Fix ChatTTS CUDA device-side assert with text sanitize and GPU recovery.

Re-enable KV cache by default, normalize digits and unsafe chars, disable per-chunk split_text, and reload ChatTTS after CUDA errors.

Co-authored-by: Cursor <cursoragent@cursor.com>

gpu_utils.py

@@ -15,6 +15,10 @@ def release_cuda_cache() -> None:
         import torch
 
         if torch.cuda.is_available():
+            try:
+                torch.cuda.synchronize()
+            except Exception:
+                pass
             torch.cuda.empty_cache()
             if hasattr(torch.cuda, "ipc_collect"):
                 torch.cuda.ipc_collect()
@@ -22,6 +26,21 @@ def release_cuda_cache() -> None:
         pass
 
 
+def is_cuda_runtime_error(exc: BaseException) -> bool:
+    msg = str(exc).lower()
+    return any(
+        k in msg
+        for k in (
+            "cuda error",
+            "device-side assert",
+            "out of memory",
+            "cublas",
+            "illegal memory access",
+            "an illegal instruction",
+        )
+    )
+
+
 def cuda_memory_summary() -> str:
     """返回简要显存占用（调试用）。"""
     try: