Fix Kokoro defaulting to CPU in worker processes

The DefaultArgs class had xtts_device hardcoded to 'cpu', which meant
all uvicorn worker processes inherited this default instead of using
auto_torch_device() to detect GPU.

Changes:
- Set DefaultArgs.xtts_device to None initially
- Call auto_torch_device() after class definition to set default
- This ensures workers use GPU if available, not hardcoded CPU
- Fixed log message to show actual device being used (not args value)
- Log moved after device calculation for accuracy

This fixes Kokoro loading on CPU even when GPU is available.

speech.py

@@ -119,7 +119,7 @@ kokoro_lang = None
 
 # Default args for worker processes (will be overridden in __main__)
 class DefaultArgs:
-    xtts_device = 'cpu'
+    xtts_device = None  # Will be set after torch import
     use_deepspeed = False
     unload_timer = None
     log_level = 'INFO'
@@ -130,6 +130,14 @@ class DefaultArgs:
 
 args = DefaultArgs()
 
+# Set default device after torch is available (for worker processes)
+# Main process will override this in __main__ with argparse
+if args.xtts_device is None:
+    try:
+        args.xtts_device = auto_torch_device()
+    except:
+        args.xtts_device = 'cpu'
+
 # Voice-to-model lookup cache (loaded at startup)
 voice_to_model_cache = {}
 
@@ -800,10 +808,10 @@ async def generate_speech(request: GenerateSpeechRequest):
             async with kokoro_load_semaphore:
                 # Double-check after acquiring lock (another request may have loaded it)
                 if kokoro_pipeline is None or kokoro_lang != lang_code:
-                    logger.info(f"Loading/switching Kokoro pipeline to language '{lang_code}' on {args.xtts_device}")
                     # Run blocking model initialization in thread pool to avoid blocking event loop
-                    # Use same device as XTTS for GPU acceleration
+                    # Use GPU if available, fallback to CPU only if explicitly disabled
                     device = args.xtts_device if args.xtts_device != 'none' else 'cpu'
+                    logger.info(f"Loading/switching Kokoro pipeline to language '{lang_code}' on device '{device}'")
                     kokoro_pipeline = await asyncio.to_thread(kokoro_wrapper, lang_code=lang_code, device=device)
                     kokoro_lang = lang_code