diff --git a/runhouse/servers/env_servlet.py b/runhouse/servers/env_servlet.py
index fc8e29fa4..941104dc3 100644
--- a/runhouse/servers/env_servlet.py
+++ b/runhouse/servers/env_servlet.py
@@ -1,4 +1,5 @@
 import logging
+import os
 import traceback
 from functools import wraps
 from typing import Any, Dict, Optional
@@ -69,6 +70,20 @@ async def __init__(self, env_name: str, *args, **kwargs):
             setup_cluster_servlet=ClusterServletSetupOption.GET_OR_FAIL,
         )
 
+        # Ray defaults to setting OMP_NUM_THREADS to 1, which unexpectedly limit parallelism in user programs.
+        # We delete it by default, but if we find that the user explicitly set it to another value, we respect that.
+        # This is really only a factor if the user set the value inside the VM or container, or inside the base_env
+        # which a cluster was initialized with. If they set it inside the env constructor and the env was sent to the
+        # cluster normally with .to, it will be set after this point.
+        # TODO this had no effect when we did it below where we set CUDA_VISIBLE_DEVICES, so we may need to move that
+        #  here and mirror the same behavior (setting it based on the detected gpus in the whole cluster may not work
+        #  for multinode, but popping it may also break things, it needs to be tested).
+        num_threads = os.environ.get("OMP_NUM_THREADS")
+        if num_threads is not None and num_threads != "1":
+            os.environ["OMP_NUM_THREADS"] = num_threads
+        else:
+            os.environ["OMP_NUM_THREADS"] = ""
+
         self.output_types = {}
         self.thread_ids = {}
 
diff --git a/runhouse/servers/obj_store.py b/runhouse/servers/obj_store.py
index f0cb81ef2..7e45cfacb 100644
--- a/runhouse/servers/obj_store.py
+++ b/runhouse/servers/obj_store.py
@@ -135,20 +135,6 @@ def __init__(self):
         self.installed_envs = {}  # TODO: consider deleting it?
         self._kv_store: Dict[Any, Any] = None
 
-        # Ray defaults to setting OMP_NUM_THREADS to 1, which unexpectedly limit parallelism in user programs.
-        # We delete it by default, but if we find that the user explicitly set it to another value, we respect that.
-        # This is really only a factor if the user set the value inside the VM or container, or inside the base_env
-        # which a cluster was initialized with. If they set it inside the env constructor and the env was sent to the
-        # cluster normally with .to, it will be set after this point.
-        # TODO this had no effect when we did it below where we set CUDA_VISIBLE_DEVICES, so we may need to move that
-        #  here and mirror the same behavior (setting it based on the detected gpus in the whole cluster may not work
-        #  for multinode, but popping it may also break things, it needs to be tested).
-        num_threads = os.environ.get("OMP_NUM_THREADS")
-        if num_threads is not None and num_threads != "1":
-            os.environ["OMP_NUM_THREADS"] = num_threads
-        else:
-            os.environ["OMP_NUM_THREADS"] = ""
-
     async def ainitialize(
         self,
         servlet_name: Optional[str] = None,