fix(mlx): release the model binding inside the sync bodies so a failing op's traceback cannot pin it

Review follow-up: the post-op drain ran in a finally, but on the error
path the exception's traceback kept the _generate_sync/_transcribe_sync
frame (and its model local) alive, so mx.clear_cache() had nothing to
return. Each sync body now drops its model (and tokenizer) binding in its
own finally before the exception propagates.

backend/backends/mlx_backend.py

@@ -241,6 +241,7 @@ class MLXTTSBackend:
             # thread mid-generation cannot turn a later self.model read into
             # None (the fallback path below runs seconds into a request).
             model = self.model
+            try:
                 # MLX generate() returns a generator yielding GenerationResult objects
                 audio_chunks = []
                 sample_rate = 24000
@@ -305,6 +306,11 @@ class MLXTTSBackend:
                     audio = np.array([], dtype=np.float32)
 
                 return audio, sample_rate
+            finally:
+                # Drop the local binding here, inside the frame a propagating
+                # traceback would keep alive, so the caller's drain can
+                # actually return the model's buffers to MLX.
+                del model
 
         def _reload_and_generate_sync():
             """Ensure the configured model is loaded, then generate — as ONE
@@ -324,9 +330,10 @@ class MLXTTSBackend:
             finally:
                 # An unload_model() that landed while we were generating only
                 # dropped the backend's reference; the model's buffers were
-                # kept alive by _generate_sync's local and have just been
+                # kept alive by _generate_sync's local, which that function
-                # returned to MLX's pool (on success or failure). Drain it now,
+                # drops in its own finally (so a propagating traceback cannot
-                # or they stay resident until the next load/unload cycle.
+                # pin it), and have just been returned to MLX's pool. Drain it
+                # now, or they stay resident until the next load/unload cycle.
                 if self.model is None:
                     empty_mlx_cache()
 
@@ -423,6 +430,7 @@ class MLXSTTBackend:
             # MLX Whisper transcription using generate method
             # The generate method accepts audio path directly
             model = self.model
+            try:
                 decode_options = {}
                 if language:
                     decode_options["language"] = language
@@ -441,6 +449,11 @@ class MLXSTTBackend:
                     return result.text.strip()
                 else:
                     return str(result).strip()
+            finally:
+                # Drop the local binding here, inside the frame a propagating
+                # traceback would keep alive, so the caller's drain can
+                # actually return the model's buffers to MLX.
+                del model
 
         def _reload_and_transcribe_sync():
             """Ensure the requested model is loaded, then transcribe — as ONE

backend/backends/qwen_llm_backend.py

@@ -305,6 +305,7 @@ class MLXQwenLLMBackend:
         from mlx_lm.sample_utils import make_sampler
 
         model, tokenizer = self.model, self.tokenizer
+        try:
             messages = _build_messages(prompt, system, examples)
             chat_prompt = tokenizer.apply_chat_template(
                 messages,
@@ -323,3 +324,8 @@ class MLXQwenLLMBackend:
                 verbose=False,
             )
             return text.strip()
+        finally:
+            # Drop the local binding here, inside the frame a propagating
+            # traceback would keep alive, so the caller's drain can
+            # actually return the model's buffers to MLX.
+            del model, tokenizer