[`NllbMoe`] Update code to properly support loss computation (#25429)

* update nllb_moe

* fix

* doc nits

* nits

* add a small test

* ficup

* remove adapted from

src/transformers/models/nllb_moe/modeling_nllb_moe.py

@@ -126,7 +126,6 @@ def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_l
     return incremental_indices.long() + padding_idx
 
 
-# Copied from transformers.models.switch_transformers.modeling_switch_transformers.load_balancing_loss_func with SwitchTransformers->NllbMoeModel
 def load_balancing_loss_func(router_probs: torch.Tensor, expert_indices: torch.Tensor) -> float:
     r"""
     Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
@@ -144,6 +143,9 @@ def load_balancing_loss_func(router_probs: torch.Tensor, expert_indices: torch.T
     Returns:
         The auxiliary loss.
     """
+    if router_probs is None:
+        return 0
+
     num_experts = router_probs.shape[-1]
 
     # cast the expert indices to int64, otherwise one-hot encoding will fail
@@ -699,7 +701,9 @@ class NllbMoeEncoderLayer(nn.Module):
         if self.is_sparse:
             hidden_states, router_states = self.ffn(hidden_states, attention_mask)
         else:
-            hidden_states = self.ffn(hidden_states)
+            # router_states set to None to track which layers have None gradients.
+            hidden_states, router_states = self.ffn(hidden_states), None
+
         hidden_states = self.ff_dropout(hidden_states)
 
         hidden_states = residual + hidden_states
@@ -830,7 +834,8 @@ class NllbMoeDecoderLayer(nn.Module):
         if self.is_sparse:
             hidden_states, router_states = self.ffn(hidden_states, attention_mask)
         else:
-            hidden_states = self.ffn(hidden_states)
+            hidden_states, router_states = self.ffn(hidden_states), None
+
         hidden_states = self.ff_dropout(hidden_states)
 
         hidden_states = residual + hidden_states
@@ -1730,7 +1735,7 @@ class NllbMoeForConditionalGeneration(NllbMoePreTrainedModel):
 
             if output_router_logits:
                 encoder_router_logits = outputs[-1]
-                decoder_router_logits = outputs[5 if output_attentions else 3]
+                decoder_router_logits = outputs[3 if output_attentions else 4]
 
                 # Compute the router loss (z_loss + auxiliary loss) for each router in the encoder and decoder
                 encoder_router_logits, encoder_expert_indexes = self._unpack_router_logits(encoder_router_logits)
@@ -1775,7 +1780,6 @@ class NllbMoeForConditionalGeneration(NllbMoePreTrainedModel):
             decoder_router_logits=outputs.decoder_router_logits,
         )
 
-    # Copied from transfomers.models.switch_transformers.SwitchTransformersForConditionalGeneration._unpack_router_logits
     def _unpack_router_logits(self, router_outputs):
         total_router_logits = []
         total_expert_indexes = []
@@ -1784,11 +1788,10 @@ class NllbMoeForConditionalGeneration(NllbMoePreTrainedModel):
                 router_logits, expert_indexes = router_output
                 total_router_logits.append(router_logits)
                 total_expert_indexes.append(expert_indexes)
-        if len(total_expert_indexes) > 0:
-            total_router_logits = torch.cat(total_router_logits, dim=1)
-        if len(total_expert_indexes) > 0:
-            torch.cat(total_expert_indexes, dim=1)
-        return torch.cat(total_router_logits, dim=1), torch.cat(total_expert_indexes, dim=1)
+
+        total_router_logits = torch.cat(total_router_logits, dim=1) if len(total_router_logits) > 0 else None
+        total_expert_indexes = torch.stack(total_expert_indexes, dim=1) if len(total_expert_indexes) > 0 else None
+        return total_router_logits, total_expert_indexes
 
     # Copied from transfomers.models.switch_transformers.SwitchTransformersForConditionalGeneration.prepare_inputs_for_generation
     def prepare_inputs_for_generation(

tests/models/nllb_moe/test_modeling_nllb_moe.py

@@ -337,6 +337,16 @@ class NllbMoeModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMi
         model.generate(input_ids, attention_mask=attention_mask)
         model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
 
+    def test_get_loss(self):
+        config, input_dict = self.model_tester.prepare_config_and_inputs()
+        input_dict["output_router_logits"] = True
+        input_dict["labels"] = input_dict["input_ids"]
+        model = NllbMoeForConditionalGeneration(config).eval().to(torch_device)
+        out = model(**input_dict)
+        self.assertIsNotNone(out.loss)
+        self.assertIsNotNone(model(**input_dict)["encoder_router_logits"][1])
+        self.assertIsNotNone(model(**input_dict)["decoder_router_logits"][0])
+
 
 @require_torch
 @require_sentencepiece