Revert to sync LLM + sentence-level streaming

AsyncLLMEngine hangs on Jetson during model loading. Reverted to sync
LLM but added fine-grained text chunking (chunk_text_fine, ~200 chars)
for the stream endpoint. Each sentence/clause generates independently,
so first audio plays after ~2-4s instead of waiting for the full text.

Not true token-level streaming, but a significant latency reduction
for multi-sentence utterances without AsyncLLMEngine dependency.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

main.py

@@ -197,6 +197,38 @@ def chunk_text(text: str, max_chars: int = 800) -> List[str]:
     return chunks if chunks else [text]
 
 
+def chunk_text_fine(text: str, max_chars: int = 200) -> List[str]:
+    """
+    Split text into fine-grained chunks for streaming — every sentence or clause.
+    Smaller chunks = faster first-audio, slight quality tradeoff at boundaries.
+    """
+    # Split on sentence boundaries AND commas/semicolons with reasonable length
+    parts = re.split(r'(?<=[.!?;])\s+', text.strip())
+
+    # Further split long parts on commas
+    chunks = []
+    for part in parts:
+        if len(part) <= max_chars:
+            chunks.append(part)
+        else:
+            # Split on commas
+            sub = re.split(r',\s+', part)
+            current = []
+            current_len = 0
+            for s in sub:
+                if current and current_len + len(s) > max_chars:
+                    chunks.append(', '.join(current))
+                    current = []
+                    current_len = 0
+                current.append(s)
+                current_len += len(s)
+            if current:
+                chunks.append(', '.join(current))
+
+    # Filter empty chunks
+    return [c.strip() for c in chunks if c.strip()]
+
+
 def generate_speech_sync(text: str, voice: str) -> bytes:
     """
     Generate speech using Orpheus model (synchronous).
@@ -374,21 +406,20 @@ async def startup():
     from vllm import AsyncLLMEngine
     from vllm.engine.arg_utils import AsyncEngineArgs
     
-    # Monkey-patch OrpheusModel to use AsyncLLMEngine for true streaming
+    # Monkey-patch OrpheusModel to use sync LLM (AsyncLLMEngine hangs on Jetson)
     original_setup_engine = OrpheusModel._setup_engine
     def patched_setup_engine(self):
         model_name = self._map_model_params(self.model_name)
-        engine_args = AsyncEngineArgs(
+        from vllm import LLM
+        return LLM(
             model=model_name,
             max_model_len=MAX_MODEL_LEN,
             gpu_memory_utilization=0.85,
             enforce_eager=False,
         )
-        return AsyncLLMEngine.from_engine_args(engine_args)
     OrpheusModel._setup_engine = patched_setup_engine
 
-    # Sync token generation (for background jobs)
-    # Uses the async engine but collects all results synchronously
+    # Sync token generation
     def patched_generate_tokens_sync(self, prompt, voice=None, request_id="req-001",
                                      temperature=0.6, top_p=0.8, max_tokens=MAX_TOKENS,
                                      stop_token_ids=[49158], repetition_penalty=1.3):
@@ -400,27 +431,9 @@ async def startup():
             temperature=temperature, top_p=top_p, max_tokens=max_tokens,
             stop_token_ids=stop_token_ids, repetition_penalty=repetition_penalty,
         )
-        req_id = f"sync-{uuid.uuid4().hex[:8]}"
-
-        # Collect from async engine — use the running loop if available, else create one
-        async def _collect_all():
-            final = None
-            async for output in self.engine.generate(prompt_string, sampling_params, req_id):
-                final = output
-            return final
-
-        try:
-            loop = asyncio.get_running_loop()
-            # We're in an async context (background task) — use asyncio.ensure_future
-            import concurrent.futures
-            with concurrent.futures.ThreadPoolExecutor() as pool:
-                final_output = pool.submit(lambda: asyncio.run(_collect_all())).result()
-        except RuntimeError:
-            # No running loop — safe to use asyncio.run
-            final_output = asyncio.run(_collect_all())
-
-        if final_output:
-            text = final_output.outputs[0].text
+        outputs = self.engine.generate([prompt_string], sampling_params)
+        for output in outputs:
+            text = output.outputs[0].text
             print(f"Raw output (first 500 chars): {text[:500]}")
             tokens = re.findall(r'<custom_token_\d+>', text)
             print(f"Found {len(tokens)} tokens")
@@ -428,31 +441,6 @@ async def startup():
                 yield token
     OrpheusModel.generate_tokens_sync = patched_generate_tokens_sync
     
-    # Async streaming token generation (for /tts/stream — yields tokens as produced)
-    async def async_generate_tokens(model_instance, prompt, voice=None,
-                                    temperature=0.6, top_p=0.8, max_tokens=MAX_TOKENS,
-                                    stop_token_ids=[49158], repetition_penalty=1.3):
-        """Async generator: yields tokens incrementally as vLLM produces them."""
-        from vllm import SamplingParams
-        prompt_string = model_instance._format_prompt(prompt, voice)
-        print(f"[streaming] {prompt}")
-        sampling_params = SamplingParams(
-            temperature=temperature, top_p=top_p, max_tokens=max_tokens,
-            stop_token_ids=stop_token_ids, repetition_penalty=repetition_penalty,
-        )
-        request_id = f"stream-{uuid.uuid4().hex[:8]}"
-        prev_text_len = 0
-
-        async for output in model_instance.engine.generate(prompt_string, sampling_params, request_id):
-            # output.outputs[0].text grows incrementally
-            text = output.outputs[0].text
-            new_text = text[prev_text_len:]
-            prev_text_len = len(text)
-            # Extract new tokens from the incremental text
-            new_tokens = re.findall(r'<custom_token_\d+>', new_text)
-            for token in new_tokens:
-                yield token
-    
     model = OrpheusModel(model_name=ORPHEUS_MODEL)
     
     print("✓ Orpheus model loaded successfully")
@@ -624,11 +612,11 @@ async def get_audio(job_id: str):
 @app.post("/tts/stream")
 async def stream_tts(request: TTSStreamRequest):
     """
-    Stream TTS audio in real-time with true token-level streaming.
+    Stream TTS audio with sentence-level chunking.
 
-    Audio starts playing within ~1-2s instead of waiting for full generation.
-    Uses vLLM AsyncLLMEngine to yield tokens incrementally, decoded to audio
-    via orpheus_tts.decoder.tokens_decoder every 7 tokens.
+    Splits text into small chunks (sentences/clauses) and generates each
+    independently. First chunk's audio starts playing while later chunks
+    are still generating. Reduces perceived latency significantly.
     """
     global model
 
@@ -639,25 +627,27 @@ async def stream_tts(request: TTSStreamRequest):
     if voice not in BUILTIN_VOICES:
         voice = DEFAULT_VOICE
 
-    async def streaming_audio_generator():
-        """Async generator: vLLM tokens → SNAC decoder → PCM audio chunks."""
-        from orpheus_tts.decoder import tokens_decoder
+    def sync_audio_generator():
+        """Generate audio per-sentence, yielding as each finishes."""
         try:
-            text_chunks = chunk_text(request.text)
+            # Split into fine-grained chunks for faster first-audio
+            text_chunks = chunk_text_fine(request.text)
+            print(f"[stream] {len(text_chunks)} chunk(s): {[c[:40] for c in text_chunks]}")
             for chunk_idx, chunk in enumerate(text_chunks):
-                print(f"Stream chunk {chunk_idx + 1}/{len(text_chunks)}: {chunk[:80]}...")
-
-                # async_generate_tokens yields tokens as vLLM produces them
-                # tokens_decoder converts them to audio every 7 tokens
-                token_gen = async_generate_tokens(model, chunk, voice=voice)
-                async for audio_chunk in tokens_decoder(token_gen):
+                print(f"  Generating chunk {chunk_idx + 1}/{len(text_chunks)}: {chunk[:60]}...")
+                syn_tokens = model.generate_speech(
+                    prompt=chunk,
+                    voice=voice,
+                    max_tokens=MAX_TOKENS,
+                )
+                for audio_chunk in syn_tokens:
                     yield audio_chunk
         except Exception as e:
             print(f"Stream error: {e}")
             raise
 
     return StreamingResponse(
-        streaming_audio_generator(),
+        sync_audio_generator(),
         media_type="audio/pcm"
     )