Fix RAG generation for T5 FiD (#3657)

I noticed a big discrepency between exact_match and token_em after I switched to using T5 FiD in public ParlAI.
Per discussion with Kurt, it's cause the generate for T5 is supposed to call TorchGeneratorAgent, rather than using the T5 generation via Hugging Face. This was in the original itnernal implementation, but got lost at some point while open sourcing.

Verified by printing out forced decoding outputs and the output from eval_model. Verified that when `token_em == 1` that the output of eval_model was also the same.

parlai/agents/rag/rag.py

@@ -3,6 +3,7 @@
 # Copyright (c) Facebook, Inc. and its affiliates.
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
+
 """
 Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.
 
@@ -27,7 +28,7 @@
 from parlai.core.opt import Opt
 from parlai.core.params import ParlaiParser
 from parlai.core.torch_agent import History, Batch
-from parlai.core.torch_generator_agent import PPLMetric, TreeSearch
+from parlai.core.torch_generator_agent import PPLMetric, TorchGeneratorAgent, TreeSearch
 from parlai.utils.distributed import sync_parameters
 from parlai.utils.io import PathManager
 import parlai.utils.logging as logging
@@ -75,6 +76,20 @@ class T5RagAgent(T5Agent, BaseGenerationAgentMixin):
     def build_rag_model(opt: Opt, dictionary: DictionaryAgent) -> T5RagModel:
         return T5RagModel(opt, dictionary)
 
+    def _generate(
+        self,
+        batch: Batch,
+        beam_size: int,
+        max_ts: int,
+        prefix_tokens: Optional[torch.LongTensor] = None,
+    ) -> Tuple[List[Tuple[torch.LongTensor, torch.Tensor]], List[TreeSearch]]:
+        """
+        Override since T5 needs to call TGA generate.
+        """
+        return TorchGeneratorAgent._generate(
+            self, batch, beam_size, max_ts, prefix_tokens
+        )
+
 
 GENERATION_AGENTS = {
     'transformer/generator': TransformerGeneratorRagAgent,
@@ -805,7 +820,12 @@ def compute_loss(
         scores, preds, enc_state, *_ = model_output
 
         self._record_retrieval_metrics(batch, enc_state)
-        loss, metric_loss, metric_correct, metric_target_tokens = self._rag_model_interface.compute_loss(
+        (
+            loss,
+            metric_loss,
+            metric_correct,
+            metric_target_tokens,
+        ) = self._rag_model_interface.compute_loss(
             self.criterion, scores, preds, enc_state, batch.label_vec
         )