[FasterGeneration] MBart supports dy2sta (#3356)

faster_generation/samples/mbart_sample.py

@@ -14,11 +14,10 @@
 import paddle
 from paddlenlp.transformers import MBartForConditionalGeneration, MBartTokenizer
 
-model_name = "mbart-large-50-one-to-many-mmt"
+model_name = "mbart-large-50-many-to-many-mmt"
 
-tokenizer = MBartTokenizer.from_pretrained(model_name)
-model = MBartForConditionalGeneration.from_pretrained(model_name,
-                                                      src_lang="en_XX")
+tokenizer = MBartTokenizer.from_pretrained(model_name, src_lang="en_XX")
+model = MBartForConditionalGeneration.from_pretrained(model_name)
 model.eval()
 
 
@@ -41,7 +40,7 @@ def postprocess_response(seq, bos_idx, eos_idx):
 
 inputs = "PaddleNLP is a powerful NLP library with Awesome pre-trained models and easy-to-use interface, supporting wide-range of NLP tasks from research to industrial applications."
 input_ids = tokenizer(inputs)["input_ids"]
-input_ids = paddle.to_tensor(input_ids, dtype='int64').unsqueeze(0)
+input_ids = paddle.to_tensor(input_ids, dtype='int32').unsqueeze(0)
 
 outputs, _ = model.generate(input_ids=input_ids,
                             forced_bos_token_id=bos_id,
@@ -53,5 +52,6 @@ def postprocess_response(seq, bos_idx, eos_idx):
 result = postprocess_response(outputs[0].numpy().tolist(), bos_id, eos_id)
 
 print("Model input:", inputs)
+
 print("Result:", result)
 # PaddleNLP是一个强大的NLP库,具有超乎寻常的预训练模型和易于使用的接口,支持从研究到工业应用的广泛的NLP任务。

paddlenlp/ops/faster_transformer/sample/mbart_export_model_sample.py

@@ -0,0 +1,147 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import argparse
+import paddle
+from pprint import pprint
+from paddlenlp.transformers import MBartForConditionalGeneration, MBartTokenizer
+from paddlenlp.ops import FasterMBART
+from paddlenlp.utils.log import logger
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_name_or_path",
+                        default="mbart-large-50-many-to-many-mmt",
+                        type=str,
+                        help="The model name to specify the bart to use. ")
+    parser.add_argument("--inference_model_dir",
+                        default="./infer_model/",
+                        type=str,
+                        help="Path to save inference model of bart. ")
+    parser.add_argument(
+        "--topk",
+        default=4,
+        type=int,
+        help="The number of candidate to procedure top_k sampling. ")
+    parser.add_argument(
+        "--topp",
+        default=1.0,
+        type=float,
+        help="The probability threshold to procedure top_p sampling. ")
+    parser.add_argument("--max_out_len",
+                        default=64,
+                        type=int,
+                        help="Maximum output length. ")
+    parser.add_argument("--temperature",
+                        default=1.0,
+                        type=float,
+                        help="The temperature to set. ")
+    parser.add_argument("--num_return_sequences",
+                        default=1,
+                        type=int,
+                        help="The number of returned sequences. ")
+    parser.add_argument("--use_fp16_decoding",
+                        action="store_true",
+                        help="Whether to use fp16 decoding to predict. ")
+    parser.add_argument("--decoding_strategy",
+                        default="beam_search",
+                        choices=["sampling", "beam_search"],
+                        type=str,
+                        help="The main strategy to decode. ")
+    parser.add_argument(
+        "--num_beams",
+        default=5,
+        type=int,
+        help="The number of candidate to procedure beam search. ")
+    parser.add_argument("--diversity_rate",
+                        default=0.0,
+                        type=float,
+                        help="The diversity rate to procedure beam search. ")
+    parser.add_argument("--repetition_penalty",
+                        default=1.0,
+                        type=float,
+                        help="The repetition_penalty to set. ")
+    parser.add_argument("--length_penalty",
+                        default=0.0,
+                        type=float,
+                        help="The length penalty to decode. ")
+    parser.add_argument("--early_stopping",
+                        action="store_true",
+                        help="Whether to do early stopping. ")
+
+    args = parser.parse_args()
+    return args
+
+
+def do_predict(args):
+    place = "gpu"
+    place = paddle.set_device(place)
+
+    model = MBartForConditionalGeneration.from_pretrained(
+        args.model_name_or_path, src_lang="en_XX")
+    tokenizer = MBartTokenizer.from_pretrained(args.model_name_or_path)
+
+    bos_id = tokenizer.lang_code_to_id["zh_CN"]
+    eos_id = model.mbart.config["eos_token_id"]
+
+    # For opening faster_encoder
+    model.eval()
+
+    faster_mbart = FasterMBART(model=model,
+                               use_fp16_decoding=args.use_fp16_decoding)
+    # Set evaluate mode
+    faster_mbart.eval()
+
+    # Convert dygraph model to static graph model
+    faster_mbart = paddle.jit.to_static(
+        faster_mbart,
+        input_spec=[
+            # input_ids
+            paddle.static.InputSpec(shape=[None, None], dtype="int32"),
+            # encoder_output
+            None,
+            # seq_len
+            None,
+            bos_id,  # forced_bos_token_id
+            args.num_beams,  # num_beams.
+            args.topk,  # top_k
+            args.topp,  # top_p
+            args.decoding_strategy,  # decode_strategy
+            tokenizer.bos_token_id,  # bos_token_id
+            tokenizer.eos_token_id,  # eos_token_id 
+            tokenizer.pad_token_id,  # pad_token_id 
+            model.mbart.
+            config["decoder_start_token_id"],  # decoder_start_token_id
+            args.max_out_len,  # max_length
+            args.diversity_rate,  # diversity_rate
+            args.length_penalty,  # length_penalty
+            args.temperature,  # temperature
+            args.num_return_sequences,  # num_return_sequences
+            args.early_stopping,  # early_stopping
+            tokenizer.eos_token_id,  #forced_eos_token_id
+        ])
+
+    # Save converted static graph model
+    paddle.jit.save(faster_mbart, os.path.join(args.inference_model_dir,
+                                               "mbart"))
+    logger.info("MBART has been saved to {}.".format(args.inference_model_dir))
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    pprint(args)
+
+    do_predict(args)

paddlenlp/ops/faster_transformer/sample/mbart_inference.py

@@ -0,0 +1,97 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import os
+import numpy as np
+from pprint import pprint
+
+import paddle
+import paddle.inference as paddle_infer
+
+from paddlenlp.transformers import MBartTokenizer
+from paddlenlp.ops.ext_utils import load
+
+
+def setup_args():
+    """Setup arguments."""
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--inference_model_dir",
+                        default="./infer_model/",
+                        type=str,
+                        help="Path to save inference model of BART. ")
+
+    args = parser.parse_args()
+
+    return args
+
+
+def postprocess_response(tokenizer, seq, bos_idx, eos_idx):
+    """Post-process the decoded sequence."""
+    eos_pos = len(seq) - 1
+    for i, idx in enumerate(seq):
+        if idx == eos_idx:
+            eos_pos = i
+            break
+    seq = [
+        idx for idx in seq[:eos_pos + 1] if idx != bos_idx and idx != eos_idx
+    ]
+    res = tokenizer.convert_ids_to_string(seq)
+    return res
+
+
+def infer(args):
+    model_name = "mbart-large-50-many-to-many-mmt"
+    tokenizer = MBartTokenizer.from_pretrained(model_name)
+
+    bos_id = tokenizer.lang_code_to_id["zh_CN"]
+    eos_id = tokenizer.eos_token_id
+
+    inputs = "PaddleNLP is a powerful NLP library with Awesome pre-trained models and easy-to-use interface, supporting wide-range of NLP tasks from research to industrial applications."
+    input_ids = tokenizer(inputs)["input_ids"]
+    input_ids = np.asarray(input_ids, dtype="int32").reshape(1, -1)
+
+    # Load FasterTransformer lib.
+    load("FasterTransformer", verbose=True)
+
+    config = paddle_infer.Config(
+        os.path.join(args.inference_model_dir, "mbart.pdmodel"),
+        os.path.join(args.inference_model_dir, "mbart.pdiparams"))
+
+    config.enable_use_gpu(100, 0)
+    config.disable_glog_info()
+    predictor = paddle_infer.create_predictor(config)
+
+    input_names = predictor.get_input_names()
+    input_handle = predictor.get_input_handle(input_names[0])
+    input_handle.copy_from_cpu(input_ids.astype("int32"))
+
+    predictor.run()
+
+    output_names = predictor.get_output_names()
+    output_handle = predictor.get_output_handle(output_names[0])
+    output_data = output_handle.copy_to_cpu()
+
+    result = postprocess_response(
+        tokenizer,
+        output_data.transpose([1, 2, 0]).tolist()[0][0], bos_id, eos_id)
+    print("Model input:", inputs)
+    print("Result:", result)
+
+
+if __name__ == "__main__":
+    args = setup_args()
+    pprint(args)
+
+    infer(args)

paddlenlp/ops/faster_transformer/transformer/decoding.py

@@ -2515,7 +2515,8 @@ def __init__(self,
         self.pos_emb = [model.decoder.decoder_embed_positions.weight]
         self.word_emb = [model.decoder.embed_tokens.weight]
 
-        self.linear_weight = [model.lm_head_weight.t()]
+        setattr(self, "lm_head_weight_", model.lm_head_weight.t())
+        self.linear_weight = [getattr(self, "lm_head_weight_")]
         self.linear_bias = [model.final_logits_bias]
 
     def forward(self,

paddlenlp/ops/faster_transformer/transformer/faster_transformer.py

@@ -1379,8 +1379,13 @@ def forward(self,
 
 
 class FasterMBART(MBartPretrainedModel):
+    enable_faster_encoder_func = enable_faster_encoder
 
-    def __init__(self, model, decoding_lib=None, use_fp16_decoding=False):
+    def __init__(self,
+                 model,
+                 decoding_lib=None,
+                 use_fp16_decoding=False,
+                 enable_faster_encoder=False):
         super(FasterMBART, self).__init__()
         self.use_fp16_decoding = use_fp16_decoding
         self._model = model
@@ -1393,13 +1398,18 @@ def __init__(self, model, decoding_lib=None, use_fp16_decoding=False):
         self.encoder = model.mbart.get_encoder()
         self.decoder = model.mbart.get_decoder()
         self.pad_token_id = model.mbart.config['pad_token_id']
+        self.enable_faster_encoder = enable_faster_encoder
 
         self.decoding = InferMBartDecoding(
             model=self._model,
             decoding_lib=decoding_lib,
             use_fp16_decoding=use_fp16_decoding,
             hidden_act=model.mbart.config['activation_function'])
 
+        if self.enable_faster_encoder:
+            # Must use `enable_faster_encoder` in `__init__` when dygraph to static graph.
+            self.encoder = FasterMBART.enable_faster_encoder_func(self.encoder)
+
     def get_encoder(self):
         return self.encoder
 
@@ -1439,11 +1449,9 @@ def forward(self,
 
         #(gongenlei) Not enable_faster_encoder temporarily
         if encoder_output is None:
-            self.encoder = enable_faster_encoder(self.encoder)
             assert input_ids is not None, "You have to specify either input_ids or encoder_output."
             encoder_output = self.prepare_encoder_decoder_kwargs_for_generation(
                 input_ids, model_kwargs)["encoder_output"]
-            self.encoder = disable_faster_encoder(self.encoder)
         batch_size = paddle.shape(encoder_output)[0]
         if seq_len is None:
             assert input_ids is not None, "You have to specify either input_ids when generating seq_len."

paddlenlp/transformers/mbart/modeling.py

@@ -203,7 +203,7 @@ def forward(self, input_ids_shape, past_key_values_length=0):
         positions = paddle.arange(past_key_values_length,
                                   past_key_values_length + seq_len,
                                   dtype="int64")
-        return super().forward(positions + self.offset)
+        return Embedding.forward(self, positions + self.offset)
 
 
 class MBartEncoder(MBartPretrainedModel):
@@ -270,7 +270,7 @@ def forward(self, input_ids=None, attention_mask=None, **kwargs):
         if input_ids is None:
             raise ValueError("Input_ids cannot be None.")
         inputs_embeds = self.d_model**0.5 * self.embed_tokens(input_ids)
-        inputs_embed_pos = self.encoder_embed_positions(input_ids.shape)
+        inputs_embed_pos = self.encoder_embed_positions(paddle.shape(input_ids))
         hidden_states = inputs_embeds + inputs_embed_pos
         hidden_states = self.encoder_layernorm_embedding(hidden_states)
         encoder_input = self.encoder_dropout(hidden_states)