From cc58a2390f64f7cded6c1de525a4d04d0357ec6e Mon Sep 17 00:00:00 2001
From: smallv0221 <33639025+smallv0221@users.noreply.github.com>
Date: Thu, 16 Dec 2021 21:53:15 +0800
Subject: [PATCH] Fix diversity rate bug (#1477)

* update perf

* fix doc and constrains for FasterGeneration

* update readme

* fix diversity rate bug
---
 examples/faster/faster_generation/samples/unimo_sample.py | 2 +-
 paddlenlp/ops/faster_transformer/transformer/decoding.py  | 6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/examples/faster/faster_generation/samples/unimo_sample.py b/examples/faster/faster_generation/samples/unimo_sample.py
index e55a23cd7f10..cd82c0b39166 100644
--- a/examples/faster/faster_generation/samples/unimo_sample.py
+++ b/examples/faster/faster_generation/samples/unimo_sample.py
@@ -41,7 +41,7 @@ def postprocess_response(token_ids, tokenizer):
     add_start_token_for_decoding=True,
     return_tensors=True,
     is_split_into_words=False)
-model.eval()
+
 outputs, _ = model.generate(
     input_ids=inputs_ids['input_ids'],
     token_type_ids=inputs_ids['token_type_ids'],
diff --git a/paddlenlp/ops/faster_transformer/transformer/decoding.py b/paddlenlp/ops/faster_transformer/transformer/decoding.py
index e0f662990b07..8848e0b0e112 100644
--- a/paddlenlp/ops/faster_transformer/transformer/decoding.py
+++ b/paddlenlp/ops/faster_transformer/transformer/decoding.py
@@ -1381,7 +1381,7 @@ def forward(self,
             _bos_id=bos_token_id,
             _eos_id=eos_token_id,
             _max_out_len=max_out_len,
-            _diversity_rate=diversity_rate,
+            _diversity_rate=-diversity_rate,
             _unk_id=self._unk_id,
             _mask_id=self._mask_id,
             _temperature=temperature,
@@ -1625,7 +1625,7 @@ def forward(self,
             self.linear_weight, self.linear_bias, self.pos_emb,
             decoding_strategy, beam_size, top_k, top_p, self._n_head,
             int(self._d_model / self._n_head), self._num_decoder_layers,
-            bos_token_id, eos_token_id, max_out_len, diversity_rate, rel_len,
+            bos_token_id, eos_token_id, max_out_len, -diversity_rate, rel_len,
             alpha, early_stopping)
 
         ids = finalize(
@@ -1877,7 +1877,7 @@ def forward(self,
             self.linear_bias, self.pos_emb, trg_word, decoding_strategy,
             beam_size, top_k, top_p, self._n_head,
             int(self._d_model / self._n_head), self._num_decoder_layers,
-            bos_token_id, eos_token_id, max_out_len, diversity_rate, rel_len,
+            bos_token_id, eos_token_id, max_out_len, -diversity_rate, rel_len,
             alpha, temperature, early_stopping, self._hidden_act)
 
         ids = finalize(