jasion commited on
Commit
a6beedb
1 Parent(s): ac7c60d

Upload 9 files

Browse files
config.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "model/new_internlm-7b",
3
+ "architectures": [
4
+ "InternLMForCausalLM"
5
+ ],
6
+ "auto_map": {
7
+ "AutoConfig": "configuration_internlm.InternLMConfig",
8
+ "AutoModel": "modeling_internlm.InternLMForCausalLM",
9
+ "AutoModelForCausalLM": "modeling_internlm.InternLMForCausalLM"
10
+ },
11
+ "bias": true,
12
+ "bos_token_id": 1,
13
+ "eos_token_id": 2,
14
+ "hidden_act": "silu",
15
+ "hidden_size": 4096,
16
+ "initializer_range": 0.02,
17
+ "intermediate_size": 11008,
18
+ "max_position_embeddings": 2048,
19
+ "model_type": "internlm",
20
+ "num_attention_heads": 32,
21
+ "num_hidden_layers": 32,
22
+ "pad_token_id": 2,
23
+ "rms_norm_eps": 1e-06,
24
+ "rotary": {
25
+ "base": 10000,
26
+ "type": "dynamic"
27
+ },
28
+ "tie_word_embeddings": false,
29
+ "torch_dtype": "float16",
30
+ "transformers_version": "4.36.2",
31
+ "use_cache": true,
32
+ "vocab_size": 103168
33
+ }
configuration_internlm.py ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+ """ InternLM model configuration"""
21
+
22
+ from transformers.configuration_utils import PretrainedConfig
23
+ from transformers.utils import logging
24
+
25
+ logger = logging.get_logger(__name__)
26
+
27
+ INTERNLM_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
28
+
29
+
30
+ class InternLMConfig(PretrainedConfig):
31
+ r"""
32
+ This is the configuration class to store the configuration of a [`InternLMModel`]. It is used to instantiate
33
+ an InternLM model according to the specified arguments, defining the model architecture. Instantiating a
34
+ configuration with the defaults will yield a similar configuration to that of the InternLM-7B.
35
+
36
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
37
+ documentation from [`PretrainedConfig`] for more information.
38
+
39
+
40
+ Args:
41
+ vocab_size (`int`, *optional*, defaults to 32000):
42
+ Vocabulary size of the InternLM model. Defines the number of different tokens that can be represented by the
43
+ `inputs_ids` passed when calling [`InternLMModel`]
44
+ hidden_size (`int`, *optional*, defaults to 4096):
45
+ Dimension of the hidden representations.
46
+ intermediate_size (`int`, *optional*, defaults to 11008):
47
+ Dimension of the MLP representations.
48
+ num_hidden_layers (`int`, *optional*, defaults to 32):
49
+ Number of hidden layers in the Transformer encoder.
50
+ num_attention_heads (`int`, *optional*, defaults to 32):
51
+ Number of attention heads for each attention layer in the Transformer encoder.
52
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
53
+ The non-linear activation function (function or string) in the decoder.
54
+ max_position_embeddings (`int`, *optional*, defaults to 2048):
55
+ The maximum sequence length that this model might ever be used with. Typically set this to something large
56
+ just in case (e.g., 512 or 1024 or 2048).
57
+ initializer_range (`float`, *optional*, defaults to 0.02):
58
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
59
+ rms_norm_eps (`float`, *optional*, defaults to 1e-12):
60
+ The epsilon used by the rms normalization layers.
61
+ use_cache (`bool`, *optional*, defaults to `True`):
62
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
63
+ relevant if `config.is_decoder=True`.
64
+ tie_word_embeddings(`bool`, *optional*, defaults to `False`):
65
+ Whether to tie weight embeddings
66
+ Example:
67
+
68
+ ```python
69
+ >>> from transformers import InternLMModel, InternLMConfig
70
+
71
+ >>> # Initializing a InternLM internlm-7b style configuration
72
+ >>> configuration = InternLMConfig()
73
+
74
+ >>> # Initializing a model from the internlm-7b style configuration
75
+ >>> model = InternLMModel(configuration)
76
+
77
+ >>> # Accessing the model configuration
78
+ >>> configuration = model.config
79
+ ```"""
80
+ model_type = "internlm"
81
+ _auto_class = "AutoConfig"
82
+
83
+ def __init__( # pylint: disable=W0102
84
+ self,
85
+ vocab_size=103168,
86
+ hidden_size=4096,
87
+ intermediate_size=11008,
88
+ num_hidden_layers=32,
89
+ num_attention_heads=32,
90
+ hidden_act="silu",
91
+ max_position_embeddings=2048,
92
+ initializer_range=0.02,
93
+ rms_norm_eps=1e-6,
94
+ use_cache=True,
95
+ pad_token_id=0,
96
+ bos_token_id=1,
97
+ eos_token_id=2,
98
+ tie_word_embeddings=False,
99
+ bias=True,
100
+ rotary={"base": 10000, "type": "dynamic"}, # pylint: disable=W0102
101
+ **kwargs,
102
+ ):
103
+ self.vocab_size = vocab_size
104
+ self.max_position_embeddings = max_position_embeddings
105
+ self.hidden_size = hidden_size
106
+ self.intermediate_size = intermediate_size
107
+ self.num_hidden_layers = num_hidden_layers
108
+ self.num_attention_heads = num_attention_heads
109
+ self.hidden_act = hidden_act
110
+ self.initializer_range = initializer_range
111
+ self.rms_norm_eps = rms_norm_eps
112
+ self.use_cache = use_cache
113
+ self.bias = bias
114
+ self.rotary = rotary
115
+ super().__init__(
116
+ pad_token_id=pad_token_id,
117
+ bos_token_id=bos_token_id,
118
+ eos_token_id=eos_token_id,
119
+ tie_word_embeddings=tie_word_embeddings,
120
+ **kwargs,
121
+ )
generation_config.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 1,
4
+ "eos_token_id": 2,
5
+ "pad_token_id": 2,
6
+ "transformers_version": "4.36.2"
7
+ }
model.safetensors.index.json ADDED
@@ -0,0 +1,458 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 14643904512
4
+ },
5
+ "weight_map": {
6
+ "lm_head.weight": "model-00003-of-00003.safetensors",
7
+ "model.embed_tokens.weight": "model-00001-of-00003.safetensors",
8
+ "model.layers.0.input_layernorm.weight": "model-00001-of-00003.safetensors",
9
+ "model.layers.0.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
10
+ "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
11
+ "model.layers.0.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
12
+ "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
13
+ "model.layers.0.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
14
+ "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
15
+ "model.layers.0.self_attn.o_proj.bias": "model-00001-of-00003.safetensors",
16
+ "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
17
+ "model.layers.0.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
18
+ "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
19
+ "model.layers.0.self_attn.rotary_emb.inv_freq": "model-00001-of-00003.safetensors",
20
+ "model.layers.0.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
21
+ "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
22
+ "model.layers.1.input_layernorm.weight": "model-00001-of-00003.safetensors",
23
+ "model.layers.1.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
24
+ "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
25
+ "model.layers.1.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
26
+ "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
27
+ "model.layers.1.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
28
+ "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
29
+ "model.layers.1.self_attn.o_proj.bias": "model-00001-of-00003.safetensors",
30
+ "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
31
+ "model.layers.1.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
32
+ "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
33
+ "model.layers.1.self_attn.rotary_emb.inv_freq": "model-00001-of-00003.safetensors",
34
+ "model.layers.1.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
35
+ "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
36
+ "model.layers.10.input_layernorm.weight": "model-00002-of-00003.safetensors",
37
+ "model.layers.10.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
38
+ "model.layers.10.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
39
+ "model.layers.10.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
40
+ "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
41
+ "model.layers.10.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
42
+ "model.layers.10.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
43
+ "model.layers.10.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
44
+ "model.layers.10.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
45
+ "model.layers.10.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
46
+ "model.layers.10.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
47
+ "model.layers.10.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
48
+ "model.layers.10.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
49
+ "model.layers.10.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
50
+ "model.layers.11.input_layernorm.weight": "model-00002-of-00003.safetensors",
51
+ "model.layers.11.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
52
+ "model.layers.11.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
53
+ "model.layers.11.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
54
+ "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
55
+ "model.layers.11.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
56
+ "model.layers.11.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
57
+ "model.layers.11.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
58
+ "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
59
+ "model.layers.11.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
60
+ "model.layers.11.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
61
+ "model.layers.11.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
62
+ "model.layers.11.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
63
+ "model.layers.11.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
64
+ "model.layers.12.input_layernorm.weight": "model-00002-of-00003.safetensors",
65
+ "model.layers.12.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
66
+ "model.layers.12.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
67
+ "model.layers.12.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
68
+ "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
69
+ "model.layers.12.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
70
+ "model.layers.12.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
71
+ "model.layers.12.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
72
+ "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
73
+ "model.layers.12.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
74
+ "model.layers.12.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
75
+ "model.layers.12.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
76
+ "model.layers.12.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
77
+ "model.layers.12.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
78
+ "model.layers.13.input_layernorm.weight": "model-00002-of-00003.safetensors",
79
+ "model.layers.13.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
80
+ "model.layers.13.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
81
+ "model.layers.13.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
82
+ "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
83
+ "model.layers.13.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
84
+ "model.layers.13.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
85
+ "model.layers.13.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
86
+ "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
87
+ "model.layers.13.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
88
+ "model.layers.13.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
89
+ "model.layers.13.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
90
+ "model.layers.13.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
91
+ "model.layers.13.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
92
+ "model.layers.14.input_layernorm.weight": "model-00002-of-00003.safetensors",
93
+ "model.layers.14.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
94
+ "model.layers.14.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
95
+ "model.layers.14.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
96
+ "model.layers.14.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
97
+ "model.layers.14.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
98
+ "model.layers.14.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
99
+ "model.layers.14.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
100
+ "model.layers.14.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
101
+ "model.layers.14.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
102
+ "model.layers.14.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
103
+ "model.layers.14.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
104
+ "model.layers.14.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
105
+ "model.layers.14.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
106
+ "model.layers.15.input_layernorm.weight": "model-00002-of-00003.safetensors",
107
+ "model.layers.15.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
108
+ "model.layers.15.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
109
+ "model.layers.15.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
110
+ "model.layers.15.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
111
+ "model.layers.15.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
112
+ "model.layers.15.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
113
+ "model.layers.15.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
114
+ "model.layers.15.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
115
+ "model.layers.15.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
116
+ "model.layers.15.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
117
+ "model.layers.15.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
118
+ "model.layers.15.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
119
+ "model.layers.15.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
120
+ "model.layers.16.input_layernorm.weight": "model-00002-of-00003.safetensors",
121
+ "model.layers.16.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
122
+ "model.layers.16.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
123
+ "model.layers.16.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
124
+ "model.layers.16.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
125
+ "model.layers.16.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
126
+ "model.layers.16.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
127
+ "model.layers.16.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
128
+ "model.layers.16.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
129
+ "model.layers.16.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
130
+ "model.layers.16.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
131
+ "model.layers.16.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
132
+ "model.layers.16.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
133
+ "model.layers.16.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
134
+ "model.layers.17.input_layernorm.weight": "model-00002-of-00003.safetensors",
135
+ "model.layers.17.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
136
+ "model.layers.17.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
137
+ "model.layers.17.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
138
+ "model.layers.17.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
139
+ "model.layers.17.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
140
+ "model.layers.17.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
141
+ "model.layers.17.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
142
+ "model.layers.17.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
143
+ "model.layers.17.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
144
+ "model.layers.17.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
145
+ "model.layers.17.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
146
+ "model.layers.17.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
147
+ "model.layers.17.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
148
+ "model.layers.18.input_layernorm.weight": "model-00002-of-00003.safetensors",
149
+ "model.layers.18.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
150
+ "model.layers.18.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
151
+ "model.layers.18.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
152
+ "model.layers.18.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
153
+ "model.layers.18.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
154
+ "model.layers.18.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
155
+ "model.layers.18.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
156
+ "model.layers.18.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
157
+ "model.layers.18.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
158
+ "model.layers.18.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
159
+ "model.layers.18.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
160
+ "model.layers.18.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
161
+ "model.layers.18.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
162
+ "model.layers.19.input_layernorm.weight": "model-00002-of-00003.safetensors",
163
+ "model.layers.19.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
164
+ "model.layers.19.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
165
+ "model.layers.19.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
166
+ "model.layers.19.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
167
+ "model.layers.19.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
168
+ "model.layers.19.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
169
+ "model.layers.19.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
170
+ "model.layers.19.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
171
+ "model.layers.19.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
172
+ "model.layers.19.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
173
+ "model.layers.19.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
174
+ "model.layers.19.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
175
+ "model.layers.19.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
176
+ "model.layers.2.input_layernorm.weight": "model-00001-of-00003.safetensors",
177
+ "model.layers.2.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
178
+ "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
179
+ "model.layers.2.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
180
+ "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
181
+ "model.layers.2.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
182
+ "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
183
+ "model.layers.2.self_attn.o_proj.bias": "model-00001-of-00003.safetensors",
184
+ "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
185
+ "model.layers.2.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
186
+ "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
187
+ "model.layers.2.self_attn.rotary_emb.inv_freq": "model-00001-of-00003.safetensors",
188
+ "model.layers.2.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
189
+ "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
190
+ "model.layers.20.input_layernorm.weight": "model-00002-of-00003.safetensors",
191
+ "model.layers.20.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
192
+ "model.layers.20.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
193
+ "model.layers.20.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
194
+ "model.layers.20.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
195
+ "model.layers.20.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
196
+ "model.layers.20.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
197
+ "model.layers.20.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
198
+ "model.layers.20.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
199
+ "model.layers.20.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
200
+ "model.layers.20.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
201
+ "model.layers.20.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
202
+ "model.layers.20.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
203
+ "model.layers.20.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
204
+ "model.layers.21.input_layernorm.weight": "model-00002-of-00003.safetensors",
205
+ "model.layers.21.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
206
+ "model.layers.21.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
207
+ "model.layers.21.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
208
+ "model.layers.21.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
209
+ "model.layers.21.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
210
+ "model.layers.21.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
211
+ "model.layers.21.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
212
+ "model.layers.21.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
213
+ "model.layers.21.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
214
+ "model.layers.21.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
215
+ "model.layers.21.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
216
+ "model.layers.21.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
217
+ "model.layers.21.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
218
+ "model.layers.22.input_layernorm.weight": "model-00003-of-00003.safetensors",
219
+ "model.layers.22.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
220
+ "model.layers.22.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
221
+ "model.layers.22.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
222
+ "model.layers.22.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
223
+ "model.layers.22.self_attn.k_proj.bias": "model-00002-of-00003.safetensors",
224
+ "model.layers.22.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
225
+ "model.layers.22.self_attn.o_proj.bias": "model-00002-of-00003.safetensors",
226
+ "model.layers.22.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
227
+ "model.layers.22.self_attn.q_proj.bias": "model-00002-of-00003.safetensors",
228
+ "model.layers.22.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
229
+ "model.layers.22.self_attn.rotary_emb.inv_freq": "model-00002-of-00003.safetensors",
230
+ "model.layers.22.self_attn.v_proj.bias": "model-00002-of-00003.safetensors",
231
+ "model.layers.22.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
232
+ "model.layers.23.input_layernorm.weight": "model-00003-of-00003.safetensors",
233
+ "model.layers.23.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
234
+ "model.layers.23.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
235
+ "model.layers.23.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
236
+ "model.layers.23.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
237
+ "model.layers.23.self_attn.k_proj.bias": "model-00003-of-00003.safetensors",
238
+ "model.layers.23.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
239
+ "model.layers.23.self_attn.o_proj.bias": "model-00003-of-00003.safetensors",
240
+ "model.layers.23.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
241
+ "model.layers.23.self_attn.q_proj.bias": "model-00003-of-00003.safetensors",
242
+ "model.layers.23.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
243
+ "model.layers.23.self_attn.rotary_emb.inv_freq": "model-00003-of-00003.safetensors",
244
+ "model.layers.23.self_attn.v_proj.bias": "model-00003-of-00003.safetensors",
245
+ "model.layers.23.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
246
+ "model.layers.24.input_layernorm.weight": "model-00003-of-00003.safetensors",
247
+ "model.layers.24.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
248
+ "model.layers.24.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
249
+ "model.layers.24.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
250
+ "model.layers.24.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
251
+ "model.layers.24.self_attn.k_proj.bias": "model-00003-of-00003.safetensors",
252
+ "model.layers.24.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
253
+ "model.layers.24.self_attn.o_proj.bias": "model-00003-of-00003.safetensors",
254
+ "model.layers.24.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
255
+ "model.layers.24.self_attn.q_proj.bias": "model-00003-of-00003.safetensors",
256
+ "model.layers.24.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
257
+ "model.layers.24.self_attn.rotary_emb.inv_freq": "model-00003-of-00003.safetensors",
258
+ "model.layers.24.self_attn.v_proj.bias": "model-00003-of-00003.safetensors",
259
+ "model.layers.24.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
260
+ "model.layers.25.input_layernorm.weight": "model-00003-of-00003.safetensors",
261
+ "model.layers.25.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
262
+ "model.layers.25.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
263
+ "model.layers.25.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
264
+ "model.layers.25.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
265
+ "model.layers.25.self_attn.k_proj.bias": "model-00003-of-00003.safetensors",
266
+ "model.layers.25.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
267
+ "model.layers.25.self_attn.o_proj.bias": "model-00003-of-00003.safetensors",
268
+ "model.layers.25.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
269
+ "model.layers.25.self_attn.q_proj.bias": "model-00003-of-00003.safetensors",
270
+ "model.layers.25.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
271
+ "model.layers.25.self_attn.rotary_emb.inv_freq": "model-00003-of-00003.safetensors",
272
+ "model.layers.25.self_attn.v_proj.bias": "model-00003-of-00003.safetensors",
273
+ "model.layers.25.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
274
+ "model.layers.26.input_layernorm.weight": "model-00003-of-00003.safetensors",
275
+ "model.layers.26.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
276
+ "model.layers.26.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
277
+ "model.layers.26.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
278
+ "model.layers.26.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
279
+ "model.layers.26.self_attn.k_proj.bias": "model-00003-of-00003.safetensors",
280
+ "model.layers.26.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
281
+ "model.layers.26.self_attn.o_proj.bias": "model-00003-of-00003.safetensors",
282
+ "model.layers.26.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
283
+ "model.layers.26.self_attn.q_proj.bias": "model-00003-of-00003.safetensors",
284
+ "model.layers.26.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
285
+ "model.layers.26.self_attn.rotary_emb.inv_freq": "model-00003-of-00003.safetensors",
286
+ "model.layers.26.self_attn.v_proj.bias": "model-00003-of-00003.safetensors",
287
+ "model.layers.26.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
288
+ "model.layers.27.input_layernorm.weight": "model-00003-of-00003.safetensors",
289
+ "model.layers.27.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
290
+ "model.layers.27.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
291
+ "model.layers.27.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
292
+ "model.layers.27.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
293
+ "model.layers.27.self_attn.k_proj.bias": "model-00003-of-00003.safetensors",
294
+ "model.layers.27.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
295
+ "model.layers.27.self_attn.o_proj.bias": "model-00003-of-00003.safetensors",
296
+ "model.layers.27.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
297
+ "model.layers.27.self_attn.q_proj.bias": "model-00003-of-00003.safetensors",
298
+ "model.layers.27.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
299
+ "model.layers.27.self_attn.rotary_emb.inv_freq": "model-00003-of-00003.safetensors",
300
+ "model.layers.27.self_attn.v_proj.bias": "model-00003-of-00003.safetensors",
301
+ "model.layers.27.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
302
+ "model.layers.28.input_layernorm.weight": "model-00003-of-00003.safetensors",
303
+ "model.layers.28.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
304
+ "model.layers.28.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
305
+ "model.layers.28.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
306
+ "model.layers.28.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
307
+ "model.layers.28.self_attn.k_proj.bias": "model-00003-of-00003.safetensors",
308
+ "model.layers.28.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
309
+ "model.layers.28.self_attn.o_proj.bias": "model-00003-of-00003.safetensors",
310
+ "model.layers.28.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
311
+ "model.layers.28.self_attn.q_proj.bias": "model-00003-of-00003.safetensors",
312
+ "model.layers.28.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
313
+ "model.layers.28.self_attn.rotary_emb.inv_freq": "model-00003-of-00003.safetensors",
314
+ "model.layers.28.self_attn.v_proj.bias": "model-00003-of-00003.safetensors",
315
+ "model.layers.28.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
316
+ "model.layers.29.input_layernorm.weight": "model-00003-of-00003.safetensors",
317
+ "model.layers.29.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
318
+ "model.layers.29.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
319
+ "model.layers.29.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
320
+ "model.layers.29.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
321
+ "model.layers.29.self_attn.k_proj.bias": "model-00003-of-00003.safetensors",
322
+ "model.layers.29.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
323
+ "model.layers.29.self_attn.o_proj.bias": "model-00003-of-00003.safetensors",
324
+ "model.layers.29.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
325
+ "model.layers.29.self_attn.q_proj.bias": "model-00003-of-00003.safetensors",
326
+ "model.layers.29.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
327
+ "model.layers.29.self_attn.rotary_emb.inv_freq": "model-00003-of-00003.safetensors",
328
+ "model.layers.29.self_attn.v_proj.bias": "model-00003-of-00003.safetensors",
329
+ "model.layers.29.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
330
+ "model.layers.3.input_layernorm.weight": "model-00001-of-00003.safetensors",
331
+ "model.layers.3.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
332
+ "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
333
+ "model.layers.3.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
334
+ "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
335
+ "model.layers.3.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
336
+ "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
337
+ "model.layers.3.self_attn.o_proj.bias": "model-00001-of-00003.safetensors",
338
+ "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
339
+ "model.layers.3.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
340
+ "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
341
+ "model.layers.3.self_attn.rotary_emb.inv_freq": "model-00001-of-00003.safetensors",
342
+ "model.layers.3.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
343
+ "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
344
+ "model.layers.30.input_layernorm.weight": "model-00003-of-00003.safetensors",
345
+ "model.layers.30.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
346
+ "model.layers.30.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
347
+ "model.layers.30.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
348
+ "model.layers.30.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
349
+ "model.layers.30.self_attn.k_proj.bias": "model-00003-of-00003.safetensors",
350
+ "model.layers.30.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
351
+ "model.layers.30.self_attn.o_proj.bias": "model-00003-of-00003.safetensors",
352
+ "model.layers.30.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
353
+ "model.layers.30.self_attn.q_proj.bias": "model-00003-of-00003.safetensors",
354
+ "model.layers.30.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
355
+ "model.layers.30.self_attn.rotary_emb.inv_freq": "model-00003-of-00003.safetensors",
356
+ "model.layers.30.self_attn.v_proj.bias": "model-00003-of-00003.safetensors",
357
+ "model.layers.30.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
358
+ "model.layers.31.input_layernorm.weight": "model-00003-of-00003.safetensors",
359
+ "model.layers.31.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
360
+ "model.layers.31.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
361
+ "model.layers.31.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
362
+ "model.layers.31.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
363
+ "model.layers.31.self_attn.k_proj.bias": "model-00003-of-00003.safetensors",
364
+ "model.layers.31.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
365
+ "model.layers.31.self_attn.o_proj.bias": "model-00003-of-00003.safetensors",
366
+ "model.layers.31.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
367
+ "model.layers.31.self_attn.q_proj.bias": "model-00003-of-00003.safetensors",
368
+ "model.layers.31.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
369
+ "model.layers.31.self_attn.rotary_emb.inv_freq": "model-00003-of-00003.safetensors",
370
+ "model.layers.31.self_attn.v_proj.bias": "model-00003-of-00003.safetensors",
371
+ "model.layers.31.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
372
+ "model.layers.4.input_layernorm.weight": "model-00001-of-00003.safetensors",
373
+ "model.layers.4.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
374
+ "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
375
+ "model.layers.4.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
376
+ "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
377
+ "model.layers.4.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
378
+ "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
379
+ "model.layers.4.self_attn.o_proj.bias": "model-00001-of-00003.safetensors",
380
+ "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
381
+ "model.layers.4.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
382
+ "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
383
+ "model.layers.4.self_attn.rotary_emb.inv_freq": "model-00001-of-00003.safetensors",
384
+ "model.layers.4.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
385
+ "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
386
+ "model.layers.5.input_layernorm.weight": "model-00001-of-00003.safetensors",
387
+ "model.layers.5.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
388
+ "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
389
+ "model.layers.5.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
390
+ "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
391
+ "model.layers.5.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
392
+ "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
393
+ "model.layers.5.self_attn.o_proj.bias": "model-00001-of-00003.safetensors",
394
+ "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
395
+ "model.layers.5.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
396
+ "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
397
+ "model.layers.5.self_attn.rotary_emb.inv_freq": "model-00001-of-00003.safetensors",
398
+ "model.layers.5.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
399
+ "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
400
+ "model.layers.6.input_layernorm.weight": "model-00001-of-00003.safetensors",
401
+ "model.layers.6.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
402
+ "model.layers.6.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
403
+ "model.layers.6.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
404
+ "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
405
+ "model.layers.6.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
406
+ "model.layers.6.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
407
+ "model.layers.6.self_attn.o_proj.bias": "model-00001-of-00003.safetensors",
408
+ "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
409
+ "model.layers.6.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
410
+ "model.layers.6.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
411
+ "model.layers.6.self_attn.rotary_emb.inv_freq": "model-00001-of-00003.safetensors",
412
+ "model.layers.6.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
413
+ "model.layers.6.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
414
+ "model.layers.7.input_layernorm.weight": "model-00001-of-00003.safetensors",
415
+ "model.layers.7.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
416
+ "model.layers.7.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
417
+ "model.layers.7.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
418
+ "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
419
+ "model.layers.7.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
420
+ "model.layers.7.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
421
+ "model.layers.7.self_attn.o_proj.bias": "model-00001-of-00003.safetensors",
422
+ "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
423
+ "model.layers.7.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
424
+ "model.layers.7.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
425
+ "model.layers.7.self_attn.rotary_emb.inv_freq": "model-00001-of-00003.safetensors",
426
+ "model.layers.7.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
427
+ "model.layers.7.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
428
+ "model.layers.8.input_layernorm.weight": "model-00001-of-00003.safetensors",
429
+ "model.layers.8.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
430
+ "model.layers.8.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
431
+ "model.layers.8.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
432
+ "model.layers.8.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
433
+ "model.layers.8.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
434
+ "model.layers.8.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
435
+ "model.layers.8.self_attn.o_proj.bias": "model-00001-of-00003.safetensors",
436
+ "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
437
+ "model.layers.8.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
438
+ "model.layers.8.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
439
+ "model.layers.8.self_attn.rotary_emb.inv_freq": "model-00001-of-00003.safetensors",
440
+ "model.layers.8.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
441
+ "model.layers.8.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
442
+ "model.layers.9.input_layernorm.weight": "model-00001-of-00003.safetensors",
443
+ "model.layers.9.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
444
+ "model.layers.9.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
445
+ "model.layers.9.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
446
+ "model.layers.9.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
447
+ "model.layers.9.self_attn.k_proj.bias": "model-00001-of-00003.safetensors",
448
+ "model.layers.9.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
449
+ "model.layers.9.self_attn.o_proj.bias": "model-00001-of-00003.safetensors",
450
+ "model.layers.9.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
451
+ "model.layers.9.self_attn.q_proj.bias": "model-00001-of-00003.safetensors",
452
+ "model.layers.9.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
453
+ "model.layers.9.self_attn.rotary_emb.inv_freq": "model-00001-of-00003.safetensors",
454
+ "model.layers.9.self_attn.v_proj.bias": "model-00001-of-00003.safetensors",
455
+ "model.layers.9.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
456
+ "model.norm.weight": "model-00003-of-00003.safetensors"
457
+ }
458
+ }
modeling_internlm.py ADDED
@@ -0,0 +1,1086 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+ """ PyTorch InternLM model."""
21
+ import math
22
+ import queue
23
+ import threading
24
+ from typing import List, Optional, Tuple, Union
25
+
26
+ import torch
27
+ import torch.utils.checkpoint
28
+ from torch import nn
29
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
30
+ from transformers.activations import ACT2FN
31
+ from transformers.generation.streamers import BaseStreamer
32
+ from transformers.modeling_outputs import (
33
+ BaseModelOutputWithPast,
34
+ CausalLMOutputWithPast,
35
+ SequenceClassifierOutputWithPast,
36
+ )
37
+ from transformers.modeling_utils import PreTrainedModel
38
+ from transformers.utils import (
39
+ add_start_docstrings,
40
+ add_start_docstrings_to_model_forward,
41
+ logging,
42
+ replace_return_docstrings,
43
+ )
44
+
45
+ from .configuration_internlm import InternLMConfig
46
+
47
+ logger = logging.get_logger(__name__)
48
+
49
+ _CONFIG_FOR_DOC = "InternLMConfig"
50
+
51
+
52
+ # Copied from transformers.models.bart.modeling_bart._make_causal_mask
53
+ def _make_causal_mask(
54
+ input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
55
+ ):
56
+ """
57
+ Make causal mask used for bi-directional self-attention.
58
+ """
59
+ bsz, tgt_len = input_ids_shape
60
+ mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
61
+ mask_cond = torch.arange(mask.size(-1), device=device)
62
+ mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
63
+ mask = mask.to(dtype)
64
+
65
+ if past_key_values_length > 0:
66
+ mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
67
+ return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
68
+
69
+
70
+ # Copied from transformers.models.bart.modeling_bart._expand_mask
71
+ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
72
+ """
73
+ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
74
+ """
75
+ bsz, src_len = mask.size()
76
+ tgt_len = tgt_len if tgt_len is not None else src_len
77
+
78
+ expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
79
+
80
+ inverted_mask = 1.0 - expanded_mask
81
+
82
+ return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
83
+
84
+
85
+ class InternLMRMSNorm(nn.Module):
86
+ """RMSNorm implemention."""
87
+
88
+ def __init__(self, hidden_size, eps=1e-6):
89
+ """
90
+ InternLMRMSNorm is equivalent to T5LayerNorm
91
+ """
92
+ super().__init__()
93
+ self.weight = nn.Parameter(torch.ones(hidden_size))
94
+ self.variance_epsilon = eps
95
+
96
+ def forward(self, hidden_states):
97
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
98
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
99
+
100
+ # convert into half-precision if necessary
101
+ if self.weight.dtype in [torch.float16, torch.bfloat16]:
102
+ hidden_states = hidden_states.to(self.weight.dtype)
103
+
104
+ return self.weight * hidden_states
105
+
106
+
107
+ class InternLMRotaryEmbedding(torch.nn.Module):
108
+ """Implement InternLM's rotary embedding.
109
+
110
+ Args:
111
+ dim (int): Characteristic dimension of each self-attentional head.
112
+ max_position_embeddings (int, optional): Model's training length. Defaults to 2048.
113
+ base (int, optional): The rotation position encodes the rotation Angle base number. Defaults to 10000.
114
+ device (Any, optional): Running device. Defaults to None.
115
+ """
116
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
117
+ super().__init__()
118
+ inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
119
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
120
+
121
+ # Build here to make `torch.jit.trace` work.
122
+ self.max_seq_len_cached = max_position_embeddings
123
+ t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
124
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
125
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
126
+ emb = torch.cat((freqs, freqs), dim=-1)
127
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
128
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
129
+
130
+ def forward(self, x, seq_len=None):
131
+ # x: [bs, num_attention_heads, seq_len, head_size]
132
+ # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
133
+ if seq_len > self.max_seq_len_cached:
134
+ self.max_seq_len_cached = seq_len
135
+ t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
136
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
137
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
138
+ emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
139
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
140
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
141
+ return (
142
+ self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
143
+ self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
144
+ )
145
+
146
+
147
+ class InternLMDynamicNTKScalingRotaryEmbedding(torch.nn.Module):
148
+ """Implement InternLM's DyanmicNTK extrapolation method, thereby broadening the model support context to 16K.
149
+
150
+ Args:
151
+ dim (int): Characteristic dimension of each self-attentional head.
152
+ max_position_embeddings (int, optional): Model's training length. Defaults to 2048.
153
+ base (int, optional): The rotation position encodes the rotation Angle base number. Defaults to 10000.
154
+ device (Any, optional): Running device. Defaults to None.
155
+ scaling_factor (float, optional): NTK method extrapolation coefficient. Defaults to 1.0.
156
+ """
157
+
158
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None, scaling_factor=1.0):
159
+ super().__init__()
160
+ inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
161
+ self.register_buffer("inv_freq", inv_freq)
162
+ self.dim = dim
163
+ self.base = base
164
+ self.scaling_factor = scaling_factor
165
+
166
+ # Build here to make `torch.jit.trace` work.
167
+ self.max_position_embeddings = max_position_embeddings
168
+ self.max_seq_len_cached = max_position_embeddings
169
+ t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
170
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
171
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
172
+ emb = torch.cat((freqs, freqs), dim=-1)
173
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
174
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
175
+
176
+ def _update_cached(self, x, seq_len=None):
177
+ self.max_seq_len_cached = max(seq_len, self.max_position_embeddings)
178
+ if seq_len > self.max_position_embeddings:
179
+ base = self.base * (
180
+ (self.scaling_factor * seq_len / self.max_position_embeddings) - (self.scaling_factor - 1)
181
+ ) ** (self.dim / (self.dim - 2))
182
+ inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2).float().to(x.device) / self.dim))
183
+ else:
184
+ inv_freq = self.inv_freq
185
+ t = torch.arange(self.max_seq_len_cached, device=inv_freq.device, dtype=inv_freq.dtype)
186
+ freqs = torch.einsum("i,j->ij", t, inv_freq)
187
+ emb = torch.cat((freqs, freqs), dim=-1)
188
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
189
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
190
+
191
+ def forward(self, x, seq_len=None):
192
+ # x: [bs, num_attention_heads, seq_len, head_size]
193
+ # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
194
+ if seq_len <= self.max_position_embeddings:
195
+ # Reset the tables if the sequence length has changed,
196
+ if self.max_seq_len_cached > self.max_position_embeddings:
197
+ self._update_cached(x, seq_len)
198
+ else:
199
+ self._update_cached(x, seq_len)
200
+
201
+ return (
202
+ self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
203
+ self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
204
+ )
205
+
206
+
207
+ def rotate_half(x):
208
+ """Rotates half the hidden dims of the input."""
209
+ x1 = x[..., : x.shape[-1] // 2]
210
+ x2 = x[..., x.shape[-1] // 2 :]
211
+ return torch.cat((-x2, x1), dim=-1)
212
+
213
+
214
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
215
+ # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
216
+ cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
217
+ sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
218
+ cos = cos.unsqueeze(0).unsqueeze(0).expand(len(position_ids), -1, -1, -1)
219
+ sin = sin.unsqueeze(0).unsqueeze(0).expand(len(position_ids), -1, -1, -1)
220
+ if q.size(2) == 1:
221
+ q_embed = (q * cos[:, :, -1, :]) + (rotate_half(q) * sin[:, :, -1, :])
222
+ else:
223
+ q_embed = (q * cos) + (rotate_half(q) * sin)
224
+
225
+ if k.size(2) == 1:
226
+ k_embed = (k * cos[:, :, -1, :]) + (rotate_half(k) * sin[:, :, -1, :])
227
+ else:
228
+ k_embed = (k * cos) + (rotate_half(k) * sin)
229
+
230
+ return q_embed, k_embed
231
+
232
+
233
+ class InternLMMLP(nn.Module):
234
+ def __init__(
235
+ self,
236
+ hidden_size: int,
237
+ intermediate_size: int,
238
+ hidden_act: str,
239
+ ):
240
+ super().__init__()
241
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
242
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
243
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
244
+ self.act_fn = ACT2FN[hidden_act]
245
+
246
+ def forward(self, x):
247
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
248
+
249
+
250
+ class InternLMAttention(nn.Module):
251
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
252
+
253
+ def __init__(self, config: InternLMConfig):
254
+ super().__init__()
255
+ self.config = config
256
+ self.hidden_size = config.hidden_size
257
+ self.num_heads = config.num_attention_heads
258
+ self.head_dim = self.hidden_size // self.num_heads
259
+ self.max_position_embeddings = config.max_position_embeddings
260
+
261
+ if (self.head_dim * self.num_heads) != self.hidden_size:
262
+ raise ValueError(
263
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
264
+ f" and `num_heads`: {self.num_heads})."
265
+ )
266
+ self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.bias)
267
+ self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.bias)
268
+ self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.bias)
269
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.bias)
270
+ self.rotary_emb = self._init_rope()
271
+
272
+ def _init_rope(self):
273
+ if self.config.rotary["type"] == "origin":
274
+ self.rotary_emb = InternLMRotaryEmbedding(
275
+ self.head_dim,
276
+ max_position_embeddings=self.max_position_embeddings,
277
+ base=self.config.rotary["base"],
278
+ )
279
+ elif self.config.rotary["type"] == "dynamic":
280
+ self.rotary_emb = InternLMDynamicNTKScalingRotaryEmbedding(
281
+ self.head_dim,
282
+ max_position_embeddings=self.max_position_embeddings,
283
+ base=self.config.rotary["base"],
284
+ scaling_factor=self.config.rotary.get("scaling_factor", 1.0),
285
+ )
286
+ else:
287
+ raise ValueError("Currently we only support rotary embedding's type being one of ('origin', 'dynamic').")
288
+ return self.rotary_emb
289
+
290
+ def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
291
+ return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
292
+
293
+ def forward(
294
+ self,
295
+ hidden_states: torch.Tensor,
296
+ attention_mask: Optional[torch.Tensor] = None,
297
+ position_ids: Optional[torch.LongTensor] = None,
298
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
299
+ output_attentions: bool = False,
300
+ use_cache: bool = False,
301
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
302
+ bsz, q_len, _ = hidden_states.size()
303
+
304
+ query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
305
+ key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
306
+ value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
307
+
308
+ if past_key_value is not None:
309
+ # reuse k, v, self_attention
310
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
311
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
312
+
313
+ # print(use_cache)
314
+ past_key_value = (key_states, value_states) if use_cache else None
315
+
316
+ kv_seq_len = key_states.shape[-2]
317
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
318
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
319
+
320
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
321
+
322
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
323
+ raise ValueError(
324
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
325
+ f" {attn_weights.size()}"
326
+ )
327
+
328
+ if attention_mask is not None:
329
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
330
+ raise ValueError(
331
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
332
+ )
333
+ attn_weights = attn_weights + attention_mask
334
+ attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
335
+
336
+ # upcast attention to fp32
337
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
338
+ attn_output = torch.matmul(attn_weights, value_states)
339
+
340
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
341
+ raise ValueError(
342
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
343
+ f" {attn_output.size()}"
344
+ )
345
+
346
+ attn_output = attn_output.transpose(1, 2)
347
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
348
+
349
+ attn_output = self.o_proj(attn_output)
350
+
351
+ if not output_attentions:
352
+ attn_weights = None
353
+
354
+ return attn_output, attn_weights, past_key_value
355
+
356
+
357
+ class InternLMDecoderLayer(nn.Module):
358
+ def __init__(self, config: InternLMConfig):
359
+ super().__init__()
360
+ self.hidden_size = config.hidden_size
361
+ self.self_attn = InternLMAttention(config=config)
362
+ self.mlp = InternLMMLP(
363
+ hidden_size=self.hidden_size,
364
+ intermediate_size=config.intermediate_size,
365
+ hidden_act=config.hidden_act,
366
+ )
367
+ self.input_layernorm = InternLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
368
+ self.post_attention_layernorm = InternLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
369
+
370
+ def forward(
371
+ self,
372
+ hidden_states: torch.Tensor,
373
+ attention_mask: Optional[torch.Tensor] = None,
374
+ position_ids: Optional[torch.LongTensor] = None,
375
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
376
+ output_attentions: Optional[bool] = False,
377
+ use_cache: Optional[bool] = False,
378
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
379
+ """
380
+ Args:
381
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
382
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
383
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
384
+ output_attentions (`bool`, *optional*):
385
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
386
+ returned tensors for more detail.
387
+ use_cache (`bool`, *optional*):
388
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
389
+ (see `past_key_values`).
390
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
391
+ """
392
+
393
+ residual = hidden_states
394
+
395
+ hidden_states = self.input_layernorm(hidden_states)
396
+
397
+ # Self Attention
398
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
399
+ hidden_states=hidden_states,
400
+ attention_mask=attention_mask,
401
+ position_ids=position_ids,
402
+ past_key_value=past_key_value,
403
+ output_attentions=output_attentions,
404
+ use_cache=use_cache,
405
+ )
406
+ hidden_states = residual + hidden_states
407
+
408
+ # Fully Connected
409
+ residual = hidden_states
410
+ hidden_states = self.post_attention_layernorm(hidden_states)
411
+ hidden_states = self.mlp(hidden_states)
412
+ hidden_states = residual + hidden_states
413
+
414
+ outputs = (hidden_states,)
415
+
416
+ if output_attentions:
417
+ outputs += (self_attn_weights,)
418
+
419
+ if use_cache:
420
+ outputs += (present_key_value,)
421
+
422
+ return outputs
423
+
424
+
425
+ INTERNLM_START_DOCSTRING = r"""
426
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
427
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
428
+ etc.)
429
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
430
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
431
+ and behavior.
432
+ Parameters:
433
+ config ([`InternLMConfig`]):
434
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
435
+ load the weights associated with the model, only the configuration. Check out the
436
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
437
+ """
438
+
439
+
440
+ @add_start_docstrings(
441
+ "The bare InternLM Model outputting raw hidden-states without any specific head on top.",
442
+ INTERNLM_START_DOCSTRING,
443
+ )
444
+ class InternLMPreTrainedModel(PreTrainedModel):
445
+ config_class = InternLMConfig
446
+ base_model_prefix = "model"
447
+ supports_gradient_checkpointing = True
448
+ _no_split_modules = ["InternLMDecoderLayer"]
449
+ _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
450
+
451
+ def _init_weights(self, module):
452
+ std = self.config.initializer_range
453
+ if isinstance(module, nn.Linear):
454
+ module.weight.data.normal_(mean=0.0, std=std)
455
+ if module.bias is not None:
456
+ module.bias.data.zero_()
457
+ elif isinstance(module, nn.Embedding):
458
+ module.weight.data.normal_(mean=0.0, std=std)
459
+ if module.padding_idx is not None:
460
+ module.weight.data[module.padding_idx].zero_()
461
+
462
+ def _set_gradient_checkpointing(self, module, value=False):
463
+ if isinstance(module, InternLMModel):
464
+ module.gradient_checkpointing = value
465
+
466
+
467
+ INTERNLM_INPUTS_DOCSTRING = r"""
468
+ Args:
469
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
470
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
471
+ it.
472
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
473
+ [`PreTrainedTokenizer.__call__`] for details.
474
+ [What are input IDs?](../glossary#input-ids)
475
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
476
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
477
+ - 1 for tokens that are **not masked**,
478
+ - 0 for tokens that are **masked**.
479
+ [What are attention masks?](../glossary#attention-mask)
480
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
481
+ [`PreTrainedTokenizer.__call__`] for details.
482
+ If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
483
+ `past_key_values`).
484
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
485
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
486
+ information on the default strategy.
487
+ - 1 indicates the head is **not masked**,
488
+ - 0 indicates the head is **masked**.
489
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
490
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
491
+ config.n_positions - 1]`.
492
+ [What are position IDs?](../glossary#position-ids)
493
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or
494
+ when `config.use_cache=True`):
495
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
496
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
497
+ `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
498
+ Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
499
+ blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
500
+ If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
501
+ don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
502
+ `decoder_input_ids` of shape `(batch_size, sequence_length)`.
503
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
504
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
505
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
506
+ model's internal embedding lookup matrix.
507
+ use_cache (`bool`, *optional*):
508
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
509
+ `past_key_values`).
510
+ output_attentions (`bool`, *optional*):
511
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
512
+ tensors for more detail.
513
+ output_hidden_states (`bool`, *optional*):
514
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
515
+ more detail.
516
+ return_dict (`bool`, *optional*):
517
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
518
+ """
519
+
520
+
521
+ @add_start_docstrings(
522
+ "The bare InternLM Model outputting raw hidden-states without any specific head on top.",
523
+ INTERNLM_START_DOCSTRING,
524
+ )
525
+ class InternLMModel(InternLMPreTrainedModel):
526
+ """
527
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`InternLMDecoderLayer`]
528
+ Args:
529
+ config: InternLMConfig
530
+ """
531
+
532
+ _auto_class = "AutoModel"
533
+
534
+ def __init__(self, config: InternLMConfig):
535
+ super().__init__(config)
536
+ self.padding_idx = config.pad_token_id
537
+ self.vocab_size = config.vocab_size
538
+
539
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
540
+ self.layers = nn.ModuleList([InternLMDecoderLayer(config) for _ in range(config.num_hidden_layers)])
541
+ self.norm = InternLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
542
+
543
+ self.gradient_checkpointing = False
544
+ # Initialize weights and apply final processing
545
+ self.post_init()
546
+
547
+ def get_input_embeddings(self):
548
+ return self.embed_tokens
549
+
550
+ def set_input_embeddings(self, value):
551
+ self.embed_tokens = value
552
+
553
+ # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
554
+ def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
555
+ # create causal mask
556
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
557
+ combined_attention_mask = None
558
+ if input_shape[-1] > 1:
559
+ combined_attention_mask = _make_causal_mask(
560
+ input_shape,
561
+ inputs_embeds.dtype,
562
+ device=inputs_embeds.device,
563
+ past_key_values_length=past_key_values_length,
564
+ )
565
+
566
+ if attention_mask is not None:
567
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
568
+ expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
569
+ inputs_embeds.device
570
+ )
571
+ combined_attention_mask = (
572
+ expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
573
+ )
574
+
575
+ return combined_attention_mask
576
+
577
+ @add_start_docstrings_to_model_forward(INTERNLM_INPUTS_DOCSTRING)
578
+ def forward(
579
+ self,
580
+ input_ids: torch.LongTensor = None,
581
+ attention_mask: Optional[torch.Tensor] = None,
582
+ position_ids: Optional[torch.LongTensor] = None,
583
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
584
+ inputs_embeds: Optional[torch.FloatTensor] = None,
585
+ use_cache: Optional[bool] = None,
586
+ output_attentions: Optional[bool] = None,
587
+ output_hidden_states: Optional[bool] = None,
588
+ return_dict: Optional[bool] = None,
589
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
590
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
591
+ output_hidden_states = (
592
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
593
+ )
594
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
595
+
596
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
597
+
598
+ # retrieve input_ids and inputs_embeds
599
+ if input_ids is not None and inputs_embeds is not None:
600
+ raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
601
+ elif input_ids is not None:
602
+ batch_size, seq_length = input_ids.shape
603
+ elif inputs_embeds is not None:
604
+ batch_size, seq_length, _ = inputs_embeds.shape
605
+ else:
606
+ raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
607
+
608
+ seq_length_with_past = seq_length
609
+ past_key_values_length = 0
610
+
611
+ if past_key_values is not None:
612
+ past_key_values_length = past_key_values[0][0].shape[2]
613
+ seq_length_with_past = seq_length_with_past + past_key_values_length
614
+
615
+ if position_ids is None:
616
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
617
+ position_ids = torch.arange(
618
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
619
+ )
620
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
621
+ else:
622
+ position_ids = position_ids.view(-1, seq_length).long()
623
+
624
+ if inputs_embeds is None:
625
+ inputs_embeds = self.embed_tokens(input_ids)
626
+ # embed positions
627
+ if attention_mask is None:
628
+ attention_mask = torch.ones(
629
+ (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
630
+ )
631
+ attention_mask = self._prepare_decoder_attention_mask(
632
+ attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
633
+ )
634
+
635
+ hidden_states = inputs_embeds
636
+
637
+ if self.gradient_checkpointing and self.training:
638
+ if use_cache:
639
+ logger.warning_once(
640
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
641
+ )
642
+ use_cache = False
643
+
644
+ # decoder layers
645
+ all_hidden_states = () if output_hidden_states else None
646
+ all_self_attns = () if output_attentions else None
647
+ next_decoder_cache = () if use_cache else None
648
+
649
+ for idx, decoder_layer in enumerate(self.layers):
650
+ if output_hidden_states:
651
+ all_hidden_states += (hidden_states,)
652
+
653
+ past_key_value = past_key_values[idx] if past_key_values is not None else None
654
+
655
+ if self.gradient_checkpointing and self.training:
656
+
657
+ def create_custom_forward(module):
658
+ def custom_forward(*inputs):
659
+ # None for past_key_value
660
+ return module(*inputs, output_attentions, None)
661
+
662
+ return custom_forward
663
+
664
+ layer_outputs = torch.utils.checkpoint.checkpoint(
665
+ create_custom_forward(decoder_layer),
666
+ hidden_states,
667
+ attention_mask,
668
+ position_ids,
669
+ None,
670
+ )
671
+ else:
672
+ layer_outputs = decoder_layer(
673
+ hidden_states,
674
+ attention_mask=attention_mask,
675
+ position_ids=position_ids,
676
+ past_key_value=past_key_value,
677
+ output_attentions=output_attentions,
678
+ use_cache=use_cache,
679
+ )
680
+
681
+ hidden_states = layer_outputs[0]
682
+
683
+ if use_cache:
684
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
685
+
686
+ if output_attentions:
687
+ all_self_attns += (layer_outputs[1],)
688
+
689
+ hidden_states = self.norm(hidden_states)
690
+
691
+ # add hidden states from the last decoder layer
692
+ if output_hidden_states:
693
+ all_hidden_states += (hidden_states,)
694
+
695
+ next_cache = next_decoder_cache if use_cache else None
696
+ if not return_dict:
697
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
698
+ return BaseModelOutputWithPast(
699
+ last_hidden_state=hidden_states,
700
+ past_key_values=next_cache,
701
+ hidden_states=all_hidden_states,
702
+ attentions=all_self_attns,
703
+ )
704
+
705
+
706
+ class InternLMForCausalLM(InternLMPreTrainedModel):
707
+ _auto_class = "AutoModelForCausalLM"
708
+
709
+ def __init__(self, config):
710
+ super().__init__(config)
711
+ self.model = InternLMModel(config)
712
+
713
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
714
+
715
+ # Initialize weights and apply final processing
716
+ self.post_init()
717
+
718
+ def get_input_embeddings(self):
719
+ return self.model.embed_tokens
720
+
721
+ def set_input_embeddings(self, value):
722
+ self.model.embed_tokens = value
723
+
724
+ def get_output_embeddings(self):
725
+ return self.lm_head
726
+
727
+ def set_output_embeddings(self, new_embeddings):
728
+ self.lm_head = new_embeddings
729
+
730
+ def set_decoder(self, decoder):
731
+ self.model = decoder
732
+
733
+ def get_decoder(self):
734
+ return self.model
735
+
736
+ @add_start_docstrings_to_model_forward(INTERNLM_INPUTS_DOCSTRING)
737
+ @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
738
+ def forward(
739
+ self,
740
+ input_ids: torch.LongTensor = None,
741
+ attention_mask: Optional[torch.Tensor] = None,
742
+ position_ids: Optional[torch.LongTensor] = None,
743
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
744
+ inputs_embeds: Optional[torch.FloatTensor] = None,
745
+ labels: Optional[torch.LongTensor] = None,
746
+ use_cache: Optional[bool] = None,
747
+ output_attentions: Optional[bool] = None,
748
+ output_hidden_states: Optional[bool] = None,
749
+ return_dict: Optional[bool] = None,
750
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
751
+ r"""
752
+ Args:
753
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
754
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
755
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
756
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
757
+ Returns:
758
+ Example:
759
+ ```python
760
+ >>> from transformers import AutoTokenizer, InternLMForCausalLM
761
+ >>> model = InternLMForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
762
+ >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
763
+ >>> prompt = "Hey, are you consciours? Can you talk to me?"
764
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
765
+ >>> # Generate
766
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
767
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
768
+ "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
769
+ ```"""
770
+
771
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
772
+ output_hidden_states = (
773
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
774
+ )
775
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
776
+
777
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
778
+ outputs = self.model(
779
+ input_ids=input_ids,
780
+ attention_mask=attention_mask,
781
+ position_ids=position_ids,
782
+ past_key_values=past_key_values,
783
+ inputs_embeds=inputs_embeds,
784
+ use_cache=use_cache,
785
+ output_attentions=output_attentions,
786
+ output_hidden_states=output_hidden_states,
787
+ return_dict=return_dict,
788
+ )
789
+
790
+ hidden_states = outputs[0]
791
+ logits = self.lm_head(hidden_states)
792
+
793
+ loss = None
794
+ if labels is not None:
795
+ # Shift so that tokens < n predict n
796
+ shift_logits = logits[..., :-1, :].contiguous()
797
+ shift_labels = labels[..., 1:].contiguous()
798
+ # Flatten the tokens
799
+ loss_fct = CrossEntropyLoss()
800
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
801
+ shift_labels = shift_labels.view(-1)
802
+ # Enable model parallelism
803
+ shift_labels = shift_labels.to(shift_logits.device)
804
+ loss = loss_fct(shift_logits, shift_labels)
805
+
806
+ if not return_dict:
807
+ output = (logits,) + outputs[1:]
808
+ return (loss,) + output if loss is not None else output
809
+
810
+ return CausalLMOutputWithPast(
811
+ loss=loss,
812
+ logits=logits,
813
+ past_key_values=outputs.past_key_values,
814
+ hidden_states=outputs.hidden_states,
815
+ attentions=outputs.attentions,
816
+ )
817
+
818
+ def prepare_inputs_for_generation(
819
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
820
+ ):
821
+ if past_key_values:
822
+ input_ids = input_ids[:, -1:]
823
+
824
+ position_ids = kwargs.get("position_ids", None)
825
+ if attention_mask is not None and position_ids is None:
826
+ # create position_ids on the fly for batch generation
827
+ position_ids = attention_mask.long().cumsum(-1) - 1
828
+ position_ids.masked_fill_(attention_mask == 0, 1)
829
+ if past_key_values:
830
+ position_ids = position_ids[:, -1].unsqueeze(-1)
831
+
832
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
833
+ if inputs_embeds is not None and past_key_values is None:
834
+ model_inputs = {"inputs_embeds": inputs_embeds}
835
+ else:
836
+ model_inputs = {"input_ids": input_ids}
837
+
838
+ model_inputs.update(
839
+ {
840
+ "position_ids": position_ids,
841
+ "past_key_values": past_key_values,
842
+ "use_cache": kwargs.get("use_cache"),
843
+ "attention_mask": attention_mask,
844
+ }
845
+ )
846
+ return model_inputs
847
+
848
+ @staticmethod
849
+ def _reorder_cache(past_key_values, beam_idx):
850
+ reordered_past = ()
851
+ for layer_past in past_key_values:
852
+ reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
853
+ return reordered_past
854
+
855
+ def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = []):
856
+ prompt = ""
857
+ for record in history:
858
+ prompt += f"""<|User|>:{record[0]}<eoh>\n<|Bot|>:{record[1]}<eoa>\n"""
859
+ prompt += f"""<|User|>:{query}<eoh>\n<|Bot|>:"""
860
+ return tokenizer([prompt], return_tensors="pt")
861
+
862
+ @torch.no_grad()
863
+ def chat(
864
+ self,
865
+ tokenizer,
866
+ query: str,
867
+ history: List[Tuple[str, str]] = [],
868
+ streamer: Optional[BaseStreamer] = None,
869
+ max_new_tokens: int = 1024,
870
+ do_sample: bool = True,
871
+ temperature: float = 0.8,
872
+ top_p: float = 0.8,
873
+ **kwargs,
874
+ ):
875
+ inputs = self.build_inputs(tokenizer, query, history)
876
+ inputs = {k: v.to(self.device) for k, v in inputs.items() if torch.is_tensor(v)}
877
+ outputs = self.generate(
878
+ **inputs,
879
+ streamer=streamer,
880
+ max_new_tokens=max_new_tokens,
881
+ do_sample=do_sample,
882
+ temperature=temperature,
883
+ top_p=top_p,
884
+ **kwargs,
885
+ )
886
+ outputs = outputs[0].cpu().tolist()[len(inputs["input_ids"][0]) :]
887
+ response = tokenizer.decode(outputs, skip_special_tokens=True)
888
+ response = response.split("<eoa>")[0]
889
+ history = history + [(query, response)]
890
+ return response, history
891
+
892
+ @torch.no_grad()
893
+ def stream_chat(
894
+ self,
895
+ tokenizer,
896
+ query: str,
897
+ history: List[Tuple[str, str]] = [],
898
+ max_new_tokens: int = 1024,
899
+ do_sample: bool = True,
900
+ temperature: float = 0.8,
901
+ top_p: float = 0.8,
902
+ **kwargs,
903
+ ):
904
+ """
905
+ Return a generator in format: (response, history)
906
+ Eg.
907
+ ('你好,有什么可以帮助您的吗', [('你好', '你好,有什么可以帮助您的吗')])
908
+ ('你好,有什么可以帮助您的吗?', [('你好', '你好,有什么可以帮助您的吗?')])
909
+ """
910
+
911
+ response_queue = queue.Queue(maxsize=20)
912
+
913
+ class ChatStreamer(BaseStreamer):
914
+ def __init__(self, tokenizer) -> None:
915
+ super().__init__()
916
+ self.tokenizer = tokenizer
917
+ self.queue = response_queue
918
+ self.query = query
919
+ self.history = history
920
+ self.response = ""
921
+ self.received_inputs = False
922
+ self.queue.put((self.response, history + [(self.query, self.response)]))
923
+
924
+ def put(self, value):
925
+ if len(value.shape) > 1 and value.shape[0] > 1:
926
+ raise ValueError("ChatStreamer only supports batch size 1")
927
+ elif len(value.shape) > 1:
928
+ value = value[0]
929
+
930
+ if not self.received_inputs:
931
+ # The first received value is input_ids, ignore here
932
+ self.received_inputs = True
933
+ return
934
+
935
+ token = self.tokenizer.decode([value[-1]], skip_special_tokens=True)
936
+ if token.strip() != "<eoa>":
937
+ self.response = self.response + token
938
+ history = self.history + [(self.query, self.response)]
939
+ self.queue.put((self.response, history))
940
+
941
+ def end(self):
942
+ self.queue.put(None)
943
+
944
+ def stream_producer():
945
+ return self.chat(
946
+ tokenizer=tokenizer,
947
+ query=query,
948
+ streamer=ChatStreamer(tokenizer=tokenizer),
949
+ history=history,
950
+ max_new_tokens=max_new_tokens,
951
+ do_sample=do_sample,
952
+ temperature=temperature,
953
+ top_p=top_p,
954
+ **kwargs,
955
+ )
956
+
957
+ def consumer():
958
+ producer = threading.Thread(target=stream_producer)
959
+ producer.start()
960
+ while True:
961
+ res = response_queue.get()
962
+ if res is None:
963
+ return
964
+ yield res
965
+
966
+ return consumer()
967
+
968
+
969
+ @add_start_docstrings(
970
+ """
971
+ The InternLM Model transformer with a sequence classification head on top (linear layer).
972
+ [`InternLMForSequenceClassification`] uses the last token in order to do the classification, as other causal models
973
+ (e.g. GPT-2) do.
974
+ Since it does classification on the last token, it requires to know the position of the last token. If a
975
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
976
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
977
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
978
+ each row of the batch).
979
+ """,
980
+ INTERNLM_START_DOCSTRING,
981
+ )
982
+ class InternLMForSequenceClassification(InternLMPreTrainedModel):
983
+ _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
984
+
985
+ def __init__(self, config):
986
+ super().__init__(config)
987
+ self.num_labels = config.num_labels
988
+ self.model = InternLMModel(config)
989
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
990
+
991
+ # Initialize weights and apply final processing
992
+ self.post_init()
993
+
994
+ def get_input_embeddings(self):
995
+ return self.model.embed_tokens
996
+
997
+ def set_input_embeddings(self, value):
998
+ self.model.embed_tokens = value
999
+
1000
+ @add_start_docstrings_to_model_forward(INTERNLM_INPUTS_DOCSTRING)
1001
+ def forward(
1002
+ self,
1003
+ input_ids: torch.LongTensor = None,
1004
+ attention_mask: Optional[torch.Tensor] = None,
1005
+ position_ids: Optional[torch.LongTensor] = None,
1006
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1007
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1008
+ labels: Optional[torch.LongTensor] = None,
1009
+ use_cache: Optional[bool] = None,
1010
+ output_attentions: Optional[bool] = None,
1011
+ output_hidden_states: Optional[bool] = None,
1012
+ return_dict: Optional[bool] = None,
1013
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1014
+ r"""
1015
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1016
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1017
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1018
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1019
+ """
1020
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1021
+
1022
+ transformer_outputs = self.model(
1023
+ input_ids,
1024
+ attention_mask=attention_mask,
1025
+ position_ids=position_ids,
1026
+ past_key_values=past_key_values,
1027
+ inputs_embeds=inputs_embeds,
1028
+ use_cache=use_cache,
1029
+ output_attentions=output_attentions,
1030
+ output_hidden_states=output_hidden_states,
1031
+ return_dict=return_dict,
1032
+ )
1033
+ hidden_states = transformer_outputs[0]
1034
+ logits = self.score(hidden_states)
1035
+
1036
+ if input_ids is not None:
1037
+ batch_size = input_ids.shape[0]
1038
+ else:
1039
+ batch_size = inputs_embeds.shape[0]
1040
+
1041
+ if self.config.pad_token_id is None and batch_size != 1:
1042
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
1043
+ if self.config.pad_token_id is None:
1044
+ sequence_lengths = -1
1045
+ else:
1046
+ if input_ids is not None:
1047
+ sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
1048
+ else:
1049
+ sequence_lengths = -1
1050
+
1051
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1052
+
1053
+ loss = None
1054
+ if labels is not None:
1055
+ labels = labels.to(logits.device)
1056
+ if self.config.problem_type is None:
1057
+ if self.num_labels == 1:
1058
+ self.config.problem_type = "regression"
1059
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1060
+ self.config.problem_type = "single_label_classification"
1061
+ else:
1062
+ self.config.problem_type = "multi_label_classification"
1063
+
1064
+ if self.config.problem_type == "regression":
1065
+ loss_fct = MSELoss()
1066
+ if self.num_labels == 1:
1067
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1068
+ else:
1069
+ loss = loss_fct(pooled_logits, labels)
1070
+ elif self.config.problem_type == "single_label_classification":
1071
+ loss_fct = CrossEntropyLoss()
1072
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1073
+ elif self.config.problem_type == "multi_label_classification":
1074
+ loss_fct = BCEWithLogitsLoss()
1075
+ loss = loss_fct(pooled_logits, labels)
1076
+ if not return_dict:
1077
+ output = (pooled_logits,) + transformer_outputs[1:]
1078
+ return ((loss,) + output) if loss is not None else output
1079
+
1080
+ return SequenceClassifierOutputWithPast(
1081
+ loss=loss,
1082
+ logits=pooled_logits,
1083
+ past_key_values=transformer_outputs.past_key_values,
1084
+ hidden_states=transformer_outputs.hidden_states,
1085
+ attentions=transformer_outputs.attentions,
1086
+ )
special_tokens_map.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "</s>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": {
17
+ "content": "</s>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ },
23
+ "unk_token": {
24
+ "content": "<unk>",
25
+ "lstrip": false,
26
+ "normalized": false,
27
+ "rstrip": false,
28
+ "single_word": false
29
+ }
30
+ }
tokenization_internlm.py ADDED
@@ -0,0 +1,242 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+
21
+ """Tokenization classes for IntermLM."""
22
+ import os
23
+ from shutil import copyfile
24
+ from typing import Any, Dict, List, Optional, Tuple
25
+
26
+ import sentencepiece as spm
27
+
28
+ from transformers.tokenization_utils import PreTrainedTokenizer
29
+ from transformers.utils import logging
30
+
31
+
32
+ logger = logging.get_logger(__name__)
33
+
34
+ VOCAB_FILES_NAMES = {"vocab_file": "./tokenizer.model"}
35
+
36
+ PRETRAINED_VOCAB_FILES_MAP = {}
37
+
38
+
39
+ class InternLMTokenizer(PreTrainedTokenizer):
40
+ """
41
+ Construct a InternLM tokenizer. Based on byte-level Byte-Pair-Encoding.
42
+
43
+ Args:
44
+ vocab_file (`str`):
45
+ Path to the vocabulary file.
46
+ """
47
+
48
+ vocab_files_names = VOCAB_FILES_NAMES
49
+ pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
50
+ model_input_names = ["input_ids", "attention_mask"]
51
+ _auto_class = "AutoTokenizer"
52
+
53
+ def __init__(
54
+ self,
55
+ vocab_file,
56
+ unk_token="<unk>",
57
+ bos_token="<s>",
58
+ eos_token="</s>",
59
+ pad_token="</s>",
60
+ sp_model_kwargs: Optional[Dict[str, Any]] = None,
61
+ add_bos_token=True,
62
+ add_eos_token=False,
63
+ decode_with_prefix_space=False,
64
+ clean_up_tokenization_spaces=False,
65
+ **kwargs,
66
+ ):
67
+ self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
68
+ super().__init__(
69
+ bos_token=bos_token,
70
+ eos_token=eos_token,
71
+ unk_token=unk_token,
72
+ pad_token=pad_token,
73
+ clean_up_tokenization_spaces=clean_up_tokenization_spaces,
74
+ **kwargs,
75
+ )
76
+ self.vocab_file = vocab_file
77
+ self.add_bos_token = add_bos_token
78
+ self.add_eos_token = add_eos_token
79
+ self.decode_with_prefix_space = decode_with_prefix_space
80
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
81
+ self.sp_model.Load(vocab_file)
82
+ self._no_prefix_space_tokens = None
83
+
84
+ """ Initialisation"""
85
+
86
+ @property
87
+ def no_prefix_space_tokens(self):
88
+ if self._no_prefix_space_tokens is None:
89
+ vocab = self.convert_ids_to_tokens(list(range(self.vocab_size)))
90
+ self._no_prefix_space_tokens = {i for i, tok in enumerate(vocab) if not tok.startswith("▁")}
91
+ return self._no_prefix_space_tokens
92
+
93
+ @property
94
+ def vocab_size(self):
95
+ """Returns vocab size"""
96
+ return self.sp_model.get_piece_size()
97
+
98
+ @property
99
+ def bos_token_id(self) -> Optional[int]:
100
+ return self.sp_model.bos_id()
101
+
102
+ @property
103
+ def eos_token_id(self) -> Optional[int]:
104
+ return self.sp_model.eos_id()
105
+
106
+ def get_vocab(self):
107
+ """Returns vocab as a dict"""
108
+ vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
109
+ vocab.update(self.added_tokens_encoder)
110
+ return vocab
111
+
112
+ def _tokenize(self, text):
113
+ """Returns a tokenized string."""
114
+ return self.sp_model.encode(text, out_type=str)
115
+
116
+ def _convert_token_to_id(self, token):
117
+ """Converts a token (str) in an id using the vocab."""
118
+ return self.sp_model.piece_to_id(token)
119
+
120
+ def _convert_id_to_token(self, index):
121
+ """Converts an index (integer) in a token (str) using the vocab."""
122
+ token = self.sp_model.IdToPiece(index)
123
+ return token
124
+
125
+ def _maybe_add_prefix_space(self, tokens, decoded):
126
+ if tokens and tokens[0] not in self.no_prefix_space_tokens:
127
+ return " " + decoded
128
+ else:
129
+ return decoded
130
+
131
+ def convert_tokens_to_string(self, tokens):
132
+ """Converts a sequence of tokens (string) in a single string."""
133
+ current_sub_tokens = []
134
+ out_string = ""
135
+ prev_is_special = False
136
+ for token in tokens:
137
+ # make sure that special tokens are not decoded using sentencepiece model
138
+ if token in self.all_special_tokens:
139
+ if not prev_is_special:
140
+ out_string += " "
141
+ out_string += self.sp_model.decode(current_sub_tokens) + token
142
+ prev_is_special = True
143
+ current_sub_tokens = []
144
+ else:
145
+ current_sub_tokens.append(token)
146
+ prev_is_special = False
147
+ out_string += self.sp_model.decode(current_sub_tokens)
148
+ out_string = self.clean_up_tokenization(out_string)
149
+ out_string = self._maybe_add_prefix_space(tokens=tokens, decoded=out_string)
150
+ return out_string[1:]
151
+
152
+ def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
153
+ """
154
+ Save the vocabulary and special tokens file to a directory.
155
+
156
+ Args:
157
+ save_directory (`str`):
158
+ The directory in which to save the vocabulary.
159
+
160
+ Returns:
161
+ `Tuple(str)`: Paths to the files saved.
162
+ """
163
+ if not os.path.isdir(save_directory):
164
+ logger.error(f"Vocabulary path ({save_directory}) should be a directory")
165
+ return
166
+ out_vocab_file = os.path.join(
167
+ save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
168
+ )
169
+
170
+ if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
171
+ copyfile(self.vocab_file, out_vocab_file)
172
+ elif not os.path.isfile(self.vocab_file):
173
+ with open(out_vocab_file, "wb") as fi:
174
+ content_spiece_model = self.sp_model.serialized_model_proto()
175
+ fi.write(content_spiece_model)
176
+
177
+ return (out_vocab_file,)
178
+
179
+ def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
180
+ if self.add_bos_token:
181
+ bos_token_ids = [self.bos_token_id]
182
+ else:
183
+ bos_token_ids = []
184
+
185
+ output = bos_token_ids + token_ids_0
186
+
187
+ if token_ids_1 is not None:
188
+ output = output + token_ids_1
189
+
190
+ if self.add_eos_token:
191
+ output = output + [self.eos_token_id]
192
+
193
+ return output
194
+
195
+ def get_special_tokens_mask(
196
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
197
+ ) -> List[int]:
198
+ """
199
+ Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
200
+ special tokens using the tokenizer `prepare_for_model` method.
201
+
202
+ Args:
203
+ token_ids_0 (`List[int]`):
204
+ List of IDs.
205
+ token_ids_1 (`List[int]`, *optional*):
206
+ Optional second list of IDs for sequence pairs.
207
+ already_has_special_tokens (`bool`, *optional*, defaults to `False`):
208
+ Whether or not the token list is already formatted with special tokens for the model.
209
+
210
+ Returns:
211
+ `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
212
+ """
213
+ if already_has_special_tokens:
214
+ return super().get_special_tokens_mask(
215
+ token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
216
+ )
217
+
218
+ if token_ids_1 is None:
219
+ return [1] + ([0] * len(token_ids_0)) + [1]
220
+ return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
221
+
222
+ def create_token_type_ids_from_sequences(
223
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
224
+ ) -> List[int]:
225
+ """
226
+ Create a mask from the two sequences passed to be used in a sequence-pair classification task. T5 does not make
227
+ use of token type ids, therefore a list of zeros is returned.
228
+
229
+ Args:
230
+ token_ids_0 (`List[int]`):
231
+ List of IDs.
232
+ token_ids_1 (`List[int]`, *optional*):
233
+ Optional second list of IDs for sequence pairs.
234
+
235
+ Returns:
236
+ `List[int]`: List of zeros.
237
+ """
238
+ eos = [self.eos_token_id]
239
+
240
+ if token_ids_1 is None:
241
+ return len(token_ids_0 + eos) * [0]
242
+ return len(token_ids_0 + eos + token_ids_1 + eos) * [0]
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aab622d98c98677a1a51f969e25765154487bf3e85c7819db105db2fcacba83f
3
+ size 1658691
tokenizer_config.json ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": true,
3
+ "add_eos_token": false,
4
+ "added_tokens_decoder": {
5
+ "0": {
6
+ "content": "<unk>",
7
+ "lstrip": false,
8
+ "normalized": false,
9
+ "rstrip": false,
10
+ "single_word": false,
11
+ "special": true
12
+ },
13
+ "1": {
14
+ "content": "<s>",
15
+ "lstrip": false,
16
+ "normalized": false,
17
+ "rstrip": false,
18
+ "single_word": false,
19
+ "special": true
20
+ },
21
+ "2": {
22
+ "content": "</s>",
23
+ "lstrip": false,
24
+ "normalized": false,
25
+ "rstrip": false,
26
+ "single_word": false,
27
+ "special": true
28
+ }
29
+ },
30
+ "additional_special_tokens": [],
31
+ "auto_map": {
32
+ "AutoTokenizer": [
33
+ "tokenization_internlm.InternLMTokenizer",
34
+ null
35
+ ]
36
+ },
37
+ "bos_token": "<s>",
38
+ "clean_up_tokenization_spaces": false,
39
+ "eos_token": "</s>",
40
+ "legacy": true,
41
+ "model_max_length": 1000000000000000019884624838656,
42
+ "pad_token": "</s>",
43
+ "sp_model_kwargs": {},
44
+ "spaces_between_special_tokens": false,
45
+ "tokenizer_class": "LlamaTokenizer",
46
+ "unk_token": "<unk>",
47
+ "use_default_system_prompt": false
48
+ }