Fix warmup text length and ve attribute for torch.compile

- Warmup now uses a ~170-char representative sentence so torch.compile
  JIT-compiles for typical token sequence lengths. Previously "Warmup."
  compiled for very short shapes, causing a full re-compile (17s) on the
  first real HA request and pushing total synthesis past 30s.
- Compile model.ve (voice encoder) in addition to s3gen — both are
  convolutional and hit the MIOpen workspace=0 bug.
- Fix _patch_timing: attribute is model.ve not model.voice_encoder,
  so the timing wrap was silently skipping the speaker embedding.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

engine.py

@@ -49,15 +49,18 @@ def load_model() -> bool:
 
         if torch.cuda.is_available():
             # torch.compile replaces MIOpen's convolution path with Triton-generated
-            # kernels, bypassing the workspace=0 fallback entirely. We compile only
+            # kernels, bypassing the workspace=0 fallback entirely. We compile s3gen
-            # s3gen (HiFiGAN vocoder + flow matching) since that's the bottleneck.
+            # (HiFiGAN vocoder + flow matching) and the voice encoder (ve) since both
+            # are convolutional and hit the workspace=0 bug.
             # suppress_errors=True falls back to eager for any op compile can't handle.
-            try:
+            torch._dynamo.config.suppress_errors = True
-                torch._dynamo.config.suppress_errors = True
+            for attr, label in [("s3gen", "s3gen"), ("ve", "ve")]:
-                chatterbox_model.s3gen = torch.compile(chatterbox_model.s3gen, dynamic=True)
+                try:
-                logger.info("s3gen compiled with torch.compile")
+                    obj = getattr(chatterbox_model, attr)
-            except Exception:
+                    setattr(chatterbox_model, attr, torch.compile(obj, dynamic=True))
-                logger.warning("torch.compile unavailable, running s3gen in eager mode", exc_info=True)
+                    logger.info(f"{label} compiled with torch.compile")
+                except Exception:
+                    logger.warning(f"torch.compile unavailable for {label}, running in eager mode", exc_info=True)
 
         _patch_timing(chatterbox_model)
         logger.info("Model loaded successfully")
@@ -87,7 +90,7 @@ def _patch_timing(model) -> None:
         pass
     try:
         # Speaker/voice encoder — xvector embedding from reference audio
-        _wrap(model.voice_encoder, "forward", "voice_encoder (speaker embedding)")
+        _wrap(model.ve, "forward", "ve (speaker embedding)")
     except AttributeError:
         pass
     try:

main.py

@@ -23,7 +23,13 @@ def _warmup(voices: dict) -> None:
     audio_prompt = resolve_voice(None, voices) if voices else None
     logger.info("Running warmup synthesis to populate MIOpen kernel cache...")
     try:
-        engine.synthesize(text="Warmup.", audio_prompt_path=audio_prompt)
+        engine.synthesize(
+            text=(
+                "This is a warmup synthesis request used to pre-compile neural network kernels "
+                "for typical text lengths, so that the first real request runs at full speed."
+            ),
+            audio_prompt_path=audio_prompt,
+        )
         logger.info("Warmup complete — MIOpen cache populated")
     except Exception:
         logger.warning("Warmup synthesis failed (non-fatal)", exc_info=True)