fix segfault in load model

llama_stack/providers/utils/inference/embedding_mixin.py

@@ -6,9 +6,12 @@
 
 import asyncio
 import base64
+import platform
 import struct
 from typing import TYPE_CHECKING
 
+import torch
+
 from llama_stack.log import get_logger
 
 if TYPE_CHECKING:
@@ -24,6 +27,8 @@ from llama_stack.apis.inference import (
 
 EMBEDDING_MODELS = {}
 
+DARWIN = "Darwin"
+
 
 log = get_logger(name=__name__, category="providers::utils")
 
@@ -83,6 +88,17 @@ class SentenceTransformerEmbeddingMixin:
         def _load_model():
             from sentence_transformers import SentenceTransformer
 
+            platform_name = platform.system()
+            if platform_name == DARWIN:
+                # PyTorch's OpenMP kernels can segfault on macOS when spawned from background
+                # threads with the default parallel settings, so force a single-threaded CPU run.
+                log.debug(f"Constraining torch threads on {platform_name} to a single worker")
+                try:
+                    torch.set_num_threads(1)
+                    torch.set_num_interop_threads(1)
+                except Exception:
+                    log.debug(f"Failed to adjust torch thread counts on {platform_name}", exc_info=True)
+
             return SentenceTransformer(model, trust_remote_code=True)
 
         loaded_model = await asyncio.to_thread(_load_model)