don't perform backward pass in evaluation mode

sentence_transformers/losses/MatryoshkaLoss.py

@@ -4,6 +4,7 @@
 from collections.abc import Iterable
 from typing import Any
 
+import torch
 import torch.nn.functional as F
 from torch import Tensor, nn
 
@@ -92,15 +93,20 @@ def __call__(self, reps: list[list[Tensor]], *args, **kwargs) -> Tensor:
             weight = self.matryoshka_weights[idx]
 
             truncated = [[shrink(r, dim) for r in minibatch] for minibatch in reps]
+            compute_gradients = torch.is_grad_enabled()
             # we need to detach the truncated embeddings,
             # otherwise the first backward pass of the underlying function will clear the computation graph of the embedding truncation
-            detached = [[r.detach().requires_grad_() for r in minibatch] for minibatch in truncated]
-            loss += weight * self.fn(detached, *args, **kwargs)
+            if compute_gradients:
+                matryoshka_reps = [[r.detach().requires_grad_() for r in minibatch] for minibatch in truncated]
+            else:
+                matryoshka_reps = truncated
+            loss += weight * self.fn(matryoshka_reps, *args, **kwargs)
             # After computing the gradients in minibatches, we need to continue the backward pass through the truncation calculation
             # the gradients must be multipied with the weights because otherwise the matryoshka weights are not considered in the backward pass
-            for t_minibatch, d_minibatch in zip(truncated, detached):
-                for t, d in zip(t_minibatch, d_minibatch):
-                    t.backward(weight * d.grad)
+            if compute_gradients:
+                for t_minibatch, d_minibatch in zip(truncated, matryoshka_reps):
+                    for t, d in zip(t_minibatch, d_minibatch):
+                        t.backward(weight * d.grad)
         return loss