solved cuda error just for now by moving the 2 tensors to cpu

openfold/utils/loss.py

@@ -38,6 +38,8 @@ import random
 from openfold.np import residue_constants as rc
 import logging 
 import procrustes 
+from openfold.utils.tensor_utils import tensor_tree_map
+import gc
 logger = logging.getLogger(__name__)
 
 def softmax_cross_entropy(logits, labels):
@@ -842,6 +844,7 @@ def between_residue_bond_loss(
     ] + next_is_proline * residue_constants.between_res_bond_length_stddev_c_n[
         1
     ]
+
     c_n_bond_length_error = torch.sqrt(eps + (c_n_bond_length - gt_length) ** 2)
     c_n_loss_per_residue = torch.nn.functional.relu(
         c_n_bond_length_error - tolerance_factor_soft * gt_stddev
@@ -1741,9 +1744,16 @@ def get_optimal_transform(
     src_center = src_atoms.mean(-2, keepdim=True)
     tgt_center = tgt_atoms.mean(-2, keepdim=True)
     r = kabsch_rotation(src_atoms,tgt_atoms)
+    del src_atoms,tgt_atoms,
+    gc.collect()
+
     tgt_center,src_center = tgt_center.to('cuda:0'),src_center.to('cuda:0')
-    x = tgt_center - src_center @ r.to('cuda:0')
-    return r, x
+    x = tgt_center.to('cpu') - src_center.to('cpu') @ r.to('cpu')
+
+    del tgt_center,src_center,mask
+    gc.collect()
+
+    return r, x.to('cuda')
 
 
 def compute_rmsd(
@@ -1756,6 +1766,9 @@ def compute_rmsd(
     true_atom_pos = true_atom_pos.to('cuda:0')
     pred_atom_pos = pred_atom_pos.to('cuda:0')
     sq_diff = torch.square(true_atom_pos - pred_atom_pos).sum(dim=-1, keepdim=False)
+    del true_atom_pos
+    del pred_atom_pos
+    gc.collect()
     if atom_mask is not None:
         sq_diff = sq_diff.to('cpu')[atom_mask.to('cpu')] # somehow it causes overflow on cuda so moved to cpu
     msd = torch.mean(sq_diff)
@@ -1830,7 +1843,7 @@ def greedy_align(
             used[i] = True
             continue
         cur_entity_ids = batch["entity_id"][asym_mask][0]
-        best_rmsd = 1e20    
+        best_rmsd = torch.inf    
         best_idx = None
         cur_asym_list = entity_2_asym_list[int(cur_entity_ids)]
         cur_residue_index = per_asym_residue_index[int(cur_asym_id)]
@@ -1847,6 +1860,7 @@ def greedy_align(
                 rmsd = compute_rmsd(
                     cropped_pos, cur_pred_pos, (cur_pred_mask.to('cuda:0') * mask.to('cuda:0')).bool()
                 )
+                print(f"rmsd is {rmsd}")
                 if rmsd < best_rmsd:   
                     best_rmsd = rmsd
                     best_idx = j
@@ -2047,13 +2061,17 @@ class AlphaFoldMultimerLoss(AlphaFoldLoss):
         anchor_pred_pos = pred_ca_pos[asym_mask]
         anchor_true_mask = true_ca_masks[anchor_gt_idx][anchor_residue_idx]
         anchor_pred_mask = pred_ca_mask[asym_mask]
+        input_mask = (anchor_true_mask.to('cuda:0') * anchor_pred_mask.to('cuda:0')).bool()
         r, x = get_optimal_transform(
             anchor_true_pos,
-            anchor_pred_pos,
-            (anchor_true_mask.to('cuda:0') * anchor_pred_mask.to('cuda:0')).bool(),
+            anchor_pred_pos,mask=input_mask
         )
+        del input_mask # just to save memory
+        del anchor_pred_mask
+        del anchor_true_mask
+        gc.collect()
 
-        aligned_true_ca_poses = [ca.to('cuda:0') @ r.to('cuda:0') + x.to('cuda:0') for ca in true_ca_poses]  # apply transforms
+        aligned_true_ca_poses = [ca.to('cpu') @ r.to('cpu') + x.to('cpu') for ca in true_ca_poses]  # apply transforms
         align = greedy_align(
                 batch,
                 per_asym_residue_index,
@@ -2064,6 +2082,11 @@ class AlphaFoldMultimerLoss(AlphaFoldLoss):
                 aligned_true_ca_poses,
                 true_ca_masks,
             )
+        
+        del aligned_true_ca_poses
+        del r,x
+        gc.collect()
+
         merged_labels = merge_labels(
             batch,
             per_asym_residue_index,
@@ -2091,6 +2114,10 @@ class AlphaFoldMultimerLoss(AlphaFoldLoss):
         permutated_labels = self.multi_chain_perm_align(out,features,labels)
         logger.info("finished multi-chain permutation")
         features.update(permutated_labels)
+        move_to_gpu = lambda t: (t.to('cuda:0'))
+        features = tensor_tree_map(move_to_gpu,features)
+        print(f"after moving features:",torch.cuda.memory_allocated(0))
+        # out = tensor_tree_map(move_to_gpu,out)
         self.loss(out,features)
         return permutated_labels
         ## TODO next need to check how the ground truth label is used