Merge branch 'main' into dev

engine.py

@@ -8,6 +8,12 @@ chatterbox_model = None
 _sample_rate = 24000
 _is_turbo = False
 
+# Cache: voice file path → prepared conditionals object.
+# prepare_conditionals loads audio, runs s3tokenizer + voice encoder, and
+# builds mel embeddings — expensive work that only depends on the reference
+# audio, not the text. Cache it so multi-chunk requests pay the cost once.
+_cond_cache: dict = {}
+
 
 def _test_cuda() -> bool:
     try:
@@ -51,6 +57,26 @@ def load_model() -> bool:
         return False
 
 
+def prepare_voice(audio_prompt_path: str) -> None:
+    """
+    Pre-compute and cache the voice conditionals for a reference audio file.
+    Calling this once avoids repeating the s3tokenizer + voice encoder work
+    on every synthesis chunk that uses the same voice.
+    """
+    if chatterbox_model is None:
+        return
+    if audio_prompt_path in _cond_cache:
+        return
+    if not _is_turbo:
+        return  # only turbo exposes prepare_conditionals
+
+    logger.info(f"Preparing voice conditionals for '{audio_prompt_path}'")
+    with torch.inference_mode():
+        chatterbox_model.prepare_conditionals(audio_prompt_path)
+    _cond_cache[audio_prompt_path] = chatterbox_model.conds
+    logger.info("Voice conditionals cached")
+
+
 def get_sample_rate() -> int:
     return _sample_rate
 
@@ -71,8 +97,16 @@ def synthesize(
         if torch.cuda.is_available():
             torch.cuda.manual_seed_all(seed)
 
+    # Restore cached conditionals so generate() skips prepare_conditionals.
+    if audio_prompt_path and _is_turbo:
+        if audio_prompt_path not in _cond_cache:
+            prepare_voice(audio_prompt_path)
+        chatterbox_model.conds = _cond_cache[audio_prompt_path]
+
     kwargs: dict = {}
-    if audio_prompt_path:
+    # Don't pass audio_prompt_path — conds are already set above.
+    # For non-turbo models there's no cache, pass path as normal.
+    if audio_prompt_path and not _is_turbo:
         kwargs["audio_prompt_path"] = audio_prompt_path
 
     if _is_turbo:

main.py

@@ -19,8 +19,18 @@ logger = logging.getLogger(__name__)
 
 
 def _warmup(voices: dict) -> None:
-    """Run one synthesis to populate MIOpen's in-memory kernel cache."""
+    """Pre-compute voice conditionals and populate MIOpen's kernel cache."""
     from wyoming_voices import resolve_voice
+
+    # Pre-compute conditionals for all discovered voices so the first real
+    # request doesn't pay the s3tokenizer + voice encoder cost.
+    for name, path in voices.items():
+        try:
+            engine.prepare_voice(path)
+        except Exception:
+            logger.warning(f"Failed to prepare voice '{name}' (non-fatal)", exc_info=True)
+
+    # Synthesis warmup to populate MIOpen's in-memory kernel cache.
     audio_prompt = resolve_voice(None, voices) if voices else None
     logger.info("Running warmup synthesis...")
     try: