ENH: Support fish speech reference audio (#2542)

doc/source/models/model_abilities/audio.rst

@@ -331,7 +331,7 @@ Clone voice, launch model ``CosyVoice-300M``.
 
     zero_shot_prompt_text = ""
     # The zero shot prompt file is the voice file
-    # the words said in the file shoule be identical to zero_shot_prompt_text
+    # the words said in the file should be identical to zero_shot_prompt_text
     with open(zero_shot_prompt_file, "rb") as f:
         zero_shot_prompt = f.read()
 
@@ -379,3 +379,34 @@ Instruction based, launch model ``CosyVoice-300M-Instruct``.
     )
 
 More instructions and examples, could be found at https://fun-audio-llm.github.io/ .
+
+
+FishSpeech Usage
+~~~~~~~~~~~~~~~~
+
+Basic usage, refer to :ref:`audio speech usage <audio_speech>`.
+
+Clone voice, launch model ``FishSpeech-1.4``. Please use `prompt_speech` instead of `reference_audio`
+to provide the reference audio to the FishSpeech model.
+
+.. code-block::
+
+    from xinference.client import Client
+
+    client = Client("http://<XINFERENCE_HOST>:<XINFERENCE_PORT>")
+
+    model = client.get_model("<MODEL_UID>")
+
+    reference_text = ""
+    # The reference audio file is the voice file
+    # the words said in the file should be identical to reference_text
+    with open(reference_audio_file, "rb") as f:
+        reference_audio = f.read()
+
+    speech_bytes = model.speech(
+        "<The text to generate audio for>",
+        reference_text=reference_text,
+        prompt_speech=reference_audio,
+        enable_reference_audio=True,
+    )
+-

xinference/client/restful/restful_client.py

@@ -704,6 +704,8 @@ def speech(
             The speed of the generated audio.
         stream: bool
             Use stream or not.
+        prompt_speech: bytes
+            The audio bytes to be provided to the model.
 
         Returns
         -------

xinference/model/audio/fish_speech.py

@@ -81,12 +81,13 @@ def load(self):
             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)
+        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,
+            compile=enable_compile,
         )
         logger.info("Llama model loaded, loading VQ-GAN model...")
 
@@ -208,11 +209,14 @@ def speech(
             logger.warning("stream mode is not implemented.")
         import torchaudio
 
+        prompt_speech = kwargs.get("prompt_speech")
         result = list(
             self._inference(
                 text=input,
-                enable_reference_audio=False,
-                reference_audio=None,
+                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),

xinference/model/audio/tests/test_fish_speech.py

@@ -22,8 +22,7 @@ def test_fish_speech(setup):
     client = Client(endpoint)
 
     model_uid = client.launch_model(
-        model_name="FishSpeech-1.4",
-        model_type="audio",
+        model_name="FishSpeech-1.4", model_type="audio", compile=False
     )
     model = client.get_model(model_uid)
     input_string = "你好，你是谁？"