fix(mlx): serialize accelerator lifecycle and inference

Route MLX load, inference, unload, reset, cache cleanup, and shutdown through a single worker. Add affinity and concurrent-unload regression coverage.\n\nVerified: 17 related backend tests; frontend CI; cargo check.

backend/services/llm.py

@@ -2,7 +2,7 @@
 LLM inference module - delegates to backend abstraction layer.
 """
 
-from ..backends import get_llm_backend, LLMBackend
+from ..backends import LLMBackend, get_llm_backend, unload_backend
 
 
 def get_llm_model() -> LLMBackend:
@@ -10,6 +10,6 @@ def get_llm_model() -> LLMBackend:
     return get_llm_backend()
 
 
-def unload_llm_model() -> None:
-    """Unload LLM model to free memory."""
-    get_llm_backend().unload_model()
+async def unload_llm_model() -> None:
+    """Unload LLM model to free memory, serialized onto the MLX worker."""
+    await unload_backend(get_llm_backend())

backend/services/mlx_thread.py

@@ -0,0 +1,39 @@
+"""Single dedicated worker thread for all MLX GPU work.
+
+MLX's Metal command encoder/stream is thread-local: it binds to whichever
+thread first touches the GPU device. ``asyncio.to_thread()`` uses the event
+loop's default executor, which hands successive calls to different worker
+threads — a model loaded on one thread and generated on another raises
+"There is no Stream(gpu, N) in current thread" (issue #699).
+
+Routing every MLX load, generate, transcribe and unload through this one
+worker keeps them on a single thread. Because the pool has a single worker,
+submitted jobs also run to completion one at a time in submission order, so a
+load-then-infer pair submitted as one job cannot be interleaved with an unload
+or a different-size load from another request.
+"""
+
+import asyncio
+from concurrent.futures import ThreadPoolExecutor
+
+_mlx_executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="mlx-worker")
+
+
+def run_on_mlx_thread(func, *args):
+    """Run ``func(*args)`` on the single dedicated MLX worker thread."""
+    loop = asyncio.get_running_loop()
+    return loop.run_in_executor(_mlx_executor, func, *args)
+
+
+def clear_mlx_cache() -> None:
+    """Return MLX's cached unified memory to the OS after a model is freed.
+
+    Must run on the MLX worker thread (call it from an unload that is already
+    routed through ``run_on_mlx_thread``). ``clear_cache`` moved out of the
+    ``mlx.core.metal`` namespace in newer MLX, so resolve it from either.
+    """
+    import mlx.core as mx
+
+    clear = getattr(mx, "clear_cache", None) or getattr(getattr(mx, "metal", None), "clear_cache", None)
+    if clear is not None:
+        clear()

backend/services/transcribe.py

@@ -2,21 +2,19 @@
 STT (Speech-to-Text) module - delegates to backend abstraction layer.
 """
 
-from typing import Optional
-from ..backends import get_stt_backend, STTBackend
+from ..backends import STTBackend, get_stt_backend, unload_backend
 
 
 def get_whisper_model() -> STTBackend:
     """
     Get STT backend instance (MLX or PyTorch based on platform).
-    
+
     Returns:
         STT backend instance
     """
     return get_stt_backend()
 
 
-def unload_whisper_model():
-    """Unload Whisper model to free memory."""
-    backend = get_stt_backend()
-    backend.unload_model()
+async def unload_whisper_model():
+    """Unload Whisper model to free memory, serialized onto the MLX worker."""
+    await unload_backend(get_stt_backend())

backend/services/tts.py

@@ -2,28 +2,27 @@
 TTS inference module - delegates to backend abstraction layer.
 """
 
-from typing import Optional
-import numpy as np
 import io
+
+import numpy as np
 import soundfile as sf
 
-from ..backends import get_tts_backend, TTSBackend
+from ..backends import TTSBackend, get_tts_backend, unload_backend
 
 
 def get_tts_model() -> TTSBackend:
     """
     Get TTS backend instance (MLX or PyTorch based on platform).
-    
+
     Returns:
         TTS backend instance
     """
     return get_tts_backend()
 
 
-def unload_tts_model():
-    """Unload TTS model to free memory."""
-    backend = get_tts_backend()
-    backend.unload_model()
+async def unload_tts_model():
+    """Unload TTS model to free memory, serialized onto the MLX worker."""
+    await unload_backend(get_tts_backend())
 
 
 def audio_to_wav_bytes(audio: np.ndarray, sample_rate: int) -> bytes: