Files
FastDeploy/scripts/convert_ep_to_safetensor.py
2025-06-29 23:29:37 +00:00

252 lines
7.9 KiB
Python

"""
# Copyright (c) 2025 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 paddle
import os
from paddlenlp.trainer import strtobool
from efficientllm.models.utils import load_checkpoint
from efficientllm.inference_args import InferenceArgs
from paddlenlp.utils.log import logger
from efficientllm.models.configuration import ErnieBotConfig
from efficientllm.models.tokenizer import ErnieBotTokenizer
from safetensors.numpy import save_file as safe_save_file
from paddlenlp.utils.env import SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_NAME
import shutil
import argparse
import importlib
import json
from paddlenlp.transformers.model_utils import shard_checkpoint
MODEL_LIB_NAMES = [
"efficientllm.models.modeling_ernie_bot",
]
def parse_arguments():
"""
parse_arguments
"""
parser = argparse.ArgumentParser()
parser.add_argument(
"--model_name_or_path",
default=None,
required=True,
help="The directory of model.",
)
parser.add_argument(
"--output_dir",
default="merged_output",
required=True,
help="The directory of merged model output.",
)
parser.add_argument(
"--safe_serialization",
type=strtobool,
default="True",
help="Whether merge the model into safetensors format.",
)
parser.add_argument(
"--predict_model_type",
type=str,
default="",
help="Quantization type for the model.",
)
parser.add_argument(
"--draft_type",
type=str,
default=None,
choices=["autoregressive", "inference_with_reference", "hydra", "mtp"],
help="Quantization type for the model.",
)
parser.add_argument(
"--moe_quant_type",
default="default",
type=str,
choices=["weight_only_int4", "weight_only_int8", "w4a8", "fp8", "default"],
help="quant type for moe part",
)
parser.add_argument(
"--use_ep",
type=strtobool,
default="True",
help="Whether merge the model into safetensors format.",
)
parser.add_argument("--dtype", type=str, default="bfloat16")
return parser.parse_args()
def get_model_cls(config):
"""
Get model class from model configuration.
"""
init_class = "ErnieBotFusedModel"
for lib_name in MODEL_LIB_NAMES:
eb_lib = importlib.import_module(lib_name)
if hasattr(eb_lib, init_class):
cls = getattr(eb_lib, init_class)
return cls
raise RuntimeError(f"Cannot find model architecture({init_class}) from eb_lib")
def save_safetensors(state_dict, args):
"""
save_safetensors
"""
logger.info("Move to numpy.")
for k in list(state_dict.keys()):
if isinstance(state_dict[k], paddle.Tensor):
state_dict[k] = state_dict.pop(k).cpu().numpy()
logger.info("Save safetensors files.")
shards, index = shard_checkpoint(
state_dict,
max_shard_size="5GB",
weights_name=SAFE_WEIGHTS_NAME,
shard_format="naive",
)
for shard_file, shard in shards.items():
save_file = os.path.join(args.output_dir, shard_file)
logger.info(f"Saving {save_file}")
safe_save_file(shard, save_file, metadata={"format": "np"})
save_index_file = os.path.join(args.output_dir, SAFE_WEIGHTS_INDEX_NAME)
with open(save_index_file, "w", encoding="utf-8") as f:
content = json.dumps(index, indent=2) + "\n"
f.write(content)
def quanted_tensor(cls, state_dict, config):
"""
quanted_tensor
"""
name_action_mappings = cls._get_tensor_quantization_mappings(config)
state_keys_map = cls._resolve_prefix_keys(
name_action_mappings.keys(), state_dict.keys()
)
for k, v in state_keys_map.items():
name_action_mappings[v] = name_action_mappings.pop(k)
state_dict_to_save = {}
from efficientllm.layers.utils import get_tensor
from tqdm import tqdm
for key in tqdm(state_dict.keys(), desc="process quantized weights "):
tensor_path = state_dict[key]
if key in name_action_mappings:
ret = state_dict[key]
action = name_action_mappings.pop(key)
quanted_weight_tensor, weight_scale_tensor = action(get_tensor(ret))
if quanted_weight_tensor._is_initialized():
state_dict_to_save[key + ".quant_weight"] = quanted_weight_tensor.cpu()
if weight_scale_tensor._is_initialized():
state_dict_to_save[key + ".quant_scale"] = weight_scale_tensor.cpu()
else:
state_dict_to_save[key] = quanted_weight_tensor.cpu()
else:
state_dict_to_save[key] = get_tensor(tensor_path).cpu()
if len(name_action_mappings) > 0:
for x in name_action_mappings.keys():
logger.debug(
f"key <{x}> need to merge tensor parallel but we can't find in model state."
)
return state_dict_to_save
def get_quant_type(args):
"""
get_quant_type
"""
quant_type = args.predict_model_type.lower()
if quant_type == "default":
quant_type = ""
moe_quant_type = args.moe_quant_type.lower()
if moe_quant_type == "default":
moe_quant_type = ""
paddle.set_default_dtype(args.dtype)
offline_args = InferenceArgs(
quant_type=quant_type,
num_layers=1,
num_attention_heads=1,
num_key_value_heads=1,
hidden_size=1,
ffn_hidden_size=1,
mp_rank=1,
mp_size=1,
)
weight_dtype, act_dtype, cachekv_dtype = (
offline_args.weight_dtype,
offline_args.act_dtype,
offline_args.cachekv_dtype,
)
return weight_dtype, act_dtype, cachekv_dtype, quant_type, moe_quant_type
def main():
"""
main
"""
args = parse_arguments()
tokenizer = ErnieBotTokenizer.from_pretrained(args.model_name_or_path)
config = ErnieBotConfig.from_pretrained(args.model_name_or_path)
(
config.weight_dtype,
config.act_dtype,
config.cachekv_dtype,
config.quant_type,
config.moe_quant_type,
) = get_quant_type(args)
config.is_mtp = args.draft_type in ["eagle", "mtp"]
config.use_ep = args.use_ep
cls = get_model_cls(config)
# load
state_dict = load_checkpoint(
args.model_name_or_path, cls, config, return_numpy=True
)
import time
start = time.perf_counter()
state_dict_to_save = quanted_tensor(cls=cls, state_dict=state_dict, config=config)
end = time.perf_counter()
logger.info("Finish Quantize.")
logger.info(f"load和量化耗时: {end - start:.6f}")
logger.info("Begin to save model")
os.makedirs(args.output_dir, exist_ok=True)
start = time.perf_counter()
if not args.safe_serialization:
paddle.save(
state_dict_to_save,
os.path.join(args.output_dir, "model_state.pdparams"),
)
else:
save_safetensors(state_dict_to_save, args)
config.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
if config.moe_quant_type == "w4a8":
# cp act_scales.json
shutil.copy(args.model_name_or_path + '/act_scales.json', args.output_dir)
shutil.copy(args.model_name_or_path + '/weight_scales.json', args.output_dir)
end = time.perf_counter()
logger.info(f"save耗时: {end - start:.6f}")
logger.info("Finish.")
if __name__ == "__main__":
main()