superlocalmemory 3.3.17 → 3.3.18

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "superlocalmemory",
-  "version": "3.3.17",
+  "version": "3.3.18",
   "description": "Information-geometric agent memory with mathematical guarantees. 4-channel retrieval, Fisher-Rao similarity, zero-LLM mode, EU AI Act compliant. Works with Claude, Cursor, Windsurf, and 17+ AI tools.",
   "keywords": [
     "ai-memory",

package/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "superlocalmemory"
-version = "3.3.17"
+version = "3.3.18"
 description = "Information-geometric agent memory with mathematical guarantees"
 readme = "README.md"
 license = {text = "MIT"}

package/src/superlocalmemory/core/config.py

@@ -155,7 +155,7 @@ class RetrievalConfig:
     # Reranking (V3.3.2: ONNX backend enabled for all modes)
     use_cross_encoder: bool = True
     cross_encoder_model: str = "cross-encoder/ms-marco-MiniLM-L-12-v2"
-    cross_encoder_backend: str = "onnx"  # "onnx" (~200MB) or "" (PyTorch, ~1.5GB)
+    cross_encoder_backend: str = ""  # "" = PyTorch (~500MB stable), "onnx" = ONNX (leaks on ARM64 CoreML)
 
     # Agentic (Mode C only)
     agentic_max_rounds: int = 3

package/src/superlocalmemory/core/embedding_worker.py

@@ -35,6 +35,8 @@ os.environ["PYTORCH_MPS_MEM_LIMIT"] = "0"
 os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 os.environ["TORCH_DEVICE"] = "cpu"
+# V3.3.17: Disable CoreML EP for ONNX Runtime — uses 3-5GB on ARM64 Mac.
+os.environ["ORT_DISABLE_COREML"] = "1"
 
 # SIGTERM bridge: Docker/systemd send SIGTERM to stop processes.
 # Without this, the worker ignores SIGTERM and becomes a zombie.
@@ -65,6 +67,34 @@ def _start_parent_watchdog() -> None:
     t.start()
 
 
+def _load_embedding_model(name: str) -> tuple:
+    """Load embedding model. ONNX first (no memory leak), PyTorch fallback.
+
+    V3.3.17: PyTorch SentenceTransformer on ARM64 Mac leaks memory —
+    grows from 300MB to 17GB after ~200 encode calls. ONNX Runtime
+    has no such issue. Same approach as CrossEncoder ONNX migration.
+
+    Returns (model, backend_name) or (None, "").
+    """
+    from sentence_transformers import SentenceTransformer
+
+    # Tier 1: ONNX (stable memory, ~200MB footprint)
+    try:
+        m = SentenceTransformer(name, backend="onnx", trust_remote_code=True)
+        return m, "onnx"
+    except Exception:
+        pass
+
+    # Tier 2: PyTorch CPU (stable at ~1.4GB after 100+ calls, verified)
+    try:
+        import torch
+        with torch.inference_mode():
+            m = SentenceTransformer(name, trust_remote_code=True, device="cpu")
+        return m, "pytorch"
+    except Exception:
+        return None, ""
+
+
 def _worker_main() -> None:
     """Main loop: read JSON requests from stdin, write responses to stdout."""
     _start_parent_watchdog()  # V3.3.7: self-terminate if parent dies
@@ -97,18 +127,17 @@ def _worker_main() -> None:
         if cmd == "load":
             name = req.get("model_name", "nomic-ai/nomic-embed-text-v1.5")
             expected_dim = req.get("dimension", 768)
-            try:
-                from sentence_transformers import SentenceTransformer
-                model = SentenceTransformer(name, trust_remote_code=True, device="cpu")
+            model, active_backend = _load_embedding_model(name)
+            if model is not None:
                 dim = model.get_sentence_embedding_dimension()
                 if dim != expected_dim:
                     _respond({"ok": False, "error": f"Dimension mismatch: {dim} != {expected_dim}"})
                     model = None
                     continue
                 model_name = name
-                _respond({"ok": True, "dim": dim, "model": name})
-            except Exception as exc:
-                _respond({"ok": False, "error": str(exc)})
+                _respond({"ok": True, "dim": dim, "model": name, "backend": active_backend})
+            else:
+                _respond({"ok": False, "error": "Model load failed"})
             continue
 
         if cmd == "embed":
@@ -117,26 +146,16 @@ def _worker_main() -> None:
                 _respond({"ok": False, "error": "No texts provided"})
                 continue
             if model is None:
-                # Auto-load if not yet loaded
                 name = req.get("model_name", "nomic-ai/nomic-embed-text-v1.5")
-                expected_dim = req.get("dimension", 768)
-                try:
-                    from sentence_transformers import SentenceTransformer
-                    model = SentenceTransformer(name, trust_remote_code=True, device="cpu")
+                model, active_backend = _load_embedding_model(name)
+                if model is not None:
                     dim = model.get_sentence_embedding_dimension()
                     model_name = name
-                except Exception as exc:
-                    _respond({"ok": False, "error": f"Model load failed: {exc}"})
+                else:
+                    _respond({"ok": False, "error": "Model load failed"})
                     continue
             try:
-                # torch.inference_mode prevents autograd graph accumulation
-                # which causes silent memory leaks over long-running sessions.
-                try:
-                    import torch
-                    with torch.inference_mode():
-                        vecs = model.encode(texts, normalize_embeddings=True)
-                except ImportError:
-                    vecs = model.encode(texts, normalize_embeddings=True)
+                vecs = model.encode(texts, normalize_embeddings=True)
                 if isinstance(vecs, np.ndarray) and vecs.ndim == 2:
                     result = [vecs[i].tolist() for i in range(vecs.shape[0])]
                 else:

package/src/superlocalmemory/core/reranker_worker.py

@@ -40,6 +40,9 @@ os.environ["PYTORCH_MPS_MEM_LIMIT"] = "0"
 os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 os.environ["TORCH_DEVICE"] = "cpu"
+# V3.3.17: Disable CoreML EP for ONNX Runtime. CoreML compiles execution
+# plans that consume 3-5GB on ARM64 Mac. CPU EP is ~500MB and fast enough.
+os.environ["ORT_DISABLE_COREML"] = "1"
 
 # SIGTERM bridge for Docker/systemd
 if sys.platform != "win32":