remove load state as np. (PaddlePaddle#7120)

Author: ZHUI

examples/benchmark/ceval/model_evaluator.py

@@ -30,7 +30,7 @@ def __init__(self, choices, k, model_name_or_path, temperature=0.2):
         super().__init__(choices, model_name_or_path, k)
         self.model_name_or_path = model_name_or_path
         self.tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
-        self.model = AutoModelForCausalLM.from_pretrained(model_name_or_path, dtype="float16", low_cpu_mem_usage=True)
+        self.model = AutoModelForCausalLM.from_pretrained(model_name_or_path, dtype="float16")
         self.model.eval()
         self.generation_config = dict(
             temperature=temperature,

examples/benchmark/peft/paddle/benchmark.py

@@ -92,7 +92,6 @@ def main():
     if model_args.model_name_or_path in ["gpt3-6.7B-en", "gpt3-13B-en"]:
         model = GPTForCausalLM.from_pretrained(
             model_args.model_name_or_path,
-            low_cpu_mem_usage=True,
             use_flash_attention=model_args.use_flash_attention,
             dtype=dtype,
             tensor_parallel_degree=training_args.tensor_parallel_degree,
@@ -104,7 +103,6 @@ def main():
     else:
         model = AutoModelForCausalLM.from_pretrained(
             model_args.model_name_or_path,
-            low_cpu_mem_usage=True,
             use_flash_attention=model_args.use_flash_attention,
             dtype=dtype,
             tensor_parallel_degree=training_args.tensor_parallel_degree,

examples/benchmark/peft/paddle/inference_benchmark.py

@@ -67,8 +67,6 @@ def predict_forward(model, inputs):
     tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
     model = AutoModelForCausalLM.from_pretrained(
         args.model_name_or_path,
-        load_state_as_np=True,
-        low_cpu_mem_usage=True,
     )
     if model.base_model_prefix == "llama":
         tokenizer.pad_token = tokenizer.unk_token

examples/code_generation/codegen/README.md

@@ -119,7 +119,6 @@ python codegen_server.py
 - `min_length`：生成的最小长度，默认为0
 - `max_length`：生成的最大长度，默认为16
 - `decode_strategy`：解码策略，默认为"greedy_search"
-- `load_state_as_np`：以numpy格式加载模型参数，可节省显存，默认为True
 - `use_fast`：是否使用FastGeneration，可加速推理，默认为True
 - `use_fp16_decoding`：是否使用fp16推理，可节省显存和加速推理，默认为True
 

examples/code_generation/codegen/codegen_server.py

@@ -35,7 +35,6 @@ class DefaultConfig:
     min_length = 0
     max_length = 16
     decode_strategy = "greedy_search"
-    load_state_as_np = True
     use_faster = True
     use_fp16_decoding = True
     default_dtype = "float16" if use_faster and use_fp16_decoding else "float32"
@@ -64,9 +63,7 @@ class Output(BaseModel):
 paddle.set_default_dtype(generate_config.default_dtype)
 
 tokenizer = CodeGenTokenizer.from_pretrained(generate_config.model_name_or_path)
-model = CodeGenForCausalLM.from_pretrained(
-    generate_config.model_name_or_path, load_state_as_np=generate_config.load_state_as_np
-)
+model = CodeGenForCausalLM.from_pretrained(generate_config.model_name_or_path)
 
 app = FastAPI()
 

examples/language_model/t5/tests/t5_mp.py

@@ -44,7 +44,6 @@ def main():
         tensor_parallel_degree=tensor_parallel_degree,
         tensor_parallel_rank=tensor_parallel_rank,
         dtype="float32",
-        low_cpu_mem_usage=True,
     )
     model.eval()
     loss = model(
@@ -63,7 +62,6 @@ def main():
             tensor_parallel_degree=tensor_parallel_degree,
             tensor_parallel_rank=tensor_parallel_rank,
             dtype="float32",
-            low_cpu_mem_usage=True,
         )
         load_model.eval()
         loss = load_model(
@@ -85,7 +83,6 @@ def main():
             tensor_parallel_degree=tensor_parallel_degree,
             tensor_parallel_rank=tensor_parallel_rank,
             dtype="float32",
-            low_cpu_mem_usage=True,
         )
         load_model.eval()
         loss = load_model(

examples/text_generation/opt/demo.py

@@ -23,7 +23,7 @@ class Demo:
     def __init__(self, model_name_or_path, max_predict_len=128):
         self.tokenizer = GPTTokenizer.from_pretrained(model_name_or_path)
         logger.info("Loading the model parameters, please wait...")
-        self.model = OPTForCausalLM.from_pretrained(model_name_or_path, load_state_as_np=True)
+        self.model = OPTForCausalLM.from_pretrained(model_name_or_path)
         self.model.eval()
         self.max_predict_len = max_predict_len
         logger.info("Model loaded.")

fast_generation/perf/codegen_perf.py

@@ -37,7 +37,7 @@ def perf_pd(args):
     place = "gpu"
     place = paddle.set_device(place)
     tokenizer = CodeGenTokenizer.from_pretrained(args.model_name_or_path)
-    model = CodeGenForCausalLM.from_pretrained(args.model_name_or_path, load_state_as_np=True)
+    model = CodeGenForCausalLM.from_pretrained(args.model_name_or_path)
     model.eval()
     load_mem = query_by_id(args.gpu_id)
 

fast_generation/perf/pegasus_perf.py

@@ -40,7 +40,7 @@ def perf_pd(args):
     place = "gpu"
     place = paddle.set_device(place)
     tokenizer = PegasusChineseTokenizer.from_pretrained(args.model_name_or_path)
-    model = PegasusForConditionalGeneration.from_pretrained(args.model_name_or_path, load_state_as_np=True)
+    model = PegasusForConditionalGeneration.from_pretrained(args.model_name_or_path)
     model.eval()
     load_mem = query_by_id(args.gpu_id)
     input_ids_np = [np.random.choice(range(len(tokenizer.vocab)), args.input_len) for _ in range(args.batch_size)]

fast_generation/samples/codegen_16b_sample.py

@@ -21,7 +21,7 @@
 model_name = "Salesforce/codegen-16B-mono"
 
 tokenizer = CodeGenTokenizer.from_pretrained(model_name)
-model = CodeGenForCausalLM.from_pretrained(model_name, load_state_as_np=True)
+model = CodeGenForCausalLM.from_pretrained(model_name)
 model.eval()
 
 inputs = "def hello"

fast_generation/samples/gpt_mp_sample.py

@@ -96,7 +96,7 @@ def main(args):
     if args.profile:
         MODEL_CLASSES[model_name][0].generate = profile(args.batch_size)(MODEL_CLASSES[model_name][0].generate)
     tokenizer = MODEL_CLASSES[model_name][-1].from_pretrained(model_name)
-    model = MODEL_CLASSES[model_name][0].from_pretrained(model_name, load_state_as_np=True)
+    model = MODEL_CLASSES[model_name][0].from_pretrained(model_name)
     model.eval()
 
     # NOTE: When using prompt, open this and replace the text with what you want.

fast_generation/samples/gptj_sample.py

@@ -20,7 +20,7 @@
 model_name = "EleutherAI/gpt-j-6B"
 
 tokenizer = GPTJTokenizer.from_pretrained(model_name)
-model = GPTJForCausalLM.from_pretrained(model_name, load_state_as_np=True)
+model = GPTJForCausalLM.from_pretrained(model_name)
 model.eval()
 
 inputs = "What is PaddleNLP?"

fast_generation/samples/plato_xl_sample.py

@@ -106,7 +106,7 @@ def main(args):
     if args.profile:
         UnifiedTransformerLMHeadModel.generate = profile(args.batch_size)(UnifiedTransformerLMHeadModel.generate)
     tokenizer = UnifiedTransformerTokenizer.from_pretrained("plato-xl")
-    model = UnifiedTransformerLMHeadModel.from_pretrained("plato-xl", load_state_as_np=True)
+    model = UnifiedTransformerLMHeadModel.from_pretrained("plato-xl")
     model.eval()
 
     history = [

llm/ernie-3.5-se/predict_generation.py

@@ -99,7 +99,6 @@ def __init__(self, args=None, tokenizer=None, model=None, **kwargs):
                 args.model_name_or_path,
                 tensor_parallel_degree=tensor_parallel_degree,
                 tensor_parallel_rank=tensor_parallel_rank,
-                load_state_as_np=True,
                 dtype=dtype,
                 use_flash_attention=use_flash_attn,
             )

llm/ernie-3.5-se/run_pretrain.py

@@ -394,7 +394,6 @@ def main():
             model_args.model_name_or_path,
             config=config,
             dtype=dtype,
-            load_state_as_np=True,
             use_progressive_seq_len=True,
         )
     else:

llm/glm/finetune_generation.py

@@ -109,7 +109,6 @@ def main():
         model_args.model_name_or_path,
         output_predict=True,
         parallel_output=True,
-        load_state_as_np=True,
         dtype=dtype,  # todo enable set dtype to avoid additional mem usage
         tensor_parallel_degree=training_args.tensor_parallel_degree,
         tensor_parallel_rank=training_args.tensor_parallel_rank,

llm/glm/predict_generation.py

@@ -80,9 +80,7 @@ def __init__(self, args):
             args.model_name_or_path,
             tensor_parallel_degree=tensor_parallel_degree,
             tensor_parallel_rank=tensor_parallel_rank,
-            load_state_as_np=True,
             dtype=dtype,
-            low_cpu_mem_usage=True,
         )
         if self.args.lora_path is not None:
             self.model = LoRAModel.from_pretrained(self.model, self.args.lora_path)

llm/gpt-3/finetune_generation.py

@@ -153,7 +153,6 @@ def main():
         model_args.model_name_or_path,
         config=config,
         dtype=dtype,
-        load_state_as_np=True,
     )
     if model_args.lora:
         if model_args.lora_path is None:

llm/gpt-3/predict_generation.py

@@ -73,8 +73,6 @@ def __init__(self, args=None, tokenizer=None, model=None, **kwargs):
 
         self.model = GPTForCausalLM.from_pretrained(
             args.model_name_or_path,
-            load_state_as_np=True,
-            low_cpu_mem_usage=True,
             dtype=dtype,
             tensor_parallel_degree=tensor_parallel_degree,
             tensor_parallel_rank=tensor_parallel_rank,

model_zoo/plato-xl/infer.py

@@ -112,7 +112,7 @@ def main(args):
     if args.profile:
         UnifiedTransformerLMHeadModel.generate = profile(args.batch_size)(UnifiedTransformerLMHeadModel.generate)
     tokenizer = UnifiedTransformerTokenizer.from_pretrained("plato-xl")
-    model = UnifiedTransformerLMHeadModel.from_pretrained("plato-xl", load_state_as_np=True)
+    model = UnifiedTransformerLMHeadModel.from_pretrained("plato-xl")
     model.eval()
 
     history = [

paddlenlp/ops/fast_transformer/sample/plato_export_model_sample.py

@@ -70,7 +70,7 @@ def do_predict(args):
         paddle.set_default_dtype("float16")
 
     model_name = "plato-xl"
-    model = UnifiedTransformerLMHeadModel.from_pretrained(model_name, load_state_as_np=True)
+    model = UnifiedTransformerLMHeadModel.from_pretrained(model_name)
     tokenizer = UnifiedTransformerTokenizer.from_pretrained(model_name)
 
     plato = FasterUnifiedTransformer(model=model, use_fp16_decoding=args.use_fp16_decoding)

paddlenlp/taskflow/text2text_generation.py

@@ -117,7 +117,6 @@ def _construct_model(self, model):
         """
         model_instance = AutoModelForCausalLM.from_pretrained(
             self.model,
-            load_state_as_np=True,
             dtype=self._dtype,
         )
         # Load the model parameter for the predict

tests/transformers/test_modeling_utils.py

@@ -18,7 +18,7 @@
 from multiprocessing import Pool
 from tempfile import TemporaryDirectory
 
-from paddlenlp.transformers import BertModel, TinyBertModel
+from paddlenlp.transformers import BertModel
 from paddlenlp.utils.env import CONFIG_NAME, MODEL_HOME, PADDLE_WEIGHTS_NAME
 from tests.testing_utils import slow
 
@@ -57,12 +57,6 @@ def test_from_pretrained_cache_dir_pretrained_init(self):
             # check against double appending model_name in cache_dir
             self.assertFalse(os.path.exists(os.path.join(tempdir, model_name, model_name)))
 
-    @slow
-    def test_from_pretrained_with_load_as_state_np_params(self):
-        """init model with `load_state_as_np` params"""
-        model = TinyBertModel.from_pretrained("tinybert-4l-312d", load_state_as_np=True)
-        self.assertIsNotNone(model)
-
     @slow
     def test_multiprocess_downloading(self):
         """test downloading with multi-process. Some errors may be triggered when downloading model