mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2025-10-05 16:48:03 +08:00
86 lines
3.2 KiB
Python
86 lines
3.2 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 paddle.distributed as dist
|
|
import pdb
|
|
from glob import glob
|
|
import os
|
|
import numpy as np
|
|
|
|
import argparse
|
|
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--model_dir", type=str, required=True)
|
|
args = parser.parse_args()
|
|
|
|
rank = dist.get_rank()
|
|
print("rank: ", rank)
|
|
# merge tpn -> tp1
|
|
model_dir = args.model_dir
|
|
model_path_pp = glob(os.path.join(model_dir, f"pp{rank}"))
|
|
model_path_pp_tp = []
|
|
for p in model_path_pp:
|
|
model_path_tp = glob(os.path.join(p, "model_state*"))
|
|
model_path_tp = sorted(model_path_tp)
|
|
save_merged_pp_path = os.path.join(p, "merged_tp1_state.pdparams")
|
|
save_merged_pp_dir = os.path.join(model_dir, "merged_tp1_state_split")
|
|
os.makedirs(save_merged_pp_dir, exist_ok=True)
|
|
print(p, model_path_tp)
|
|
|
|
state_dicts = [paddle.load(path, return_numpy=True) for path in model_path_tp]
|
|
state = state_dicts[0]
|
|
|
|
print("merge tp")
|
|
print("p: ", p)
|
|
for k, v in state.items():
|
|
save_split_path = os.path.join(save_merged_pp_dir, k)
|
|
state_now = []
|
|
for i in range(len(state_dicts)):
|
|
state_now.append(state_dicts[i][k])
|
|
print("k: ", k, ", v.shape: ", v.shape)
|
|
if "qkv_proj" in k:
|
|
"""not need prmt"""
|
|
# qkv not prmt
|
|
ori_q = [s[:, :1024] for s in state_now]
|
|
ori_k = [s[:, 1024:1152] for s in state_now]
|
|
ori_v = [s[:, 1152:] for s in state_now]
|
|
new_q = np.concatenate(ori_q, axis=1)
|
|
new_k = np.concatenate(ori_k, axis=1)
|
|
new_v = np.concatenate(ori_v, axis=1)
|
|
print(new_q.shape)
|
|
print(new_k.shape)
|
|
print(new_v.shape)
|
|
new_w = np.concatenate([new_q, new_k, new_v], axis=1)
|
|
# new_w = np.concatenate(state_now, axis=1)
|
|
elif "o_proj" in k or "down_proj" in k:
|
|
new_w = np.concatenate(state_now, axis=0)
|
|
elif "embed_tokens" in k:
|
|
new_w = np.concatenate(state_now, axis=0)
|
|
elif "up_gate_proj" in k:
|
|
dim = state_now[0].shape[1]
|
|
half_ffn1_1 = [s[:, :(dim // 2)] for s in state_now]
|
|
half_ffn1_2 = [s[:, (dim // 2):] for s in state_now]
|
|
new_ffn1_1 = np.concatenate(half_ffn1_1, axis=1)
|
|
new_ffn1_2 = np.concatenate(half_ffn1_2, axis=1)
|
|
new_w = np.concatenate([new_ffn1_1, new_ffn1_2], axis=1)
|
|
elif "lm_head" in k or "mtp_linear_proj" in k:
|
|
new_w = np.concatenate(state_now, axis=1)
|
|
else:
|
|
new_w = v
|
|
print("merged_shape: ", new_w.shape)
|
|
paddle.save(paddle.to_tensor(new_w), save_split_path)
|
|
print("merge end") |