Skip to content

Commit 113b883

Browse files
committed
add Qwen2Moe
1 parent 110983d commit 113b883

File tree

9 files changed

+2204
-2
lines changed

9 files changed

+2204
-2
lines changed

llm/data.py

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -44,11 +44,11 @@ def get_convert_example(model):
4444

4545
if base_model_prefix == "chatglm":
4646
return convert_example_chatglm
47-
elif base_model_prefix in ["chatglm_v2", "llama", "bloom", "opt", "qwen", "mixtral"]:
47+
elif base_model_prefix in ["chatglm_v2", "llama", "bloom", "opt", "qwen", "mixtral", "qwen2moe"]:
4848
return convert_example_common
4949
else:
5050
raise ValueError(
51-
f"Unknown base_model_prefix: {model.base_model_prefix}. Supported base_model_prefix list: chatglm, bloom, llama, qwen, mixtral"
51+
f"Unknown base_model_prefix: {model.base_model_prefix}. Supported base_model_prefix list: chatglm, bloom, llama, qwen, mixtral, qwen2moe"
5252
)
5353

5454

llm/qwen2moe/lora_argument.json

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,32 @@
1+
{
2+
"model_name_or_path": "Qwen/Qwen1.5-MoE-A2.7B",
3+
"dataset_name_or_path": "./data",
4+
"output_dir": "./checkpoints/qwen2moe_lora_ckpts",
5+
"per_device_train_batch_size": 4,
6+
"gradient_accumulation_steps": 4,
7+
"per_device_eval_batch_size": 8,
8+
"eval_accumulation_steps":16,
9+
"num_train_epochs": 3,
10+
"learning_rate": 3e-04,
11+
"warmup_steps": 30,
12+
"logging_steps": 1,
13+
"evaluation_strategy": "epoch",
14+
"save_strategy": "epoch",
15+
"src_length": 1024,
16+
"max_length": 2048,
17+
"fp16": true,
18+
"fp16_opt_level": "O2",
19+
"do_train": true,
20+
"do_eval": true,
21+
"disable_tqdm": true,
22+
"load_best_model_at_end": true,
23+
"eval_with_do_generation": false,
24+
"metric_for_best_model": "accuracy",
25+
"recompute": true,
26+
"save_total_limit": 1,
27+
"tensor_parallel_degree": 8,
28+
"pipeline_parallel_degree": 1,
29+
"lora": true,
30+
"zero_padding": false,
31+
"use_flash_attention": false
32+
}

llm/qwen2moe/sft_argument.json

Lines changed: 30 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,30 @@
1+
{
2+
"model_name_or_path": "Qwen/Qwen1.5-MoE-A2.7B",
3+
"dataset_name_or_path": "./data",
4+
"output_dir": "./checkpoints/qwen2moe_sft_ckpts",
5+
"per_device_train_batch_size": 4,
6+
"gradient_accumulation_steps": 4,
7+
"per_device_eval_batch_size": 8,
8+
"eval_accumulation_steps":16,
9+
"num_train_epochs": 3,
10+
"learning_rate": 3e-05,
11+
"warmup_steps": 30,
12+
"logging_steps": 1,
13+
"evaluation_strategy": "epoch",
14+
"save_strategy": "epoch",
15+
"src_length": 1024,
16+
"max_length": 2048,
17+
"bf16": true,
18+
"fp16_opt_level": "O2",
19+
"do_train": true,
20+
"do_eval": true,
21+
"disable_tqdm": true,
22+
"load_best_model_at_end": true,
23+
"eval_with_do_generation": false,
24+
"metric_for_best_model": "accuracy",
25+
"recompute": true,
26+
"save_total_limit": 1,
27+
"tensor_parallel_degree": 8,
28+
"sharding": "stage2",
29+
"pipeline_parallel_degree": 1
30+
}

paddlenlp/transformers/auto/modeling.py

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -118,6 +118,7 @@
118118
("Bloom", "bloom"),
119119
("QWen", "qwen"),
120120
("Mixtral", "mixtral"),
121+
("QWen2Moe", "qwen2moe"),
121122
]
122123
)
123124

Lines changed: 16 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,16 @@
1+
# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
2+
#
3+
# Licensed under the Apache License, Version 2.0 (the "License");
4+
# you may not use this file except in compliance with the License.
5+
# You may obtain a copy of the License at
6+
#
7+
# http://www.apache.org/licenses/LICENSE-2.0
8+
#
9+
# Unless required by applicable law or agreed to in writing, software
10+
# distributed under the License is distributed on an "AS IS" BASIS,
11+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+
# See the License for the specific language governing permissions and
13+
# limitations under the License.
14+
15+
from .configuration import Qwen2MoeConfig
16+
from .modeling import Qwen2MoeForCausalLM
Lines changed: 203 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,203 @@
1+
# coding=utf-8
2+
# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
3+
#
4+
# Licensed under the Apache License, Version 2.0 (the "License");
5+
# you may not use this file except in compliance with the License.
6+
# You may obtain a copy of the License at
7+
#
8+
# http://www.apache.org/licenses/LICENSE-2.0
9+
#
10+
# Unless required by applicable law or agreed to in writing, software
11+
# distributed under the License is distributed on an "AS IS" BASIS,
12+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13+
# See the License for the specific language governing permissions and
14+
# limitations under the License.
15+
""" Qwen2MoE model configuration"""
16+
17+
from paddlenlp.transformers.configuration_utils import PretrainedConfig
18+
19+
__all__ = [
20+
"Qwen2MoeConfig",
21+
]
22+
23+
24+
class Qwen2MoeConfig(PretrainedConfig):
25+
r"""
26+
This is the configuration class to store the configuration of a [`Qwen2MoeModel`]. It is used to instantiate a
27+
Qwen2MoE model according to the specified arguments, defining the model architecture. Instantiating a configuration
28+
with the defaults will yield a similar configuration to that of
29+
Qwen1.5-MoE-A2.7B" [Qwen/Qwen1.5-MoE-A2.7B"](https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B").
30+
31+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
32+
documentation from [`PretrainedConfig`] for more information.
33+
34+
35+
Args:
36+
vocab_size (`int`, *optional*, defaults to 151936):
37+
Vocabulary size of the Qwen2MoE model. Defines the number of different tokens that can be represented by the
38+
`inputs_ids` passed when calling [`Qwen2MoeModel`]
39+
hidden_size (`int`, *optional*, defaults to 2048):
40+
Dimension of the hidden representations.
41+
intermediate_size (`int`, *optional*, defaults to 5632):
42+
Dimension of the MLP representations.
43+
num_hidden_layers (`int`, *optional*, defaults to 24):
44+
Number of hidden layers in the Transformer encoder.
45+
num_attention_heads (`int`, *optional*, defaults to 16):
46+
Number of attention heads for each attention layer in the Transformer encoder.
47+
num_key_value_heads (`int`, *optional*, defaults to 16):
48+
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
49+
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
50+
`num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
51+
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
52+
by meanpooling all the original heads within that group. For more details checkout [this
53+
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
54+
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
55+
The non-linear activation function (function or string) in the decoder.
56+
max_position_embeddings (`int`, *optional*, defaults to 32768):
57+
The maximum sequence length that this model might ever be used with.
58+
initializer_range (`float`, *optional*, defaults to 0.02):
59+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
60+
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
61+
The epsilon used by the rms normalization layers.
62+
use_cache (`bool`, *optional*, defaults to `True`):
63+
Whether or not the model should return the last key/values attentions (not used by all models). Only
64+
relevant if `config.is_decoder=True`.
65+
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
66+
Whether the model's input and output word embeddings should be tied.
67+
rope_theta (`float`, *optional*, defaults to 10000.0):
68+
The base period of the RoPE embeddings.
69+
use_sliding_window (`bool`, *optional*, defaults to `False`):
70+
Whether to use sliding window attention.
71+
sliding_window (`int`, *optional*, defaults to 4096):
72+
Sliding window attention (SWA) window size. If not specified, will default to `4096`.
73+
max_window_layers (`int`, *optional*, defaults to 28):
74+
The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
75+
attention_dropout (`float`, *optional*, defaults to 0.0):
76+
The dropout ratio for the attention probabilities.
77+
decoder_sparse_step (`int`, *optional*, defaults to 1):
78+
The frequency of the MoE layer.
79+
moe_intermediate_size (`int`, *optional*, defaults to 1408):
80+
Intermediate size of the routed expert.
81+
shared_expert_intermediate_size (`int`, *optional*, defaults to 5632):
82+
Intermediate size of the shared expert.
83+
num_experts_per_tok (`int`, *optional*, defaults to 4):
84+
Number of selected experts.
85+
num_experts (`int`, *optional*, defaults to 60):
86+
Number of routed experts.
87+
norm_topk_prob (`bool`, *optional*, defaults to `False`):
88+
Whether to normalize the topk probabilities.
89+
output_router_logits (`bool`, *optional*, defaults to `False`):
90+
Whether or not the router logits should be returned by the model. Enabeling this will also
91+
allow the model to output the auxiliary loss, including load balancing loss and router z-loss.
92+
router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
93+
The aux loss factor for the total loss.
94+
95+
```python
96+
>>> from paddlenlp.transformers import Qwen2MoeModel, Qwen2MoeConfig
97+
98+
>>> # Initializing a Qwen2MoE style configuration
99+
>>> configuration = Qwen2MoeConfig()
100+
101+
>>> # Initializing a model from the Qwen1.5-MoE-A2.7B" style configuration
102+
>>> model = Qwen2MoeModel(configuration)
103+
104+
>>> # Accessing the model configuration
105+
>>> configuration = model.config
106+
```"""
107+
108+
model_type = "qwen2_moe"
109+
keys_to_ignore_at_inference = ["past_key_values"]
110+
111+
def __init__(
112+
self,
113+
vocab_size=151936,
114+
hidden_size=2048,
115+
intermediate_size=5632,
116+
num_hidden_layers=24,
117+
num_attention_heads=16,
118+
num_key_value_heads=16,
119+
hidden_act="silu",
120+
max_position_embeddings=32768,
121+
seq_length=2048,
122+
initializer_range=0.02,
123+
rms_norm_eps=1e-6,
124+
use_cache=True,
125+
use_recompute=False,
126+
recompute_granularity="full",
127+
no_recompute_layers=None,
128+
use_flash_attention=False,
129+
attention_dropout=0.0,
130+
use_fused_rope=False,
131+
rope_theta=10000.0,
132+
tensor_parallel_output=True,
133+
sequence_parallel=False,
134+
fuse_sequence_parallel_allreduce=False,
135+
pad_token_id=0,
136+
bos_token_id=1,
137+
eos_token_id=2,
138+
tie_word_embeddings=False,
139+
use_sliding_window=False,
140+
sliding_window=4096,
141+
max_window_layers=28,
142+
decoder_sparse_step=1,
143+
moe_intermediate_size=1408,
144+
shared_expert_intermediate_size=5632,
145+
num_experts_per_tok=4,
146+
num_experts=60,
147+
norm_topk_prob=False,
148+
output_router_logits=False,
149+
router_aux_loss_coef=0.001,
150+
**kwargs,
151+
):
152+
self.vocab_size = vocab_size
153+
self.max_position_embeddings = max_position_embeddings
154+
self.seq_length = seq_length
155+
self.hidden_size = hidden_size
156+
self.intermediate_size = intermediate_size
157+
self.num_hidden_layers = num_hidden_layers
158+
self.num_attention_heads = num_attention_heads
159+
self.use_sliding_window = use_sliding_window
160+
self.sliding_window = sliding_window
161+
self.max_window_layers = max_window_layers
162+
163+
self.num_key_value_heads = num_key_value_heads
164+
self.hidden_act = hidden_act
165+
166+
self.initializer_range = initializer_range
167+
self.rms_norm_eps = rms_norm_eps
168+
169+
self.use_cache = use_cache
170+
self.use_recompute = use_recompute
171+
self.recompute_granularity = recompute_granularity
172+
self.no_recompute_layers = no_recompute_layers
173+
self.use_flash_attention = use_flash_attention
174+
self.tensor_parallel_output = tensor_parallel_output
175+
self.sequence_parallel = sequence_parallel
176+
self.fuse_sequence_parallel_allreduce = fuse_sequence_parallel_allreduce
177+
178+
self.pad_token_id = pad_token_id
179+
self.bos_token_id = bos_token_id
180+
self.eos_token_id = eos_token_id
181+
182+
self.use_fused_rope = use_fused_rope
183+
self.rope_theta = rope_theta
184+
self.attention_dropout = attention_dropout
185+
186+
# MoE arguments
187+
self.decoder_sparse_step = decoder_sparse_step
188+
self.moe_intermediate_size = moe_intermediate_size
189+
self.shared_expert_intermediate_size = shared_expert_intermediate_size
190+
self.num_experts_per_tok = num_experts_per_tok
191+
self.num_experts = num_experts
192+
self.norm_topk_prob = norm_topk_prob
193+
self.output_router_logits = output_router_logits
194+
self.router_aux_loss_coef = router_aux_loss_coef
195+
196+
super().__init__(
197+
pad_token_id=pad_token_id,
198+
bos_token_id=bos_token_id,
199+
eos_token_id=eos_token_id,
200+
tie_word_embeddings=tie_word_embeddings,
201+
tensor_parallel_output=tensor_parallel_output,
202+
**kwargs,
203+
)

0 commit comments

Comments
 (0)