feat: FastAPI ingress for TTS — GET /api/tts returns raw WAV

- Add FastAPI ingress to TTSDeployment with two routes:
  POST / — JSON API with base64 audio (backward compat)
  GET /api/tts?text=&language_id= — raw WAV bytes (zero overhead)
- GET /speakers endpoint for speaker listing
- Properly uses _fastapi naming to avoid collision with Ray binding
- app = TTSDeployment.bind() for rayservice.yaml compatibility

ray_serve/serve_tts.py

@@ -1,6 +1,10 @@
 """
 Ray Serve deployment for Coqui TTS.
 Runs on: elminster (RTX 2070 8GB, CUDA)
+
+Provides two API styles:
+  POST /tts          — JSON body → JSON response with base64 audio
+  GET  /tts/api/tts  — Coqui-compatible query params → raw WAV bytes
 """
 
 import base64
@@ -9,6 +13,8 @@ import os
 import time
 from typing import Any
 
+from fastapi import FastAPI, Query
+from fastapi.responses import Response
 from ray import serve
 
 try:
@@ -16,8 +22,11 @@ try:
 except ImportError:
     InferenceLogger = None
 
+_fastapi = FastAPI()
+
 
 @serve.deployment(name="TTSDeployment", num_replicas=1)
+@serve.ingress(_fastapi)
 class TTSDeployment:
     def __init__(self):
         import torch
@@ -50,100 +59,109 @@ class TTSDeployment:
         else:
             self._mlflow = None
 
-    async def __call__(self, request: dict[str, Any]) -> dict[str, Any]:
-        """
-        Handle text-to-speech requests.
+    # ── internal synthesis helpers ────────────────────────────────────────
 
-        Expected request format:
-        {
-            "text": "Text to synthesize",
-            "speaker": "speaker_name",
-            "language": "en",
-            "speed": 1.0,
-            "output_format": "wav",
-            "return_base64": true
-        }
-        """
+    def _synthesize(self, text: str, speaker: str | None = None,
+                    language: str | None = None, speed: float = 1.0):
+        """Return (wav_bytes: bytes, sample_rate: int, duration: float)."""
         import numpy as np
         from scipy.io import wavfile
 
+        wav = self.tts.tts(text=text, speaker=speaker, language=language, speed=speed)
+        if not isinstance(wav, np.ndarray):
+            wav = np.array(wav)
+
+        wav_int16 = (wav * 32767).astype(np.int16)
+        sample_rate = (
+            self.tts.synthesizer.output_sample_rate
+            if hasattr(self.tts, "synthesizer")
+            else 22050
+        )
+        buf = io.BytesIO()
+        wavfile.write(buf, sample_rate, wav_int16)
+        return buf.getvalue(), sample_rate, len(wav) / sample_rate
+
+    def _log(self, start: float, duration: float, text_len: int):
+        if self._mlflow:
+            elapsed = time.time() - start
+            self._mlflow.log_request(
+                latency_s=elapsed,
+                audio_duration_s=duration,
+                text_chars=text_len,
+                realtime_factor=elapsed / duration if duration > 0 else 0,
+            )
+
+    # ── POST / — JSON API (base64 audio in response) ────────────────────
+
+    @_fastapi.post("/")
+    async def generate_json(self, request: dict[str, Any]) -> dict[str, Any]:
+        """
+        JSON API — POST body:
+        {"text": "...", "speaker": "...", "language": "en", "speed": 1.0,
+         "output_format": "wav", "return_base64": true}
+        """
         _start = time.time()
         text = request.get("text", "")
+        if not text:
+            return {"error": "No text provided"}
+
         speaker = request.get("speaker")
         language = request.get("language")
         speed = request.get("speed", 1.0)
         output_format = request.get("output_format", "wav")
         return_base64 = request.get("return_base64", True)
 
-        if not text:
-            return {"error": "No text provided"}
-
-        # Generate speech
         try:
-            # TTS.tts returns a numpy array of audio samples
-            wav = self.tts.tts(
-                text=text,
-                speaker=speaker,
-                language=language,
-                speed=speed,
+            audio_bytes, sample_rate, duration = self._synthesize(
+                text, speaker, language, speed
             )
+            self._log(_start, duration, len(text))
 
-            # Convert to numpy array if needed
-            if not isinstance(wav, np.ndarray):
-                wav = np.array(wav)
-
-            # Normalize to int16
-            wav_int16 = (wav * 32767).astype(np.int16)
-
-            # Get sample rate from model config
-            sample_rate = (
-                self.tts.synthesizer.output_sample_rate
-                if hasattr(self.tts, "synthesizer")
-                else 22050
-            )
-
-            # Write to buffer
-            buffer = io.BytesIO()
-            wavfile.write(buffer, sample_rate, wav_int16)
-            audio_bytes = buffer.getvalue()
-
-            duration = len(wav) / sample_rate
-
-            # Log to MLflow
-            if self._mlflow:
-                self._mlflow.log_request(
-                    latency_s=time.time() - _start,
-                    audio_duration_s=duration,
-                    text_chars=len(text),
-                    realtime_factor=(time.time() - _start) / duration if duration > 0 else 0,
-                )
-
-            response = {
+            resp: dict[str, Any] = {
                 "model": self.model_name,
                 "sample_rate": sample_rate,
                 "duration": duration,
                 "format": output_format,
             }
-
             if return_base64:
-                response["audio"] = base64.b64encode(audio_bytes).decode("utf-8")
+                resp["audio"] = base64.b64encode(audio_bytes).decode("utf-8")
             else:
-                response["audio_bytes"] = audio_bytes
-
-            return response
-
+                resp["audio_bytes"] = audio_bytes
+            return resp
         except Exception as e:
-            return {
-                "error": str(e),
-                "model": self.model_name,
-            }
+            return {"error": str(e), "model": self.model_name}
 
+    # ── GET /api/tts — Coqui-compatible raw WAV endpoint ─────────────────
+
+    @_fastapi.get("/api/tts")
+    async def generate_raw(
+        self,
+        text: str = Query(..., description="Text to synthesize"),
+        language_id: str = Query("en", description="Language code"),
+        speaker_id: str | None = Query(None, description="Speaker name"),
+    ) -> Response:
+        """Coqui XTTS-compatible endpoint — returns raw WAV bytes."""
+        _start = time.time()
+        if not text:
+            return Response(content="text parameter required", status_code=400)
+
+        try:
+            audio_bytes, _sr, duration = self._synthesize(
+                text, speaker_id, language_id
+            )
+            self._log(_start, duration, len(text))
+            return Response(content=audio_bytes, media_type="audio/wav")
+        except Exception as e:
+            return Response(content=str(e), status_code=500)
+
+    # ── GET /speakers — list available speakers ──────────────────────────
+
+    @_fastapi.get("/speakers")
     def list_speakers(self) -> dict[str, Any]:
         """List available speakers for multi-speaker models."""
         speakers = []
         if hasattr(self.tts, "speakers") and self.tts.speakers:
             speakers = self.tts.speakers
-
         return {
             "model": self.model_name,
             "speakers": speakers,