diff --git a/ocpmodels/models/equiformer_v2/equiformer_v2_oc20.py b/ocpmodels/models/equiformer_v2/equiformer_v2_oc20.py
index f8aa5a258..59fabeda5 100644
--- a/ocpmodels/models/equiformer_v2/equiformer_v2_oc20.py
+++ b/ocpmodels/models/equiformer_v2/equiformer_v2_oc20.py
@@ -95,6 +95,7 @@ class EquiformerV2_OC20(BaseModel):
         proj_drop (float):          Dropout rate for outputs of attention and FFN in Transformer blocks
 
         weight_init (str):          ['normal', 'uniform'] initialization of weights of linear layers except those in radial functions
+        enforce_max_neighbors_strictly (bool):      When edges are subselected based on the `max_neighbors` arg, arbitrarily select amongst equidistant / degenerate edges to have exactly the correct number.
     """
 
     def __init__(
@@ -137,6 +138,7 @@ def __init__(
         drop_path_rate: float = 0.05,
         proj_drop: float = 0.0,
         weight_init: str = "normal",
+        enforce_max_neighbors_strictly: bool = True,
     ):
         super().__init__()
 
@@ -198,6 +200,8 @@ def __init__(
         self.weight_init = weight_init
         assert self.weight_init in ["normal", "uniform"]
 
+        self.enforce_max_neighbors_strictly = enforce_max_neighbors_strictly
+
         self.device = "cpu"  # torch.cuda.current_device()
 
         self.grad_forces = False
@@ -381,7 +385,10 @@ def forward(self, data):
             cell_offsets,
             _,  # cell offset distances
             neighbors,
-        ) = self.generate_graph(data)
+        ) = self.generate_graph(
+            data,
+            enforce_max_neighbors_strictly=self.enforce_max_neighbors_strictly,
+        )
 
         ###############################################################
         # Initialize data structures