            enable_reference_audio=enable_reference_audio,
-        )
-
-        # LLAMA Inference
-        request = dict(
-            device=self._model.device,
-            max_new_tokens=max_new_tokens,
-            text=text,
-            top_p=top_p,
-            repetition_penalty=repetition_penalty,
-            temperature=temperature,
-            compile=self._kwargs.get("compile", False),
-            iterative_prompt=chunk_length > 0,
-            chunk_length=chunk_length,
-            max_length=2048,
-            prompt_tokens=prompt_tokens if enable_reference_audio else None,
-            prompt_text=reference_text if enable_reference_audio else None,
-        )
-
-        response_queue = queue.Queue()
-        self._llama_queue.put(
-            GenerateRequest(
-                request=request,
-                response_queue=response_queue,
-            )
+        self._engine = TTSInferenceEngine(
+            self._llama_queue, self._model, precision, enable_compile
         )
 
-        segments = []
-
-        while True:
-            result: WrappedGenerateResponse = response_queue.get()
-            if result.status == "error":
-                raise result.response
-
-            result: GenerateResponse = result.response
-            if result.action == "next":
-                break
-
-            with autocast_exclude_mps(
-                device_type=self._model.device.type,
-                dtype=self._kwargs.get("precision", torch.bfloat16),
-            ):
-                fake_audios = decode_vq_tokens(
-                    decoder_model=self._model,
-                    codes=result.codes,
-                )
-
-            fake_audios = fake_audios.float().cpu().numpy()
-            segments.append(fake_audios)
-
-            if streaming:
-                yield fake_audios, None, None
-
-        if len(segments) == 0:
-            raise Exception("No audio generated, please check the input text.")
-
-        # No matter streaming or not, we need to return the final audio
-        audio = np.concatenate(segments, axis=0)
-        yield None, (self._model.spec_transform.sample_rate, audio), None
-
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-            gc.collect()
-
     def speech(
         self,
         input: str,
@@ -211,21 +124,31 @@ class FishSpeechModel:
         if speed != 1.0:
             logger.warning("Fish speech does not support setting speed: %s.", speed)
         import torchaudio
+        from tools.schema import ServeReferenceAudio, ServeTTSRequest
 
         prompt_speech = kwargs.get("prompt_speech")
-        result = self._inference(
-            text=input,
-            enable_reference_audio=kwargs.get(
-                "enable_reference_audio", prompt_speech is not None
-            ),
-            reference_audio=prompt_speech,
-            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),
-            streaming=stream,
+        prompt_text = kwargs.get("prompt_text", kwargs.get("reference_text", ""))
+        if prompt_speech is not None:
+            r = ServeReferenceAudio(audio=prompt_speech, text=prompt_text)
+            references = [r]
+        else:
+            references = []
+
+        assert self._engine is not None
+        result = self._engine.inference(
+            ServeTTSRequest(
+                text=input,
+                references=references,
+                reference_id=kwargs.get("reference_id"),
+                seed=kwargs.get("seed"),
+                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,
+                format=response_format,
+            )
         )
 
         if stream:
@@ -241,7 +164,9 @@ class FishSpeechModel:
                     last_pos = 0
                     with writer.open():
                         for chunk in result:
-                            chunk = chunk[0]
+                            if chunk.code == "final":
+                                continue
+                            chunk = chunk.audio[1]
                             if chunk is not None:
                                 chunk = chunk.reshape((chunk.shape[0], 1))
                                 trans_chunk = torch.from_numpy(chunk)
@@ -256,7 +181,7 @@ class FishSpeechModel:
             return _stream_generator()
         else:
             result = list(result)
-            sample_rate, audio = result[0][1]
+            sample_rate, audio = result[0].audio
             audio = np.array([audio])
 
             # Save the generated audio

xinference/model/audio/model_spec.json

@@ -236,10 +236,34 @@
     "multilingual": true
   },
   {
-    "model_name": "FishSpeech-1.4",
+    "model_name": "CosyVoice2-0.5B",
+    "model_family": "CosyVoice",
+    "model_id": "mrfakename/CosyVoice2-0.5B",
+    "model_revision": "5676baabc8a76dc93ef60a88bbd2420deaa2f644",
+    "model_ability": "text-to-audio",
+    "multilingual": true
+  },
+  {
+    "model_name": "FishSpeech-1.5",
     "model_family": "FishAudio",
-    "model_id": "fishaudio/fish-speech-1.4",
-    "model_revision": "069c573759936b35191d3380deb89183c0656f59",
+    "model_id": "fishaudio/fish-speech-1.5",
+    "model_revision": "268b6ec86243dd683bc78dab7e9a6cedf9191f2a",
+    "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
+  },
+  {
+    "model_name": "F5-TTS-MLX",
+    "model_family": "F5-TTS-MLX",
+    "model_id": "lucasnewman/f5-tts-mlx",
+    "model_revision": "7642bb232e3fcacf92c51c786edebb8624da6b93",
     "model_ability": "text-to-audio",
     "multilingual": true
   }

xinference/model/audio/model_spec_modelscope.json

@@ -73,5 +73,23 @@
     "model_revision": "master",
     "model_ability": "text-to-audio",
     "multilingual": true
+  },
+  {
+    "model_name": "CosyVoice2-0.5B",
+    "model_family": "CosyVoice",
+    "model_hub": "modelscope",
+    "model_id": "iic/CosyVoice2-0.5B",
+    "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/utils.py

@@ -11,8 +11,40 @@
 # 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 io
+
+import numpy as np
+
 from .core import AudioModelFamilyV1
 
 
 def get_model_version(audio_model: AudioModelFamilyV1) -> str:
     return audio_model.model_name
+
+
+def ensure_sample_rate(
+    audio: np.ndarray, old_sample_rate: int, sample_rate: int
+) -> np.ndarray:
+    import soundfile as sf
+    from scipy.signal import resample
+
+    if old_sample_rate != sample_rate:
+        # Calculate the new data length
+        new_length = int(len(audio) * sample_rate / old_sample_rate)
+
+        # Resample the data
+        resampled_data = resample(audio, new_length)
+
+        # Use BytesIO to save the resampled data to memory
+        with io.BytesIO() as buffer:
+            # Write the resampled data to the memory buffer
+            sf.write(buffer, resampled_data, sample_rate, format="WAV")
+
+            # Reset the buffer position to the beginning
+            buffer.seek(0)
+
+            # Read the data from the memory buffer
+            audio, sr = sf.read(buffer, dtype="float32")
+
+    return audio