Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[INFER][LLM] Add the AutoPredictor for inference #9445

Merged
merged 9 commits into from
Dec 3, 2024
Merged
Changes from 1 commit
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
decoupling the model loading and predictor loading
zeroRains committed Nov 26, 2024
commit 483fff732b3287ca4daf7f34987fb3f12ebc772f
142 changes: 74 additions & 68 deletions llm/predict/predictor.py
Original file line number Diff line number Diff line change
@@ -1182,10 +1182,8 @@ def create_predictor(
Returns:
Predictor: The predictor.
"""
tensor_parallel_degree = kwargs.pop("tensor_parallel_degree", 1)
tensor_parallel_rank = kwargs.pop("tensor_parallel_rank", 0)
model = None
cache_kvs_shape = None
model = kwargs.pop("model",None)
cache_kvs_shape = kwargs.pop("cache_kvs_shape",None)

# static or dynamic
execute_mode = "Dygraph" if predictor_args.mode == "dynamic" else "StaticGraph"
@@ -1194,70 +1192,9 @@ def create_predictor(
if predictor_args.inference_model:
# block/no block
inference_mode = f"{'Block' if predictor_args.block_attn else ''}Inference"
if execute_mode == "Dygraph":
# AutoInferenceModel
model = AutoInferenceModelForCausalLM.from_pretrained(
predictor_args.model_name_or_path,
config=config,
predictor_args=predictor_args,
model_args=model_args,
dtype=predictor_args.dtype,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
)
model.eval()
else:
# cache_kvs_shape compute
model = AutoInferenceModelForCausalLM.from_pretrained(
predictor_args.model_name_or_path,
config=config,
predictor_args=predictor_args,
model_args=model_args,
dtype=predictor_args.dtype,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
)
cache_kvs_shape = model.get_cache_kvs_shape(
config, predictor_args.batch_size, predictor_args.total_max_length
)
else:
inference_mode = ""
if execute_mode == "Dygraph":
# model import (gpt-3,ernie) or AutoModel
if model_args.model_type == "gpt-3":
sys.path.append("./gpt-3")
from modeling import GPTForCausalLM

model = GPTForCausalLM.from_pretrained(
predictor_args.model_name_or_path,
dtype=predictor_args.dtype,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
tensor_parallel_output=False,
)
elif model_args.model_type == "ernie-3.5-se":
sys.path.append("./ernie-3.5-se")
from modeling import Ernie35ForCausalLM

tensor_parallel_degree = paddle.distributed.get_world_size()
tensor_parallel_rank = paddle.distributed.get_rank()
model = Ernie35ForCausalLM.from_pretrained(
predictor_args.model_name_or_path,
dtype=predictor_args.dtype,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
tensor_parallel_output=False,
)
else:
model = AutoModelForCausalLM.from_pretrained(
predictor_args.model_name_or_path,
dtype=predictor_args.dtype,
use_flash_attention=predictor_args.use_flash_attention,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
tensor_parallel_output=False,
)


predictor_class_name = execute_mode + inference_mode + "Predictor"

import_class = sys.modules[__name__]
@@ -1308,13 +1245,82 @@ def create_predictor(
predictor_args.temperature = 1.0

tensor_parallel_rank, tensor_parallel_degree = llm_utils.init_dist_env()

model = None
cache_kvs_shape = None

# model loading
if predictor_args.inference_model:
if predictor_args.mode == "dynamic":
# AutoInferenceModel
model = AutoInferenceModelForCausalLM.from_pretrained(
predictor_args.model_name_or_path,
config=config,
predictor_args=predictor_args,
model_args=model_args,
dtype=predictor_args.dtype,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
)
model.eval()
else:
# cache_kvs_shape compute
model = AutoInferenceModelForCausalLM.from_pretrained(
predictor_args.model_name_or_path,
config=config,
predictor_args=predictor_args,
model_args=model_args,
dtype=predictor_args.dtype,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
)
cache_kvs_shape = model.get_cache_kvs_shape(
config, predictor_args.batch_size, predictor_args.total_max_length
)
else:
if predictor_args.mode == "dynamic":
# model import (gpt-3,ernie) or AutoModel
if model_args.model_type == "gpt-3":
sys.path.append("./gpt-3")
from modeling import GPTForCausalLM

model = GPTForCausalLM.from_pretrained(
predictor_args.model_name_or_path,
dtype=predictor_args.dtype,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
tensor_parallel_output=False,
)
elif model_args.model_type == "ernie-3.5-se":
sys.path.append("./ernie-3.5-se")
from modeling import Ernie35ForCausalLM

tensor_parallel_degree = paddle.distributed.get_world_size()
tensor_parallel_rank = paddle.distributed.get_rank()
model = Ernie35ForCausalLM.from_pretrained(
predictor_args.model_name_or_path,
dtype=predictor_args.dtype,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
tensor_parallel_output=False,
)
else:
model = AutoModelForCausalLM.from_pretrained(
predictor_args.model_name_or_path,
dtype=predictor_args.dtype,
use_flash_attention=predictor_args.use_flash_attention,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
tensor_parallel_output=False,
)

predictor = AutoPredictor.create_predictor(
predictor_args,
config,
model_args,
tokenizer,
tensor_parallel_degree=tensor_parallel_degree,
tensor_parallel_rank=tensor_parallel_rank,
model=model,
cache_kvs_shape=cache_kvs_shape
)

return predictor
2 changes: 1 addition & 1 deletion paddlenlp/transformers/auto/modeling.py
Original file line number Diff line number Diff line change
@@ -837,7 +837,7 @@ def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):

# Import the InferenceModel
import_class = importlib.import_module(
f"paddlenlp.experimental.transformers.{cls._name_mapping[config.architectures[0]]}.modeling"
f"paddlenlp.experimental.transformers.{config.model_type}.modeling"
)

model_class_name = f"{model_name}InferenceModel"