Enable nemo-retriever install and CLI on non-Linux platforms

api/src/nv_ingest_api/internal/primitives/nim/model_interface/decorators.py

@@ -9,12 +9,20 @@
 
 logger = logging.getLogger(__name__)
 
-# Create a shared manager and lock for thread-safe access
-manager = Manager()
-global_cache = manager.dict()
+# Lazily initialized on first use so that importing this module does not
+# spawn a subprocess at import time (which breaks Windows multiprocessing).
+_manager = None
+global_cache = None
 lock = Lock()
 
 
+def _ensure_manager():
+    global _manager, global_cache
+    if _manager is None:
+        _manager = Manager()
+        global_cache = _manager.dict()
+
+
 def multiprocessing_cache(max_calls):
     """
     A decorator that creates a global cache shared between multiple processes.
@@ -28,10 +36,14 @@ def multiprocessing_cache(max_calls):
     """
 
     def decorator(func):
-        call_count = manager.Value("i", 0)  # Shared integer for call counting
+        call_count = None  # Initialized lazily alongside the manager
 
         @wraps(func)
         def wrapper(*args, **kwargs):
+            nonlocal call_count
+            _ensure_manager()
+            if call_count is None:
+                call_count = _manager.Value("i", 0)
             key = (func.__name__, args, frozenset(kwargs.items()))
 
             with lock:

nemo_retriever/pyproject.toml

@@ -63,7 +63,7 @@ dependencies = [
   "nemotron-page-elements-v3>=0.dev0",
   "nemotron-graphic-elements-v1>=0.dev0",
   "nemotron-table-structure-v1>=0.dev0",
-  "nemotron-ocr>=0.dev0",
+  "nemotron-ocr>=0.dev0; sys_platform == 'linux'",
   "markitdown",
   "timm==1.0.22",
   "tqdm",
@@ -76,7 +76,7 @@ dependencies = [
   "soundfile>=0.12.0",
   "scipy>=1.11.0",
   "nvidia-ml-py",
-  "vllm==0.16.0",
+  "vllm==0.16.0; sys_platform == 'linux'",
 ]
 
 [project.optional-dependencies]