clonefy commited on
Commit
7c82a23
1 Parent(s): 503ec12

Upload folder using huggingface_hub

Browse files
Files changed (47) hide show
  1. config.json +32 -0
  2. configuration.json +1 -0
  3. configuration_qwen2_rm.py +140 -0
  4. generation_config.json +14 -0
  5. merges.txt +0 -0
  6. model-00001-of-00037.safetensors +3 -0
  7. model-00002-of-00037.safetensors +3 -0
  8. model-00003-of-00037.safetensors +3 -0
  9. model-00004-of-00037.safetensors +3 -0
  10. model-00005-of-00037.safetensors +3 -0
  11. model-00006-of-00037.safetensors +3 -0
  12. model-00007-of-00037.safetensors +3 -0
  13. model-00008-of-00037.safetensors +3 -0
  14. model-00009-of-00037.safetensors +3 -0
  15. model-00010-of-00037.safetensors +3 -0
  16. model-00011-of-00037.safetensors +3 -0
  17. model-00012-of-00037.safetensors +3 -0
  18. model-00013-of-00037.safetensors +3 -0
  19. model-00014-of-00037.safetensors +3 -0
  20. model-00015-of-00037.safetensors +3 -0
  21. model-00016-of-00037.safetensors +3 -0
  22. model-00017-of-00037.safetensors +3 -0
  23. model-00018-of-00037.safetensors +3 -0
  24. model-00019-of-00037.safetensors +3 -0
  25. model-00020-of-00037.safetensors +3 -0
  26. model-00021-of-00037.safetensors +3 -0
  27. model-00022-of-00037.safetensors +3 -0
  28. model-00023-of-00037.safetensors +3 -0
  29. model-00024-of-00037.safetensors +3 -0
  30. model-00025-of-00037.safetensors +3 -0
  31. model-00026-of-00037.safetensors +3 -0
  32. model-00027-of-00037.safetensors +3 -0
  33. model-00028-of-00037.safetensors +3 -0
  34. model-00029-of-00037.safetensors +3 -0
  35. model-00030-of-00037.safetensors +3 -0
  36. model-00031-of-00037.safetensors +3 -0
  37. model-00032-of-00037.safetensors +3 -0
  38. model-00033-of-00037.safetensors +3 -0
  39. model-00034-of-00037.safetensors +3 -0
  40. model-00035-of-00037.safetensors +3 -0
  41. model-00036-of-00037.safetensors +3 -0
  42. model-00037-of-00037.safetensors +3 -0
  43. model.safetensors.index.json +974 -0
  44. modeling_qwen2_rm.py +1549 -0
  45. tokenizer.json +0 -0
  46. tokenizer_config.json +207 -0
  47. vocab.json +0 -0
config.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen2ForRewardModel"
4
+ ],
5
+ "attention_dropout": 0.0,
6
+ "auto_map": {
7
+ "AutoConfig": "configuration_qwen2_rm.Qwen2RMConfig",
8
+ "AutoModel": "modeling_qwen2_rm.Qwen2ForRewardModel"
9
+ },
10
+ "bos_token_id": 151643,
11
+ "eos_token_id": 151645,
12
+ "hidden_act": "silu",
13
+ "hidden_size": 8192,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 29568,
16
+ "max_position_embeddings": 4096,
17
+ "max_window_layers": 70,
18
+ "model_type": "qwen2",
19
+ "num_attention_heads": 64,
20
+ "num_hidden_layers": 80,
21
+ "num_key_value_heads": 8,
22
+ "rms_norm_eps": 1e-06,
23
+ "rope_theta": 10000.0,
24
+ "sliding_window": 4096,
25
+ "tie_word_embeddings": false,
26
+ "torch_dtype": "bfloat16",
27
+ "transformers_version": "4.37.0",
28
+ "use_cache": true,
29
+ "use_mrope": false,
30
+ "use_sliding_window": false,
31
+ "vocab_size": 152064
32
+ }
configuration.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"framework":"Pytorch","task":"text-generation"}
configuration_qwen2_rm.py ADDED
@@ -0,0 +1,140 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ """Qwen2 model configuration"""
16
+
17
+ from transformers.configuration_utils import PretrainedConfig
18
+ from transformers.utils import logging
19
+
20
+
21
+ logger = logging.get_logger(__name__)
22
+
23
+
24
+ class Qwen2RMConfig(PretrainedConfig):
25
+ r"""
26
+ This is the configuration class to store the configuration of a [`Qwen2Model`]. It is used to instantiate a
27
+ Qwen2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
28
+ with the defaults will yield a similar configuration to that of
29
+ Qwen2-7B-beta [Qwen/Qwen2-7B-beta](https://huggingface.co/Qwen/Qwen2-7B-beta).
30
+
31
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
32
+ documentation from [`PretrainedConfig`] for more information.
33
+
34
+
35
+ Args:
36
+ vocab_size (`int`, *optional*, defaults to 151936):
37
+ Vocabulary size of the Qwen2 model. Defines the number of different tokens that can be represented by the
38
+ `inputs_ids` passed when calling [`Qwen2Model`]
39
+ hidden_size (`int`, *optional*, defaults to 4096):
40
+ Dimension of the hidden representations.
41
+ intermediate_size (`int`, *optional*, defaults to 22016):
42
+ Dimension of the MLP representations.
43
+ num_hidden_layers (`int`, *optional*, defaults to 32):
44
+ Number of hidden layers in the Transformer encoder.
45
+ num_attention_heads (`int`, *optional*, defaults to 32):
46
+ Number of attention heads for each attention layer in the Transformer encoder.
47
+ num_key_value_heads (`int`, *optional*, defaults to 32):
48
+ This is the number of key_value heads that should be used to implement Grouped Query Attention. If
49
+ `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
50
+ `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
51
+ converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
52
+ by meanpooling all the original heads within that group. For more details checkout [this
53
+ paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
54
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
55
+ The non-linear activation function (function or string) in the decoder.
56
+ max_position_embeddings (`int`, *optional*, defaults to 32768):
57
+ The maximum sequence length that this model might ever be used with.
58
+ initializer_range (`float`, *optional*, defaults to 0.02):
59
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
60
+ rms_norm_eps (`float`, *optional*, defaults to 1e-06):
61
+ The epsilon used by the rms normalization layers.
62
+ use_cache (`bool`, *optional*, defaults to `True`):
63
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
64
+ relevant if `config.is_decoder=True`.
65
+ tie_word_embeddings (`bool`, *optional*, defaults to `False`):
66
+ Whether the model's input and output word embeddings should be tied.
67
+ rope_theta (`float`, *optional*, defaults to 10000.0):
68
+ The base period of the RoPE embeddings.
69
+ use_sliding_window (`bool`, *optional*, defaults to `False`):
70
+ Whether to use sliding window attention.
71
+ sliding_window (`int`, *optional*, defaults to 4096):
72
+ Sliding window attention (SWA) window size. If not specified, will default to `4096`.
73
+ max_window_layers (`int`, *optional*, defaults to 28):
74
+ The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
75
+ attention_dropout (`float`, *optional*, defaults to 0.0):
76
+ The dropout ratio for the attention probabilities.
77
+
78
+ ```python
79
+ >>> from transformers import Qwen2Model, Qwen2Config
80
+
81
+ >>> # Initializing a Qwen2 style configuration
82
+ >>> configuration = Qwen2Config()
83
+
84
+ >>> # Initializing a model from the Qwen2-7B style configuration
85
+ >>> model = Qwen2Model(configuration)
86
+
87
+ >>> # Accessing the model configuration
88
+ >>> configuration = model.config
89
+ ```"""
90
+
91
+ model_type = "qwen2"
92
+ keys_to_ignore_at_inference = ["past_key_values"]
93
+
94
+ def __init__(
95
+ self,
96
+ vocab_size=151936,
97
+ hidden_size=4096,
98
+ intermediate_size=22016,
99
+ num_hidden_layers=32,
100
+ num_attention_heads=32,
101
+ num_key_value_heads=32,
102
+ hidden_act="silu",
103
+ max_position_embeddings=32768,
104
+ initializer_range=0.02,
105
+ rms_norm_eps=1e-6,
106
+ use_cache=True,
107
+ tie_word_embeddings=False,
108
+ rope_theta=10000.0,
109
+ use_sliding_window=False,
110
+ sliding_window=4096,
111
+ max_window_layers=28,
112
+ attention_dropout=0.0,
113
+ **kwargs,
114
+ ):
115
+ self.vocab_size = vocab_size
116
+ self.max_position_embeddings = max_position_embeddings
117
+ self.hidden_size = hidden_size
118
+ self.intermediate_size = intermediate_size
119
+ self.num_hidden_layers = num_hidden_layers
120
+ self.num_attention_heads = num_attention_heads
121
+ self.use_sliding_window = use_sliding_window
122
+ self.sliding_window = sliding_window if use_sliding_window else None
123
+ self.max_window_layers = max_window_layers
124
+
125
+ # for backward compatibility
126
+ if num_key_value_heads is None:
127
+ num_key_value_heads = num_attention_heads
128
+
129
+ self.num_key_value_heads = num_key_value_heads
130
+ self.hidden_act = hidden_act
131
+ self.initializer_range = initializer_range
132
+ self.rms_norm_eps = rms_norm_eps
133
+ self.use_cache = use_cache
134
+ self.rope_theta = rope_theta
135
+ self.attention_dropout = attention_dropout
136
+
137
+ super().__init__(
138
+ tie_word_embeddings=tie_word_embeddings,
139
+ **kwargs,
140
+ )
generation_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 151643,
3
+ "pad_token_id": 151643,
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 151645,
7
+ 151643
8
+ ],
9
+ "repetition_penalty": 1.05,
10
+ "temperature": 0.7,
11
+ "top_p": 0.8,
12
+ "top_k": 20,
13
+ "transformers_version": "4.37.0"
14
+ }
merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
model-00001-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a4ff4e49797d72ac12cff8b3ebe8488a00d98bdf202b4dd0d87fed9fe9166276
3
+ size 3896596186
model-00002-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:07b6891108872fc38ab1aea229915abb59a2eba9e69880f648e3cc8ff748edee
3
+ size 3995200440
model-00003-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1cc4cddbf816d130b04530704cc7c9d38ba2e3a955174ce2c1510204eacf82c4
3
+ size 3812769392
model-00004-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c8aaf78a5799ab58a90a13f6580905d765c923762e918549d5a48534e109e02f
3
+ size 3995183944
model-00005-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f74b1212f4455b678d9bc236050fc8fe9d2b560ffc425d43b3245ca053d71e5e
3
+ size 3995183944
model-00006-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:395dd26a2ced72b7ec796e08ba4b2564132b8a0d8e54790ede5a7767d8902815
3
+ size 3995200456
model-00007-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:725687ad273a3a6242dee3f9892ac53d1bf072d8885163e28da69b2ac2cba407
3
+ size 3812769424
model-00008-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f8cf70b85514693e88e3bf2dc7ec2bea2505eefc628989b03a12c45f1b60a2a9
3
+ size 3995183968
model-00009-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a81fa5fcfdf62bf74e0b7697b264caee44d059d4447c8b6efe2fc3dcc67d37e7
3
+ size 3995183968
model-00010-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:154c01e23dd96ecf4947222394b21a056b48dc01c17f16edbd953f6c839a6437
3
+ size 3995200464
model-00011-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b56bd37a8890a6d14f3c8b764ed135933444f854d517e901a0b513f316205f41
3
+ size 3812769424
model-00012-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b0eb2d70107a95ce7cb43fc7d53e3bdc8fbf36e9f3e3a55ce4207acbc36036e4
3
+ size 3995183968
model-00013-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:36150cdb2324a4c0c0fb44a6d490c4a74ce414aef4a1e9b47d5ef100a127b45b
3
+ size 3995183968
model-00014-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ab9ea0058b43b204b109e7154b738112cfad681f903aafb34b452b6d5518df58
3
+ size 3995200464
model-00015-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ff0649985d41cf95a58caba7ecbcc76551555e25d7e7ba37a5f33bc2341c899a
3
+ size 3812769424
model-00016-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c909a856b0c40efbb4774671f31fb533d93d29d32f38c2fcf8b4a88a4d33f11b
3
+ size 3995183968
model-00017-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dca882ba953f08e2084df51a88fe24092ec9dee58ca8edd9909a2965cbe7286b
3
+ size 3995183968
model-00018-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6958e000731e462f334e72678be3356064d01a4841c00807f85a66e2720b6cb
3
+ size 3995200464
model-00019-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3a1a2d371ed9b6e3e32c6334690c89df004b75704097078d717e9530deffc71c
3
+ size 3812769424
model-00020-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5711c5b7b2c2828f4f94d27778ced388f835153424d75f67ca987de7301d00af
3
+ size 3995183968
model-00021-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8982522cb15675c2b2250f39eeb9c23a69f816e65ca90fb885e0040cdd5abbb0
3
+ size 3995183968
model-00022-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bbdc3a59afdd70ceb27159a992ad5f640d4ce4109acfd332a0dc7d0f987c490c
3
+ size 3995200464
model-00023-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e8a83e1742c3d1902eae9120e3eeb480326ef842ac87c1602ea1193e27ea3bbb
3
+ size 3812769424
model-00024-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3110154c0c919d2505df3d3f4f3dad152536a3d2e26e2355bdc01271712a6e5f
3
+ size 3995183968
model-00025-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:04d2b66c37a14cbcb1f21c7053a7527d4d18cda424d29e584aa443a2bf695e69
3
+ size 3995183968
model-00026-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:75e44563fa0126205baf0ac44f4979f0e88c1d911eef1782f1bba875300d858e
3
+ size 3995200464
model-00027-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:58007df9233375734b68e6f6d3d6063699c273ec0e232fb4fd4b3f0c966cb863
3
+ size 3812769424
model-00028-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1a6e49db60f0157293816a0866cd31c10b0af88d4ed9af84ae464da88aa2a151
3
+ size 3995183968
model-00029-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:11812a923332a7c46213c4432cf6b33456842d0625eb05d1bd12ab77bb27f9c6
3
+ size 3995183968
model-00030-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0e7e3eb9cd2d5d7ec2a9ec3e4681b55270c2eb344966872b0370a4496f9aae31
3
+ size 3995200464
model-00031-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:30a8f2cb09d02368dec9abe72afc601975a54e33d5c394831e5eed0ae2102316
3
+ size 3812769424
model-00032-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:18ba78df4a02489d11b29f2605915b7061c0a6fae49d3f5c836cff238e82e378
3
+ size 3995183968
model-00033-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a0f1196282308fc69537d268a3c6c07492916d85054a34c2bafa7346932880d0
3
+ size 3995183968
model-00034-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cf2866f23272a8d59804024e4ab8fdad24c0d697e42a4b1780228712c1542589
3
+ size 3995200464
model-00035-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0ddf03d616cee7d2b58f09f31cccf0f132532ac6980156ccc3b41f6cfc48d69d
3
+ size 3812769424
model-00036-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2a25a398d9eb190a3e485c012716d6f6e73d3c611aaee7b090e4df1e22c15be0
3
+ size 3995183968
model-00037-of-00037.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b1138966cbbeb25b25e787ebd0ab9e6371032f05b0ed5fa3e90abf3c7364399a
3
+ size 3460317640
model.safetensors.index.json ADDED
@@ -0,0 +1,974 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 145546657794
4
+ },
5
+ "weight_map": {
6
+ "lm_head.weight": "model-00037-of-00037.safetensors",
7
+ "model.embed_tokens.weight": "model-00001-of-00037.safetensors",
8
+ "model.layers.0.input_layernorm.weight": "model-00001-of-00037.safetensors",
9
+ "model.layers.0.mlp.down_proj.weight": "model-00002-of-00037.safetensors",
10
+ "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00037.safetensors",
11
+ "model.layers.0.mlp.up_proj.weight": "model-00001-of-00037.safetensors",
12
+ "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00037.safetensors",
13
+ "model.layers.0.self_attn.k_proj.bias": "model-00001-of-00037.safetensors",
14
+ "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00037.safetensors",
15
+ "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00037.safetensors",
16
+ "model.layers.0.self_attn.q_proj.bias": "model-00001-of-00037.safetensors",
17
+ "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00037.safetensors",
18
+ "model.layers.0.self_attn.v_proj.bias": "model-00001-of-00037.safetensors",
19
+ "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00037.safetensors",
20
+ "model.layers.1.input_layernorm.weight": "model-00002-of-00037.safetensors",
21
+ "model.layers.1.mlp.down_proj.weight": "model-00002-of-00037.safetensors",
22
+ "model.layers.1.mlp.gate_proj.weight": "model-00002-of-00037.safetensors",
23
+ "model.layers.1.mlp.up_proj.weight": "model-00002-of-00037.safetensors",
24
+ "model.layers.1.post_attention_layernorm.weight": "model-00002-of-00037.safetensors",
25
+ "model.layers.1.self_attn.k_proj.bias": "model-00002-of-00037.safetensors",
26
+ "model.layers.1.self_attn.k_proj.weight": "model-00002-of-00037.safetensors",
27
+ "model.layers.1.self_attn.o_proj.weight": "model-00002-of-00037.safetensors",
28
+ "model.layers.1.self_attn.q_proj.bias": "model-00002-of-00037.safetensors",
29
+ "model.layers.1.self_attn.q_proj.weight": "model-00002-of-00037.safetensors",
30
+ "model.layers.1.self_attn.v_proj.bias": "model-00002-of-00037.safetensors",
31
+ "model.layers.1.self_attn.v_proj.weight": "model-00002-of-00037.safetensors",
32
+ "model.layers.10.input_layernorm.weight": "model-00006-of-00037.safetensors",
33
+ "model.layers.10.mlp.down_proj.weight": "model-00006-of-00037.safetensors",
34
+ "model.layers.10.mlp.gate_proj.weight": "model-00006-of-00037.safetensors",
35
+ "model.layers.10.mlp.up_proj.weight": "model-00006-of-00037.safetensors",
36
+ "model.layers.10.post_attention_layernorm.weight": "model-00006-of-00037.safetensors",
37
+ "model.layers.10.self_attn.k_proj.bias": "model-00006-of-00037.safetensors",
38
+ "model.layers.10.self_attn.k_proj.weight": "model-00006-of-00037.safetensors",
39
+ "model.layers.10.self_attn.o_proj.weight": "model-00006-of-00037.safetensors",
40
+ "model.layers.10.self_attn.q_proj.bias": "model-00006-of-00037.safetensors",
41
+ "model.layers.10.self_attn.q_proj.weight": "model-00006-of-00037.safetensors",
42
+ "model.layers.10.self_attn.v_proj.bias": "model-00006-of-00037.safetensors",
43
+ "model.layers.10.self_attn.v_proj.weight": "model-00006-of-00037.safetensors",
44
+ "model.layers.11.input_layernorm.weight": "model-00006-of-00037.safetensors",
45
+ "model.layers.11.mlp.down_proj.weight": "model-00006-of-00037.safetensors",
46
+ "model.layers.11.mlp.gate_proj.weight": "model-00006-of-00037.safetensors",
47
+ "model.layers.11.mlp.up_proj.weight": "model-00006-of-00037.safetensors",
48
+ "model.layers.11.post_attention_layernorm.weight": "model-00006-of-00037.safetensors",
49
+ "model.layers.11.self_attn.k_proj.bias": "model-00006-of-00037.safetensors",
50
+ "model.layers.11.self_attn.k_proj.weight": "model-00006-of-00037.safetensors",
51
+ "model.layers.11.self_attn.o_proj.weight": "model-00006-of-00037.safetensors",
52
+ "model.layers.11.self_attn.q_proj.bias": "model-00006-of-00037.safetensors",
53
+ "model.layers.11.self_attn.q_proj.weight": "model-00006-of-00037.safetensors",
54
+ "model.layers.11.self_attn.v_proj.bias": "model-00006-of-00037.safetensors",
55
+ "model.layers.11.self_attn.v_proj.weight": "model-00006-of-00037.safetensors",
56
+ "model.layers.12.input_layernorm.weight": "model-00006-of-00037.safetensors",
57
+ "model.layers.12.mlp.down_proj.weight": "model-00007-of-00037.safetensors",
58
+ "model.layers.12.mlp.gate_proj.weight": "model-00007-of-00037.safetensors",
59
+ "model.layers.12.mlp.up_proj.weight": "model-00007-of-00037.safetensors",
60
+ "model.layers.12.post_attention_layernorm.weight": "model-00007-of-00037.safetensors",
61
+ "model.layers.12.self_attn.k_proj.bias": "model-00007-of-00037.safetensors",
62
+ "model.layers.12.self_attn.k_proj.weight": "model-00007-of-00037.safetensors",
63
+ "model.layers.12.self_attn.o_proj.weight": "model-00007-of-00037.safetensors",
64
+ "model.layers.12.self_attn.q_proj.bias": "model-00007-of-00037.safetensors",
65
+ "model.layers.12.self_attn.q_proj.weight": "model-00007-of-00037.safetensors",
66
+ "model.layers.12.self_attn.v_proj.bias": "model-00007-of-00037.safetensors",
67
+ "model.layers.12.self_attn.v_proj.weight": "model-00007-of-00037.safetensors",
68
+ "model.layers.13.input_layernorm.weight": "model-00007-of-00037.safetensors",
69
+ "model.layers.13.mlp.down_proj.weight": "model-00007-of-00037.safetensors",
70
+ "model.layers.13.mlp.gate_proj.weight": "model-00007-of-00037.safetensors",
71
+ "model.layers.13.mlp.up_proj.weight": "model-00007-of-00037.safetensors",
72
+ "model.layers.13.post_attention_layernorm.weight": "model-00007-of-00037.safetensors",
73
+ "model.layers.13.self_attn.k_proj.bias": "model-00007-of-00037.safetensors",
74
+ "model.layers.13.self_attn.k_proj.weight": "model-00007-of-00037.safetensors",
75
+ "model.layers.13.self_attn.o_proj.weight": "model-00007-of-00037.safetensors",
76
+ "model.layers.13.self_attn.q_proj.bias": "model-00007-of-00037.safetensors",
77
+ "model.layers.13.self_attn.q_proj.weight": "model-00007-of-00037.safetensors",
78
+ "model.layers.13.self_attn.v_proj.bias": "model-00007-of-00037.safetensors",
79
+ "model.layers.13.self_attn.v_proj.weight": "model-00007-of-00037.safetensors",
80
+ "model.layers.14.input_layernorm.weight": "model-00007-of-00037.safetensors",
81
+ "model.layers.14.mlp.down_proj.weight": "model-00008-of-00037.safetensors",
82
+ "model.layers.14.mlp.gate_proj.weight": "model-00008-of-00037.safetensors",
83
+ "model.layers.14.mlp.up_proj.weight": "model-00008-of-00037.safetensors",
84
+ "model.layers.14.post_attention_layernorm.weight": "model-00007-of-00037.safetensors",
85
+ "model.layers.14.self_attn.k_proj.bias": "model-00007-of-00037.safetensors",
86
+ "model.layers.14.self_attn.k_proj.weight": "model-00007-of-00037.safetensors",
87
+ "model.layers.14.self_attn.o_proj.weight": "model-00007-of-00037.safetensors",
88
+ "model.layers.14.self_attn.q_proj.bias": "model-00007-of-00037.safetensors",
89
+ "model.layers.14.self_attn.q_proj.weight": "model-00007-of-00037.safetensors",
90
+ "model.layers.14.self_attn.v_proj.bias": "model-00007-of-00037.safetensors",
91
+ "model.layers.14.self_attn.v_proj.weight": "model-00007-of-00037.safetensors",
92
+ "model.layers.15.input_layernorm.weight": "model-00008-of-00037.safetensors",
93
+ "model.layers.15.mlp.down_proj.weight": "model-00008-of-00037.safetensors",
94
+ "model.layers.15.mlp.gate_proj.weight": "model-00008-of-00037.safetensors",
95
+ "model.layers.15.mlp.up_proj.weight": "model-00008-of-00037.safetensors",
96
+ "model.layers.15.post_attention_layernorm.weight": "model-00008-of-00037.safetensors",
97
+ "model.layers.15.self_attn.k_proj.bias": "model-00008-of-00037.safetensors",
98
+ "model.layers.15.self_attn.k_proj.weight": "model-00008-of-00037.safetensors",
99
+ "model.layers.15.self_attn.o_proj.weight": "model-00008-of-00037.safetensors",
100
+ "model.layers.15.self_attn.q_proj.bias": "model-00008-of-00037.safetensors",
101
+ "model.layers.15.self_attn.q_proj.weight": "model-00008-of-00037.safetensors",
102
+ "model.layers.15.self_attn.v_proj.bias": "model-00008-of-00037.safetensors",
103
+ "model.layers.15.self_attn.v_proj.weight": "model-00008-of-00037.safetensors",
104
+ "model.layers.16.input_layernorm.weight": "model-00008-of-00037.safetensors",
105
+ "model.layers.16.mlp.down_proj.weight": "model-00009-of-00037.safetensors",
106
+ "model.layers.16.mlp.gate_proj.weight": "model-00009-of-00037.safetensors",
107
+ "model.layers.16.mlp.up_proj.weight": "model-00008-of-00037.safetensors",
108
+ "model.layers.16.post_attention_layernorm.weight": "model-00008-of-00037.safetensors",
109
+ "model.layers.16.self_attn.k_proj.bias": "model-00008-of-00037.safetensors",
110
+ "model.layers.16.self_attn.k_proj.weight": "model-00008-of-00037.safetensors",
111
+ "model.layers.16.self_attn.o_proj.weight": "model-00008-of-00037.safetensors",
112
+ "model.layers.16.self_attn.q_proj.bias": "model-00008-of-00037.safetensors",
113
+ "model.layers.16.self_attn.q_proj.weight": "model-00008-of-00037.safetensors",
114
+ "model.layers.16.self_attn.v_proj.bias": "model-00008-of-00037.safetensors",
115
+ "model.layers.16.self_attn.v_proj.weight": "model-00008-of-00037.safetensors",
116
+ "model.layers.17.input_layernorm.weight": "model-00009-of-00037.safetensors",
117
+ "model.layers.17.mlp.down_proj.weight": "model-00009-of-00037.safetensors",
118
+ "model.layers.17.mlp.gate_proj.weight": "model-00009-of-00037.safetensors",
119
+ "model.layers.17.mlp.up_proj.weight": "model-00009-of-00037.safetensors",
120
+ "model.layers.17.post_attention_layernorm.weight": "model-00009-of-00037.safetensors",
121
+ "model.layers.17.self_attn.k_proj.bias": "model-00009-of-00037.safetensors",
122
+ "model.layers.17.self_attn.k_proj.weight": "model-00009-of-00037.safetensors",
123
+ "model.layers.17.self_attn.o_proj.weight": "model-00009-of-00037.safetensors",
124
+ "model.layers.17.self_attn.q_proj.bias": "model-00009-of-00037.safetensors",
125
+ "model.layers.17.self_attn.q_proj.weight": "model-00009-of-00037.safetensors",
126
+ "model.layers.17.self_attn.v_proj.bias": "model-00009-of-00037.safetensors",
127
+ "model.layers.17.self_attn.v_proj.weight": "model-00009-of-00037.safetensors",
128
+ "model.layers.18.input_layernorm.weight": "model-00009-of-00037.safetensors",
129
+ "model.layers.18.mlp.down_proj.weight": "model-00010-of-00037.safetensors",
130
+ "model.layers.18.mlp.gate_proj.weight": "model-00009-of-00037.safetensors",
131
+ "model.layers.18.mlp.up_proj.weight": "model-00009-of-00037.safetensors",
132
+ "model.layers.18.post_attention_layernorm.weight": "model-00009-of-00037.safetensors",
133
+ "model.layers.18.self_attn.k_proj.bias": "model-00009-of-00037.safetensors",
134
+ "model.layers.18.self_attn.k_proj.weight": "model-00009-of-00037.safetensors",
135
+ "model.layers.18.self_attn.o_proj.weight": "model-00009-of-00037.safetensors",
136
+ "model.layers.18.self_attn.q_proj.bias": "model-00009-of-00037.safetensors",
137
+ "model.layers.18.self_attn.q_proj.weight": "model-00009-of-00037.safetensors",
138
+ "model.layers.18.self_attn.v_proj.bias": "model-00009-of-00037.safetensors",
139
+ "model.layers.18.self_attn.v_proj.weight": "model-00009-of-00037.safetensors",
140
+ "model.layers.19.input_layernorm.weight": "model-00010-of-00037.safetensors",
141
+ "model.layers.19.mlp.down_proj.weight": "model-00010-of-00037.safetensors",
142
+ "model.layers.19.mlp.gate_proj.weight": "model-00010-of-00037.safetensors",
143
+ "model.layers.19.mlp.up_proj.weight": "model-00010-of-00037.safetensors",
144
+ "model.layers.19.post_attention_layernorm.weight": "model-00010-of-00037.safetensors",
145
+ "model.layers.19.self_attn.k_proj.bias": "model-00010-of-00037.safetensors",
146
+ "model.layers.19.self_attn.k_proj.weight": "model-00010-of-00037.safetensors",
147
+ "model.layers.19.self_attn.o_proj.weight": "model-00010-of-00037.safetensors",
148
+ "model.layers.19.self_attn.q_proj.bias": "model-00010-of-00037.safetensors",
149
+ "model.layers.19.self_attn.q_proj.weight": "model-00010-of-00037.safetensors",
150
+ "model.layers.19.self_attn.v_proj.bias": "model-00010-of-00037.safetensors",
151
+ "model.layers.19.self_attn.v_proj.weight": "model-00010-of-00037.safetensors",
152
+ "model.layers.2.input_layernorm.weight": "model-00002-of-00037.safetensors",
153
+ "model.layers.2.mlp.down_proj.weight": "model-00002-of-00037.safetensors",
154
+ "model.layers.2.mlp.gate_proj.weight": "model-00002-of-00037.safetensors",
155
+ "model.layers.2.mlp.up_proj.weight": "model-00002-of-00037.safetensors",
156
+ "model.layers.2.post_attention_layernorm.weight": "model-00002-of-00037.safetensors",
157
+ "model.layers.2.self_attn.k_proj.bias": "model-00002-of-00037.safetensors",
158
+ "model.layers.2.self_attn.k_proj.weight": "model-00002-of-00037.safetensors",
159
+ "model.layers.2.self_attn.o_proj.weight": "model-00002-of-00037.safetensors",
160
+ "model.layers.2.self_attn.q_proj.bias": "model-00002-of-00037.safetensors",
161
+ "model.layers.2.self_attn.q_proj.weight": "model-00002-of-00037.safetensors",
162
+ "model.layers.2.self_attn.v_proj.bias": "model-00002-of-00037.safetensors",
163
+ "model.layers.2.self_attn.v_proj.weight": "model-00002-of-00037.safetensors",
164
+ "model.layers.20.input_layernorm.weight": "model-00010-of-00037.safetensors",
165
+ "model.layers.20.mlp.down_proj.weight": "model-00010-of-00037.safetensors",
166
+ "model.layers.20.mlp.gate_proj.weight": "model-00010-of-00037.safetensors",
167
+ "model.layers.20.mlp.up_proj.weight": "model-00010-of-00037.safetensors",
168
+ "model.layers.20.post_attention_layernorm.weight": "model-00010-of-00037.safetensors",
169
+ "model.layers.20.self_attn.k_proj.bias": "model-00010-of-00037.safetensors",
170
+ "model.layers.20.self_attn.k_proj.weight": "model-00010-of-00037.safetensors",
171
+ "model.layers.20.self_attn.o_proj.weight": "model-00010-of-00037.safetensors",
172
+ "model.layers.20.self_attn.q_proj.bias": "model-00010-of-00037.safetensors",
173
+ "model.layers.20.self_attn.q_proj.weight": "model-00010-of-00037.safetensors",
174
+ "model.layers.20.self_attn.v_proj.bias": "model-00010-of-00037.safetensors",
175
+ "model.layers.20.self_attn.v_proj.weight": "model-00010-of-00037.safetensors",
176
+ "model.layers.21.input_layernorm.weight": "model-00010-of-00037.safetensors",
177
+ "model.layers.21.mlp.down_proj.weight": "model-00011-of-00037.safetensors",
178
+ "model.layers.21.mlp.gate_proj.weight": "model-00011-of-00037.safetensors",
179
+ "model.layers.21.mlp.up_proj.weight": "model-00011-of-00037.safetensors",
180
+ "model.layers.21.post_attention_layernorm.weight": "model-00011-of-00037.safetensors",
181
+ "model.layers.21.self_attn.k_proj.bias": "model-00011-of-00037.safetensors",
182
+ "model.layers.21.self_attn.k_proj.weight": "model-00011-of-00037.safetensors",
183
+ "model.layers.21.self_attn.o_proj.weight": "model-00011-of-00037.safetensors",
184
+ "model.layers.21.self_attn.q_proj.bias": "model-00011-of-00037.safetensors",
185
+ "model.layers.21.self_attn.q_proj.weight": "model-00011-of-00037.safetensors",
186
+ "model.layers.21.self_attn.v_proj.bias": "model-00011-of-00037.safetensors",
187
+ "model.layers.21.self_attn.v_proj.weight": "model-00011-of-00037.safetensors",
188
+ "model.layers.22.input_layernorm.weight": "model-00011-of-00037.safetensors",
189
+ "model.layers.22.mlp.down_proj.weight": "model-00011-of-00037.safetensors",
190
+ "model.layers.22.mlp.gate_proj.weight": "model-00011-of-00037.safetensors",
191
+ "model.layers.22.mlp.up_proj.weight": "model-00011-of-00037.safetensors",
192
+ "model.layers.22.post_attention_layernorm.weight": "model-00011-of-00037.safetensors",
193
+ "model.layers.22.self_attn.k_proj.bias": "model-00011-of-00037.safetensors",
194
+ "model.layers.22.self_attn.k_proj.weight": "model-00011-of-00037.safetensors",
195
+ "model.layers.22.self_attn.o_proj.weight": "model-00011-of-00037.safetensors",
196
+ "model.layers.22.self_attn.q_proj.bias": "model-00011-of-00037.safetensors",
197
+ "model.layers.22.self_attn.q_proj.weight": "model-00011-of-00037.safetensors",
198
+ "model.layers.22.self_attn.v_proj.bias": "model-00011-of-00037.safetensors",
199
+ "model.layers.22.self_attn.v_proj.weight": "model-00011-of-00037.safetensors",
200
+ "model.layers.23.input_layernorm.weight": "model-00011-of-00037.safetensors",
201
+ "model.layers.23.mlp.down_proj.weight": "model-00012-of-00037.safetensors",
202
+ "model.layers.23.mlp.gate_proj.weight": "model-00012-of-00037.safetensors",
203
+ "model.layers.23.mlp.up_proj.weight": "model-00012-of-00037.safetensors",
204
+ "model.layers.23.post_attention_layernorm.weight": "model-00011-of-00037.safetensors",
205
+ "model.layers.23.self_attn.k_proj.bias": "model-00011-of-00037.safetensors",
206
+ "model.layers.23.self_attn.k_proj.weight": "model-00011-of-00037.safetensors",
207
+ "model.layers.23.self_attn.o_proj.weight": "model-00011-of-00037.safetensors",
208
+ "model.layers.23.self_attn.q_proj.bias": "model-00011-of-00037.safetensors",
209
+ "model.layers.23.self_attn.q_proj.weight": "model-00011-of-00037.safetensors",
210
+ "model.layers.23.self_attn.v_proj.bias": "model-00011-of-00037.safetensors",
211
+ "model.layers.23.self_attn.v_proj.weight": "model-00011-of-00037.safetensors",
212
+ "model.layers.24.input_layernorm.weight": "model-00012-of-00037.safetensors",
213
+ "model.layers.24.mlp.down_proj.weight": "model-00012-of-00037.safetensors",
214
+ "model.layers.24.mlp.gate_proj.weight": "model-00012-of-00037.safetensors",
215
+ "model.layers.24.mlp.up_proj.weight": "model-00012-of-00037.safetensors",
216
+ "model.layers.24.post_attention_layernorm.weight": "model-00012-of-00037.safetensors",
217
+ "model.layers.24.self_attn.k_proj.bias": "model-00012-of-00037.safetensors",
218
+ "model.layers.24.self_attn.k_proj.weight": "model-00012-of-00037.safetensors",
219
+ "model.layers.24.self_attn.o_proj.weight": "model-00012-of-00037.safetensors",
220
+ "model.layers.24.self_attn.q_proj.bias": "model-00012-of-00037.safetensors",
221
+ "model.layers.24.self_attn.q_proj.weight": "model-00012-of-00037.safetensors",
222
+ "model.layers.24.self_attn.v_proj.bias": "model-00012-of-00037.safetensors",
223
+ "model.layers.24.self_attn.v_proj.weight": "model-00012-of-00037.safetensors",
224
+ "model.layers.25.input_layernorm.weight": "model-00012-of-00037.safetensors",
225
+ "model.layers.25.mlp.down_proj.weight": "model-00013-of-00037.safetensors",
226
+ "model.layers.25.mlp.gate_proj.weight": "model-00013-of-00037.safetensors",
227
+ "model.layers.25.mlp.up_proj.weight": "model-00012-of-00037.safetensors",
228
+ "model.layers.25.post_attention_layernorm.weight": "model-00012-of-00037.safetensors",
229
+ "model.layers.25.self_attn.k_proj.bias": "model-00012-of-00037.safetensors",
230
+ "model.layers.25.self_attn.k_proj.weight": "model-00012-of-00037.safetensors",
231
+ "model.layers.25.self_attn.o_proj.weight": "model-00012-of-00037.safetensors",
232
+ "model.layers.25.self_attn.q_proj.bias": "model-00012-of-00037.safetensors",
233
+ "model.layers.25.self_attn.q_proj.weight": "model-00012-of-00037.safetensors",
234
+ "model.layers.25.self_attn.v_proj.bias": "model-00012-of-00037.safetensors",
235
+ "model.layers.25.self_attn.v_proj.weight": "model-00012-of-00037.safetensors",
236
+ "model.layers.26.input_layernorm.weight": "model-00013-of-00037.safetensors",
237
+ "model.layers.26.mlp.down_proj.weight": "model-00013-of-00037.safetensors",
238
+ "model.layers.26.mlp.gate_proj.weight": "model-00013-of-00037.safetensors",
239
+ "model.layers.26.mlp.up_proj.weight": "model-00013-of-00037.safetensors",
240
+ "model.layers.26.post_attention_layernorm.weight": "model-00013-of-00037.safetensors",
241
+ "model.layers.26.self_attn.k_proj.bias": "model-00013-of-00037.safetensors",
242
+ "model.layers.26.self_attn.k_proj.weight": "model-00013-of-00037.safetensors",
243
+ "model.layers.26.self_attn.o_proj.weight": "model-00013-of-00037.safetensors",
244
+ "model.layers.26.self_attn.q_proj.bias": "model-00013-of-00037.safetensors",
245
+ "model.layers.26.self_attn.q_proj.weight": "model-00013-of-00037.safetensors",
246
+ "model.layers.26.self_attn.v_proj.bias": "model-00013-of-00037.safetensors",
247
+ "model.layers.26.self_attn.v_proj.weight": "model-00013-of-00037.safetensors",
248
+ "model.layers.27.input_layernorm.weight": "model-00013-of-00037.safetensors",
249
+ "model.layers.27.mlp.down_proj.weight": "model-00014-of-00037.safetensors",
250
+ "model.layers.27.mlp.gate_proj.weight": "model-00013-of-00037.safetensors",
251
+ "model.layers.27.mlp.up_proj.weight": "model-00013-of-00037.safetensors",
252
+ "model.layers.27.post_attention_layernorm.weight": "model-00013-of-00037.safetensors",
253
+ "model.layers.27.self_attn.k_proj.bias": "model-00013-of-00037.safetensors",
254
+ "model.layers.27.self_attn.k_proj.weight": "model-00013-of-00037.safetensors",
255
+ "model.layers.27.self_attn.o_proj.weight": "model-00013-of-00037.safetensors",
256
+ "model.layers.27.self_attn.q_proj.bias": "model-00013-of-00037.safetensors",
257
+ "model.layers.27.self_attn.q_proj.weight": "model-00013-of-00037.safetensors",
258
+ "model.layers.27.self_attn.v_proj.bias": "model-00013-of-00037.safetensors",
259
+ "model.layers.27.self_attn.v_proj.weight": "model-00013-of-00037.safetensors",
260
+ "model.layers.28.input_layernorm.weight": "model-00014-of-00037.safetensors",
261
+ "model.layers.28.mlp.down_proj.weight": "model-00014-of-00037.safetensors",
262
+ "model.layers.28.mlp.gate_proj.weight": "model-00014-of-00037.safetensors",
263
+ "model.layers.28.mlp.up_proj.weight": "model-00014-of-00037.safetensors",
264
+ "model.layers.28.post_attention_layernorm.weight": "model-00014-of-00037.safetensors",
265
+ "model.layers.28.self_attn.k_proj.bias": "model-00014-of-00037.safetensors",
266
+ "model.layers.28.self_attn.k_proj.weight": "model-00014-of-00037.safetensors",
267
+ "model.layers.28.self_attn.o_proj.weight": "model-00014-of-00037.safetensors",
268
+ "model.layers.28.self_attn.q_proj.bias": "model-00014-of-00037.safetensors",
269
+ "model.layers.28.self_attn.q_proj.weight": "model-00014-of-00037.safetensors",
270
+ "model.layers.28.self_attn.v_proj.bias": "model-00014-of-00037.safetensors",
271
+ "model.layers.28.self_attn.v_proj.weight": "model-00014-of-00037.safetensors",
272
+ "model.layers.29.input_layernorm.weight": "model-00014-of-00037.safetensors",
273
+ "model.layers.29.mlp.down_proj.weight": "model-00014-of-00037.safetensors",
274
+ "model.layers.29.mlp.gate_proj.weight": "model-00014-of-00037.safetensors",
275
+ "model.layers.29.mlp.up_proj.weight": "model-00014-of-00037.safetensors",
276
+ "model.layers.29.post_attention_layernorm.weight": "model-00014-of-00037.safetensors",
277
+ "model.layers.29.self_attn.k_proj.bias": "model-00014-of-00037.safetensors",
278
+ "model.layers.29.self_attn.k_proj.weight": "model-00014-of-00037.safetensors",
279
+ "model.layers.29.self_attn.o_proj.weight": "model-00014-of-00037.safetensors",
280
+ "model.layers.29.self_attn.q_proj.bias": "model-00014-of-00037.safetensors",
281
+ "model.layers.29.self_attn.q_proj.weight": "model-00014-of-00037.safetensors",
282
+ "model.layers.29.self_attn.v_proj.bias": "model-00014-of-00037.safetensors",
283
+ "model.layers.29.self_attn.v_proj.weight": "model-00014-of-00037.safetensors",
284
+ "model.layers.3.input_layernorm.weight": "model-00002-of-00037.safetensors",
285
+ "model.layers.3.mlp.down_proj.weight": "model-00003-of-00037.safetensors",
286
+ "model.layers.3.mlp.gate_proj.weight": "model-00003-of-00037.safetensors",
287
+ "model.layers.3.mlp.up_proj.weight": "model-00003-of-00037.safetensors",
288
+ "model.layers.3.post_attention_layernorm.weight": "model-00003-of-00037.safetensors",
289
+ "model.layers.3.self_attn.k_proj.bias": "model-00003-of-00037.safetensors",
290
+ "model.layers.3.self_attn.k_proj.weight": "model-00003-of-00037.safetensors",
291
+ "model.layers.3.self_attn.o_proj.weight": "model-00003-of-00037.safetensors",
292
+ "model.layers.3.self_attn.q_proj.bias": "model-00003-of-00037.safetensors",
293
+ "model.layers.3.self_attn.q_proj.weight": "model-00003-of-00037.safetensors",
294
+ "model.layers.3.self_attn.v_proj.bias": "model-00003-of-00037.safetensors",
295
+ "model.layers.3.self_attn.v_proj.weight": "model-00003-of-00037.safetensors",
296
+ "model.layers.30.input_layernorm.weight": "model-00014-of-00037.safetensors",
297
+ "model.layers.30.mlp.down_proj.weight": "model-00015-of-00037.safetensors",
298
+ "model.layers.30.mlp.gate_proj.weight": "model-00015-of-00037.safetensors",
299
+ "model.layers.30.mlp.up_proj.weight": "model-00015-of-00037.safetensors",
300
+ "model.layers.30.post_attention_layernorm.weight": "model-00015-of-00037.safetensors",
301
+ "model.layers.30.self_attn.k_proj.bias": "model-00015-of-00037.safetensors",
302
+ "model.layers.30.self_attn.k_proj.weight": "model-00015-of-00037.safetensors",
303
+ "model.layers.30.self_attn.o_proj.weight": "model-00015-of-00037.safetensors",
304
+ "model.layers.30.self_attn.q_proj.bias": "model-00015-of-00037.safetensors",
305
+ "model.layers.30.self_attn.q_proj.weight": "model-00015-of-00037.safetensors",
306
+ "model.layers.30.self_attn.v_proj.bias": "model-00015-of-00037.safetensors",
307
+ "model.layers.30.self_attn.v_proj.weight": "model-00015-of-00037.safetensors",
308
+ "model.layers.31.input_layernorm.weight": "model-00015-of-00037.safetensors",
309
+ "model.layers.31.mlp.down_proj.weight": "model-00015-of-00037.safetensors",
310
+ "model.layers.31.mlp.gate_proj.weight": "model-00015-of-00037.safetensors",
311
+ "model.layers.31.mlp.up_proj.weight": "model-00015-of-00037.safetensors",
312
+ "model.layers.31.post_attention_layernorm.weight": "model-00015-of-00037.safetensors",
313
+ "model.layers.31.self_attn.k_proj.bias": "model-00015-of-00037.safetensors",
314
+ "model.layers.31.self_attn.k_proj.weight": "model-00015-of-00037.safetensors",
315
+ "model.layers.31.self_attn.o_proj.weight": "model-00015-of-00037.safetensors",
316
+ "model.layers.31.self_attn.q_proj.bias": "model-00015-of-00037.safetensors",
317
+ "model.layers.31.self_attn.q_proj.weight": "model-00015-of-00037.safetensors",
318
+ "model.layers.31.self_attn.v_proj.bias": "model-00015-of-00037.safetensors",
319
+ "model.layers.31.self_attn.v_proj.weight": "model-00015-of-00037.safetensors",
320
+ "model.layers.32.input_layernorm.weight": "model-00015-of-00037.safetensors",
321
+ "model.layers.32.mlp.down_proj.weight": "model-00016-of-00037.safetensors",
322
+ "model.layers.32.mlp.gate_proj.weight": "model-00016-of-00037.safetensors",
323
+ "model.layers.32.mlp.up_proj.weight": "model-00016-of-00037.safetensors",
324
+ "model.layers.32.post_attention_layernorm.weight": "model-00015-of-00037.safetensors",
325
+ "model.layers.32.self_attn.k_proj.bias": "model-00015-of-00037.safetensors",
326
+ "model.layers.32.self_attn.k_proj.weight": "model-00015-of-00037.safetensors",
327
+ "model.layers.32.self_attn.o_proj.weight": "model-00015-of-00037.safetensors",
328
+ "model.layers.32.self_attn.q_proj.bias": "model-00015-of-00037.safetensors",
329
+ "model.layers.32.self_attn.q_proj.weight": "model-00015-of-00037.safetensors",
330
+ "model.layers.32.self_attn.v_proj.bias": "model-00015-of-00037.safetensors",
331
+ "model.layers.32.self_attn.v_proj.weight": "model-00015-of-00037.safetensors",
332
+ "model.layers.33.input_layernorm.weight": "model-00016-of-00037.safetensors",
333
+ "model.layers.33.mlp.down_proj.weight": "model-00016-of-00037.safetensors",
334
+ "model.layers.33.mlp.gate_proj.weight": "model-00016-of-00037.safetensors",
335
+ "model.layers.33.mlp.up_proj.weight": "model-00016-of-00037.safetensors",
336
+ "model.layers.33.post_attention_layernorm.weight": "model-00016-of-00037.safetensors",
337
+ "model.layers.33.self_attn.k_proj.bias": "model-00016-of-00037.safetensors",
338
+ "model.layers.33.self_attn.k_proj.weight": "model-00016-of-00037.safetensors",
339
+ "model.layers.33.self_attn.o_proj.weight": "model-00016-of-00037.safetensors",
340
+ "model.layers.33.self_attn.q_proj.bias": "model-00016-of-00037.safetensors",
341
+ "model.layers.33.self_attn.q_proj.weight": "model-00016-of-00037.safetensors",
342
+ "model.layers.33.self_attn.v_proj.bias": "model-00016-of-00037.safetensors",
343
+ "model.layers.33.self_attn.v_proj.weight": "model-00016-of-00037.safetensors",
344
+ "model.layers.34.input_layernorm.weight": "model-00016-of-00037.safetensors",
345
+ "model.layers.34.mlp.down_proj.weight": "model-00017-of-00037.safetensors",
346
+ "model.layers.34.mlp.gate_proj.weight": "model-00017-of-00037.safetensors",
347
+ "model.layers.34.mlp.up_proj.weight": "model-00016-of-00037.safetensors",
348
+ "model.layers.34.post_attention_layernorm.weight": "model-00016-of-00037.safetensors",
349
+ "model.layers.34.self_attn.k_proj.bias": "model-00016-of-00037.safetensors",
350
+ "model.layers.34.self_attn.k_proj.weight": "model-00016-of-00037.safetensors",
351
+ "model.layers.34.self_attn.o_proj.weight": "model-00016-of-00037.safetensors",
352
+ "model.layers.34.self_attn.q_proj.bias": "model-00016-of-00037.safetensors",
353
+ "model.layers.34.self_attn.q_proj.weight": "model-00016-of-00037.safetensors",
354
+ "model.layers.34.self_attn.v_proj.bias": "model-00016-of-00037.safetensors",
355
+ "model.layers.34.self_attn.v_proj.weight": "model-00016-of-00037.safetensors",
356
+ "model.layers.35.input_layernorm.weight": "model-00017-of-00037.safetensors",
357
+ "model.layers.35.mlp.down_proj.weight": "model-00017-of-00037.safetensors",
358
+ "model.layers.35.mlp.gate_proj.weight": "model-00017-of-00037.safetensors",
359
+ "model.layers.35.mlp.up_proj.weight": "model-00017-of-00037.safetensors",
360
+ "model.layers.35.post_attention_layernorm.weight": "model-00017-of-00037.safetensors",
361
+ "model.layers.35.self_attn.k_proj.bias": "model-00017-of-00037.safetensors",
362
+ "model.layers.35.self_attn.k_proj.weight": "model-00017-of-00037.safetensors",
363
+ "model.layers.35.self_attn.o_proj.weight": "model-00017-of-00037.safetensors",
364
+ "model.layers.35.self_attn.q_proj.bias": "model-00017-of-00037.safetensors",
365
+ "model.layers.35.self_attn.q_proj.weight": "model-00017-of-00037.safetensors",
366
+ "model.layers.35.self_attn.v_proj.bias": "model-00017-of-00037.safetensors",
367
+ "model.layers.35.self_attn.v_proj.weight": "model-00017-of-00037.safetensors",
368
+ "model.layers.36.input_layernorm.weight": "model-00017-of-00037.safetensors",
369
+ "model.layers.36.mlp.down_proj.weight": "model-00018-of-00037.safetensors",
370
+ "model.layers.36.mlp.gate_proj.weight": "model-00017-of-00037.safetensors",
371
+ "model.layers.36.mlp.up_proj.weight": "model-00017-of-00037.safetensors",
372
+ "model.layers.36.post_attention_layernorm.weight": "model-00017-of-00037.safetensors",
373
+ "model.layers.36.self_attn.k_proj.bias": "model-00017-of-00037.safetensors",
374
+ "model.layers.36.self_attn.k_proj.weight": "model-00017-of-00037.safetensors",
375
+ "model.layers.36.self_attn.o_proj.weight": "model-00017-of-00037.safetensors",
376
+ "model.layers.36.self_attn.q_proj.bias": "model-00017-of-00037.safetensors",
377
+ "model.layers.36.self_attn.q_proj.weight": "model-00017-of-00037.safetensors",
378
+ "model.layers.36.self_attn.v_proj.bias": "model-00017-of-00037.safetensors",
379
+ "model.layers.36.self_attn.v_proj.weight": "model-00017-of-00037.safetensors",
380
+ "model.layers.37.input_layernorm.weight": "model-00018-of-00037.safetensors",
381
+ "model.layers.37.mlp.down_proj.weight": "model-00018-of-00037.safetensors",
382
+ "model.layers.37.mlp.gate_proj.weight": "model-00018-of-00037.safetensors",
383
+ "model.layers.37.mlp.up_proj.weight": "model-00018-of-00037.safetensors",
384
+ "model.layers.37.post_attention_layernorm.weight": "model-00018-of-00037.safetensors",
385
+ "model.layers.37.self_attn.k_proj.bias": "model-00018-of-00037.safetensors",
386
+ "model.layers.37.self_attn.k_proj.weight": "model-00018-of-00037.safetensors",
387
+ "model.layers.37.self_attn.o_proj.weight": "model-00018-of-00037.safetensors",
388
+ "model.layers.37.self_attn.q_proj.bias": "model-00018-of-00037.safetensors",
389
+ "model.layers.37.self_attn.q_proj.weight": "model-00018-of-00037.safetensors",
390
+ "model.layers.37.self_attn.v_proj.bias": "model-00018-of-00037.safetensors",
391
+ "model.layers.37.self_attn.v_proj.weight": "model-00018-of-00037.safetensors",
392
+ "model.layers.38.input_layernorm.weight": "model-00018-of-00037.safetensors",
393
+ "model.layers.38.mlp.down_proj.weight": "model-00018-of-00037.safetensors",
394
+ "model.layers.38.mlp.gate_proj.weight": "model-00018-of-00037.safetensors",
395
+ "model.layers.38.mlp.up_proj.weight": "model-00018-of-00037.safetensors",
396
+ "model.layers.38.post_attention_layernorm.weight": "model-00018-of-00037.safetensors",
397
+ "model.layers.38.self_attn.k_proj.bias": "model-00018-of-00037.safetensors",
398
+ "model.layers.38.self_attn.k_proj.weight": "model-00018-of-00037.safetensors",
399
+ "model.layers.38.self_attn.o_proj.weight": "model-00018-of-00037.safetensors",
400
+ "model.layers.38.self_attn.q_proj.bias": "model-00018-of-00037.safetensors",
401
+ "model.layers.38.self_attn.q_proj.weight": "model-00018-of-00037.safetensors",
402
+ "model.layers.38.self_attn.v_proj.bias": "model-00018-of-00037.safetensors",
403
+ "model.layers.38.self_attn.v_proj.weight": "model-00018-of-00037.safetensors",
404
+ "model.layers.39.input_layernorm.weight": "model-00018-of-00037.safetensors",
405
+ "model.layers.39.mlp.down_proj.weight": "model-00019-of-00037.safetensors",
406
+ "model.layers.39.mlp.gate_proj.weight": "model-00019-of-00037.safetensors",
407
+ "model.layers.39.mlp.up_proj.weight": "model-00019-of-00037.safetensors",
408
+ "model.layers.39.post_attention_layernorm.weight": "model-00019-of-00037.safetensors",
409
+ "model.layers.39.self_attn.k_proj.bias": "model-00019-of-00037.safetensors",
410
+ "model.layers.39.self_attn.k_proj.weight": "model-00019-of-00037.safetensors",
411
+ "model.layers.39.self_attn.o_proj.weight": "model-00019-of-00037.safetensors",
412
+ "model.layers.39.self_attn.q_proj.bias": "model-00019-of-00037.safetensors",
413
+ "model.layers.39.self_attn.q_proj.weight": "model-00019-of-00037.safetensors",
414
+ "model.layers.39.self_attn.v_proj.bias": "model-00019-of-00037.safetensors",
415
+ "model.layers.39.self_attn.v_proj.weight": "model-00019-of-00037.safetensors",
416
+ "model.layers.4.input_layernorm.weight": "model-00003-of-00037.safetensors",
417
+ "model.layers.4.mlp.down_proj.weight": "model-00003-of-00037.safetensors",
418
+ "model.layers.4.mlp.gate_proj.weight": "model-00003-of-00037.safetensors",
419
+ "model.layers.4.mlp.up_proj.weight": "model-00003-of-00037.safetensors",
420
+ "model.layers.4.post_attention_layernorm.weight": "model-00003-of-00037.safetensors",
421
+ "model.layers.4.self_attn.k_proj.bias": "model-00003-of-00037.safetensors",
422
+ "model.layers.4.self_attn.k_proj.weight": "model-00003-of-00037.safetensors",
423
+ "model.layers.4.self_attn.o_proj.weight": "model-00003-of-00037.safetensors",
424
+ "model.layers.4.self_attn.q_proj.bias": "model-00003-of-00037.safetensors",
425
+ "model.layers.4.self_attn.q_proj.weight": "model-00003-of-00037.safetensors",
426
+ "model.layers.4.self_attn.v_proj.bias": "model-00003-of-00037.safetensors",
427
+ "model.layers.4.self_attn.v_proj.weight": "model-00003-of-00037.safetensors",
428
+ "model.layers.40.input_layernorm.weight": "model-00019-of-00037.safetensors",
429
+ "model.layers.40.mlp.down_proj.weight": "model-00019-of-00037.safetensors",
430
+ "model.layers.40.mlp.gate_proj.weight": "model-00019-of-00037.safetensors",
431
+ "model.layers.40.mlp.up_proj.weight": "model-00019-of-00037.safetensors",
432
+ "model.layers.40.post_attention_layernorm.weight": "model-00019-of-00037.safetensors",
433
+ "model.layers.40.self_attn.k_proj.bias": "model-00019-of-00037.safetensors",
434
+ "model.layers.40.self_attn.k_proj.weight": "model-00019-of-00037.safetensors",
435
+ "model.layers.40.self_attn.o_proj.weight": "model-00019-of-00037.safetensors",
436
+ "model.layers.40.self_attn.q_proj.bias": "model-00019-of-00037.safetensors",
437
+ "model.layers.40.self_attn.q_proj.weight": "model-00019-of-00037.safetensors",
438
+ "model.layers.40.self_attn.v_proj.bias": "model-00019-of-00037.safetensors",
439
+ "model.layers.40.self_attn.v_proj.weight": "model-00019-of-00037.safetensors",
440
+ "model.layers.41.input_layernorm.weight": "model-00019-of-00037.safetensors",
441
+ "model.layers.41.mlp.down_proj.weight": "model-00020-of-00037.safetensors",
442
+ "model.layers.41.mlp.gate_proj.weight": "model-00020-of-00037.safetensors",
443
+ "model.layers.41.mlp.up_proj.weight": "model-00020-of-00037.safetensors",
444
+ "model.layers.41.post_attention_layernorm.weight": "model-00019-of-00037.safetensors",
445
+ "model.layers.41.self_attn.k_proj.bias": "model-00019-of-00037.safetensors",
446
+ "model.layers.41.self_attn.k_proj.weight": "model-00019-of-00037.safetensors",
447
+ "model.layers.41.self_attn.o_proj.weight": "model-00019-of-00037.safetensors",
448
+ "model.layers.41.self_attn.q_proj.bias": "model-00019-of-00037.safetensors",
449
+ "model.layers.41.self_attn.q_proj.weight": "model-00019-of-00037.safetensors",
450
+ "model.layers.41.self_attn.v_proj.bias": "model-00019-of-00037.safetensors",
451
+ "model.layers.41.self_attn.v_proj.weight": "model-00019-of-00037.safetensors",
452
+ "model.layers.42.input_layernorm.weight": "model-00020-of-00037.safetensors",
453
+ "model.layers.42.mlp.down_proj.weight": "model-00020-of-00037.safetensors",
454
+ "model.layers.42.mlp.gate_proj.weight": "model-00020-of-00037.safetensors",
455
+ "model.layers.42.mlp.up_proj.weight": "model-00020-of-00037.safetensors",
456
+ "model.layers.42.post_attention_layernorm.weight": "model-00020-of-00037.safetensors",
457
+ "model.layers.42.self_attn.k_proj.bias": "model-00020-of-00037.safetensors",
458
+ "model.layers.42.self_attn.k_proj.weight": "model-00020-of-00037.safetensors",
459
+ "model.layers.42.self_attn.o_proj.weight": "model-00020-of-00037.safetensors",
460
+ "model.layers.42.self_attn.q_proj.bias": "model-00020-of-00037.safetensors",
461
+ "model.layers.42.self_attn.q_proj.weight": "model-00020-of-00037.safetensors",
462
+ "model.layers.42.self_attn.v_proj.bias": "model-00020-of-00037.safetensors",
463
+ "model.layers.42.self_attn.v_proj.weight": "model-00020-of-00037.safetensors",
464
+ "model.layers.43.input_layernorm.weight": "model-00020-of-00037.safetensors",
465
+ "model.layers.43.mlp.down_proj.weight": "model-00021-of-00037.safetensors",
466
+ "model.layers.43.mlp.gate_proj.weight": "model-00021-of-00037.safetensors",
467
+ "model.layers.43.mlp.up_proj.weight": "model-00020-of-00037.safetensors",
468
+ "model.layers.43.post_attention_layernorm.weight": "model-00020-of-00037.safetensors",
469
+ "model.layers.43.self_attn.k_proj.bias": "model-00020-of-00037.safetensors",
470
+ "model.layers.43.self_attn.k_proj.weight": "model-00020-of-00037.safetensors",
471
+ "model.layers.43.self_attn.o_proj.weight": "model-00020-of-00037.safetensors",
472
+ "model.layers.43.self_attn.q_proj.bias": "model-00020-of-00037.safetensors",
473
+ "model.layers.43.self_attn.q_proj.weight": "model-00020-of-00037.safetensors",
474
+ "model.layers.43.self_attn.v_proj.bias": "model-00020-of-00037.safetensors",
475
+ "model.layers.43.self_attn.v_proj.weight": "model-00020-of-00037.safetensors",
476
+ "model.layers.44.input_layernorm.weight": "model-00021-of-00037.safetensors",
477
+ "model.layers.44.mlp.down_proj.weight": "model-00021-of-00037.safetensors",
478
+ "model.layers.44.mlp.gate_proj.weight": "model-00021-of-00037.safetensors",
479
+ "model.layers.44.mlp.up_proj.weight": "model-00021-of-00037.safetensors",
480
+ "model.layers.44.post_attention_layernorm.weight": "model-00021-of-00037.safetensors",
481
+ "model.layers.44.self_attn.k_proj.bias": "model-00021-of-00037.safetensors",
482
+ "model.layers.44.self_attn.k_proj.weight": "model-00021-of-00037.safetensors",
483
+ "model.layers.44.self_attn.o_proj.weight": "model-00021-of-00037.safetensors",
484
+ "model.layers.44.self_attn.q_proj.bias": "model-00021-of-00037.safetensors",
485
+ "model.layers.44.self_attn.q_proj.weight": "model-00021-of-00037.safetensors",
486
+ "model.layers.44.self_attn.v_proj.bias": "model-00021-of-00037.safetensors",
487
+ "model.layers.44.self_attn.v_proj.weight": "model-00021-of-00037.safetensors",
488
+ "model.layers.45.input_layernorm.weight": "model-00021-of-00037.safetensors",
489
+ "model.layers.45.mlp.down_proj.weight": "model-00022-of-00037.safetensors",
490
+ "model.layers.45.mlp.gate_proj.weight": "model-00021-of-00037.safetensors",
491
+ "model.layers.45.mlp.up_proj.weight": "model-00021-of-00037.safetensors",
492
+ "model.layers.45.post_attention_layernorm.weight": "model-00021-of-00037.safetensors",
493
+ "model.layers.45.self_attn.k_proj.bias": "model-00021-of-00037.safetensors",
494
+ "model.layers.45.self_attn.k_proj.weight": "model-00021-of-00037.safetensors",
495
+ "model.layers.45.self_attn.o_proj.weight": "model-00021-of-00037.safetensors",
496
+ "model.layers.45.self_attn.q_proj.bias": "model-00021-of-00037.safetensors",
497
+ "model.layers.45.self_attn.q_proj.weight": "model-00021-of-00037.safetensors",
498
+ "model.layers.45.self_attn.v_proj.bias": "model-00021-of-00037.safetensors",
499
+ "model.layers.45.self_attn.v_proj.weight": "model-00021-of-00037.safetensors",
500
+ "model.layers.46.input_layernorm.weight": "model-00022-of-00037.safetensors",
501
+ "model.layers.46.mlp.down_proj.weight": "model-00022-of-00037.safetensors",
502
+ "model.layers.46.mlp.gate_proj.weight": "model-00022-of-00037.safetensors",
503
+ "model.layers.46.mlp.up_proj.weight": "model-00022-of-00037.safetensors",
504
+ "model.layers.46.post_attention_layernorm.weight": "model-00022-of-00037.safetensors",
505
+ "model.layers.46.self_attn.k_proj.bias": "model-00022-of-00037.safetensors",
506
+ "model.layers.46.self_attn.k_proj.weight": "model-00022-of-00037.safetensors",
507
+ "model.layers.46.self_attn.o_proj.weight": "model-00022-of-00037.safetensors",
508
+ "model.layers.46.self_attn.q_proj.bias": "model-00022-of-00037.safetensors",
509
+ "model.layers.46.self_attn.q_proj.weight": "model-00022-of-00037.safetensors",
510
+ "model.layers.46.self_attn.v_proj.bias": "model-00022-of-00037.safetensors",
511
+ "model.layers.46.self_attn.v_proj.weight": "model-00022-of-00037.safetensors",
512
+ "model.layers.47.input_layernorm.weight": "model-00022-of-00037.safetensors",
513
+ "model.layers.47.mlp.down_proj.weight": "model-00022-of-00037.safetensors",
514
+ "model.layers.47.mlp.gate_proj.weight": "model-00022-of-00037.safetensors",
515
+ "model.layers.47.mlp.up_proj.weight": "model-00022-of-00037.safetensors",
516
+ "model.layers.47.post_attention_layernorm.weight": "model-00022-of-00037.safetensors",
517
+ "model.layers.47.self_attn.k_proj.bias": "model-00022-of-00037.safetensors",
518
+ "model.layers.47.self_attn.k_proj.weight": "model-00022-of-00037.safetensors",
519
+ "model.layers.47.self_attn.o_proj.weight": "model-00022-of-00037.safetensors",
520
+ "model.layers.47.self_attn.q_proj.bias": "model-00022-of-00037.safetensors",
521
+ "model.layers.47.self_attn.q_proj.weight": "model-00022-of-00037.safetensors",
522
+ "model.layers.47.self_attn.v_proj.bias": "model-00022-of-00037.safetensors",
523
+ "model.layers.47.self_attn.v_proj.weight": "model-00022-of-00037.safetensors",
524
+ "model.layers.48.input_layernorm.weight": "model-00022-of-00037.safetensors",
525
+ "model.layers.48.mlp.down_proj.weight": "model-00023-of-00037.safetensors",
526
+ "model.layers.48.mlp.gate_proj.weight": "model-00023-of-00037.safetensors",
527
+ "model.layers.48.mlp.up_proj.weight": "model-00023-of-00037.safetensors",
528
+ "model.layers.48.post_attention_layernorm.weight": "model-00023-of-00037.safetensors",
529
+ "model.layers.48.self_attn.k_proj.bias": "model-00023-of-00037.safetensors",
530
+ "model.layers.48.self_attn.k_proj.weight": "model-00023-of-00037.safetensors",
531
+ "model.layers.48.self_attn.o_proj.weight": "model-00023-of-00037.safetensors",
532
+ "model.layers.48.self_attn.q_proj.bias": "model-00023-of-00037.safetensors",
533
+ "model.layers.48.self_attn.q_proj.weight": "model-00023-of-00037.safetensors",
534
+ "model.layers.48.self_attn.v_proj.bias": "model-00023-of-00037.safetensors",
535
+ "model.layers.48.self_attn.v_proj.weight": "model-00023-of-00037.safetensors",
536
+ "model.layers.49.input_layernorm.weight": "model-00023-of-00037.safetensors",
537
+ "model.layers.49.mlp.down_proj.weight": "model-00023-of-00037.safetensors",
538
+ "model.layers.49.mlp.gate_proj.weight": "model-00023-of-00037.safetensors",
539
+ "model.layers.49.mlp.up_proj.weight": "model-00023-of-00037.safetensors",
540
+ "model.layers.49.post_attention_layernorm.weight": "model-00023-of-00037.safetensors",
541
+ "model.layers.49.self_attn.k_proj.bias": "model-00023-of-00037.safetensors",
542
+ "model.layers.49.self_attn.k_proj.weight": "model-00023-of-00037.safetensors",
543
+ "model.layers.49.self_attn.o_proj.weight": "model-00023-of-00037.safetensors",
544
+ "model.layers.49.self_attn.q_proj.bias": "model-00023-of-00037.safetensors",
545
+ "model.layers.49.self_attn.q_proj.weight": "model-00023-of-00037.safetensors",
546
+ "model.layers.49.self_attn.v_proj.bias": "model-00023-of-00037.safetensors",
547
+ "model.layers.49.self_attn.v_proj.weight": "model-00023-of-00037.safetensors",
548
+ "model.layers.5.input_layernorm.weight": "model-00003-of-00037.safetensors",
549
+ "model.layers.5.mlp.down_proj.weight": "model-00004-of-00037.safetensors",
550
+ "model.layers.5.mlp.gate_proj.weight": "model-00004-of-00037.safetensors",
551
+ "model.layers.5.mlp.up_proj.weight": "model-00004-of-00037.safetensors",
552
+ "model.layers.5.post_attention_layernorm.weight": "model-00003-of-00037.safetensors",
553
+ "model.layers.5.self_attn.k_proj.bias": "model-00003-of-00037.safetensors",
554
+ "model.layers.5.self_attn.k_proj.weight": "model-00003-of-00037.safetensors",
555
+ "model.layers.5.self_attn.o_proj.weight": "model-00003-of-00037.safetensors",
556
+ "model.layers.5.self_attn.q_proj.bias": "model-00003-of-00037.safetensors",
557
+ "model.layers.5.self_attn.q_proj.weight": "model-00003-of-00037.safetensors",
558
+ "model.layers.5.self_attn.v_proj.bias": "model-00003-of-00037.safetensors",
559
+ "model.layers.5.self_attn.v_proj.weight": "model-00003-of-00037.safetensors",
560
+ "model.layers.50.input_layernorm.weight": "model-00023-of-00037.safetensors",
561
+ "model.layers.50.mlp.down_proj.weight": "model-00024-of-00037.safetensors",
562
+ "model.layers.50.mlp.gate_proj.weight": "model-00024-of-00037.safetensors",
563
+ "model.layers.50.mlp.up_proj.weight": "model-00024-of-00037.safetensors",
564
+ "model.layers.50.post_attention_layernorm.weight": "model-00023-of-00037.safetensors",
565
+ "model.layers.50.self_attn.k_proj.bias": "model-00023-of-00037.safetensors",
566
+ "model.layers.50.self_attn.k_proj.weight": "model-00023-of-00037.safetensors",
567
+ "model.layers.50.self_attn.o_proj.weight": "model-00023-of-00037.safetensors",
568
+ "model.layers.50.self_attn.q_proj.bias": "model-00023-of-00037.safetensors",
569
+ "model.layers.50.self_attn.q_proj.weight": "model-00023-of-00037.safetensors",
570
+ "model.layers.50.self_attn.v_proj.bias": "model-00023-of-00037.safetensors",
571
+ "model.layers.50.self_attn.v_proj.weight": "model-00023-of-00037.safetensors",
572
+ "model.layers.51.input_layernorm.weight": "model-00024-of-00037.safetensors",
573
+ "model.layers.51.mlp.down_proj.weight": "model-00024-of-00037.safetensors",
574
+ "model.layers.51.mlp.gate_proj.weight": "model-00024-of-00037.safetensors",
575
+ "model.layers.51.mlp.up_proj.weight": "model-00024-of-00037.safetensors",
576
+ "model.layers.51.post_attention_layernorm.weight": "model-00024-of-00037.safetensors",
577
+ "model.layers.51.self_attn.k_proj.bias": "model-00024-of-00037.safetensors",
578
+ "model.layers.51.self_attn.k_proj.weight": "model-00024-of-00037.safetensors",
579
+ "model.layers.51.self_attn.o_proj.weight": "model-00024-of-00037.safetensors",
580
+ "model.layers.51.self_attn.q_proj.bias": "model-00024-of-00037.safetensors",
581
+ "model.layers.51.self_attn.q_proj.weight": "model-00024-of-00037.safetensors",
582
+ "model.layers.51.self_attn.v_proj.bias": "model-00024-of-00037.safetensors",
583
+ "model.layers.51.self_attn.v_proj.weight": "model-00024-of-00037.safetensors",
584
+ "model.layers.52.input_layernorm.weight": "model-00024-of-00037.safetensors",
585
+ "model.layers.52.mlp.down_proj.weight": "model-00025-of-00037.safetensors",
586
+ "model.layers.52.mlp.gate_proj.weight": "model-00025-of-00037.safetensors",
587
+ "model.layers.52.mlp.up_proj.weight": "model-00024-of-00037.safetensors",
588
+ "model.layers.52.post_attention_layernorm.weight": "model-00024-of-00037.safetensors",
589
+ "model.layers.52.self_attn.k_proj.bias": "model-00024-of-00037.safetensors",
590
+ "model.layers.52.self_attn.k_proj.weight": "model-00024-of-00037.safetensors",
591
+ "model.layers.52.self_attn.o_proj.weight": "model-00024-of-00037.safetensors",
592
+ "model.layers.52.self_attn.q_proj.bias": "model-00024-of-00037.safetensors",
593
+ "model.layers.52.self_attn.q_proj.weight": "model-00024-of-00037.safetensors",
594
+ "model.layers.52.self_attn.v_proj.bias": "model-00024-of-00037.safetensors",
595
+ "model.layers.52.self_attn.v_proj.weight": "model-00024-of-00037.safetensors",
596
+ "model.layers.53.input_layernorm.weight": "model-00025-of-00037.safetensors",
597
+ "model.layers.53.mlp.down_proj.weight": "model-00025-of-00037.safetensors",
598
+ "model.layers.53.mlp.gate_proj.weight": "model-00025-of-00037.safetensors",
599
+ "model.layers.53.mlp.up_proj.weight": "model-00025-of-00037.safetensors",
600
+ "model.layers.53.post_attention_layernorm.weight": "model-00025-of-00037.safetensors",
601
+ "model.layers.53.self_attn.k_proj.bias": "model-00025-of-00037.safetensors",
602
+ "model.layers.53.self_attn.k_proj.weight": "model-00025-of-00037.safetensors",
603
+ "model.layers.53.self_attn.o_proj.weight": "model-00025-of-00037.safetensors",
604
+ "model.layers.53.self_attn.q_proj.bias": "model-00025-of-00037.safetensors",
605
+ "model.layers.53.self_attn.q_proj.weight": "model-00025-of-00037.safetensors",
606
+ "model.layers.53.self_attn.v_proj.bias": "model-00025-of-00037.safetensors",
607
+ "model.layers.53.self_attn.v_proj.weight": "model-00025-of-00037.safetensors",
608
+ "model.layers.54.input_layernorm.weight": "model-00025-of-00037.safetensors",
609
+ "model.layers.54.mlp.down_proj.weight": "model-00026-of-00037.safetensors",
610
+ "model.layers.54.mlp.gate_proj.weight": "model-00025-of-00037.safetensors",
611
+ "model.layers.54.mlp.up_proj.weight": "model-00025-of-00037.safetensors",
612
+ "model.layers.54.post_attention_layernorm.weight": "model-00025-of-00037.safetensors",
613
+ "model.layers.54.self_attn.k_proj.bias": "model-00025-of-00037.safetensors",
614
+ "model.layers.54.self_attn.k_proj.weight": "model-00025-of-00037.safetensors",
615
+ "model.layers.54.self_attn.o_proj.weight": "model-00025-of-00037.safetensors",
616
+ "model.layers.54.self_attn.q_proj.bias": "model-00025-of-00037.safetensors",
617
+ "model.layers.54.self_attn.q_proj.weight": "model-00025-of-00037.safetensors",
618
+ "model.layers.54.self_attn.v_proj.bias": "model-00025-of-00037.safetensors",
619
+ "model.layers.54.self_attn.v_proj.weight": "model-00025-of-00037.safetensors",
620
+ "model.layers.55.input_layernorm.weight": "model-00026-of-00037.safetensors",
621
+ "model.layers.55.mlp.down_proj.weight": "model-00026-of-00037.safetensors",
622
+ "model.layers.55.mlp.gate_proj.weight": "model-00026-of-00037.safetensors",
623
+ "model.layers.55.mlp.up_proj.weight": "model-00026-of-00037.safetensors",
624
+ "model.layers.55.post_attention_layernorm.weight": "model-00026-of-00037.safetensors",
625
+ "model.layers.55.self_attn.k_proj.bias": "model-00026-of-00037.safetensors",
626
+ "model.layers.55.self_attn.k_proj.weight": "model-00026-of-00037.safetensors",
627
+ "model.layers.55.self_attn.o_proj.weight": "model-00026-of-00037.safetensors",
628
+ "model.layers.55.self_attn.q_proj.bias": "model-00026-of-00037.safetensors",
629
+ "model.layers.55.self_attn.q_proj.weight": "model-00026-of-00037.safetensors",
630
+ "model.layers.55.self_attn.v_proj.bias": "model-00026-of-00037.safetensors",
631
+ "model.layers.55.self_attn.v_proj.weight": "model-00026-of-00037.safetensors",
632
+ "model.layers.56.input_layernorm.weight": "model-00026-of-00037.safetensors",
633
+ "model.layers.56.mlp.down_proj.weight": "model-00026-of-00037.safetensors",
634
+ "model.layers.56.mlp.gate_proj.weight": "model-00026-of-00037.safetensors",
635
+ "model.layers.56.mlp.up_proj.weight": "model-00026-of-00037.safetensors",
636
+ "model.layers.56.post_attention_layernorm.weight": "model-00026-of-00037.safetensors",
637
+ "model.layers.56.self_attn.k_proj.bias": "model-00026-of-00037.safetensors",
638
+ "model.layers.56.self_attn.k_proj.weight": "model-00026-of-00037.safetensors",
639
+ "model.layers.56.self_attn.o_proj.weight": "model-00026-of-00037.safetensors",
640
+ "model.layers.56.self_attn.q_proj.bias": "model-00026-of-00037.safetensors",
641
+ "model.layers.56.self_attn.q_proj.weight": "model-00026-of-00037.safetensors",
642
+ "model.layers.56.self_attn.v_proj.bias": "model-00026-of-00037.safetensors",
643
+ "model.layers.56.self_attn.v_proj.weight": "model-00026-of-00037.safetensors",
644
+ "model.layers.57.input_layernorm.weight": "model-00026-of-00037.safetensors",
645
+ "model.layers.57.mlp.down_proj.weight": "model-00027-of-00037.safetensors",
646
+ "model.layers.57.mlp.gate_proj.weight": "model-00027-of-00037.safetensors",
647
+ "model.layers.57.mlp.up_proj.weight": "model-00027-of-00037.safetensors",
648
+ "model.layers.57.post_attention_layernorm.weight": "model-00027-of-00037.safetensors",
649
+ "model.layers.57.self_attn.k_proj.bias": "model-00027-of-00037.safetensors",
650
+ "model.layers.57.self_attn.k_proj.weight": "model-00027-of-00037.safetensors",
651
+ "model.layers.57.self_attn.o_proj.weight": "model-00027-of-00037.safetensors",
652
+ "model.layers.57.self_attn.q_proj.bias": "model-00027-of-00037.safetensors",
653
+ "model.layers.57.self_attn.q_proj.weight": "model-00027-of-00037.safetensors",
654
+ "model.layers.57.self_attn.v_proj.bias": "model-00027-of-00037.safetensors",
655
+ "model.layers.57.self_attn.v_proj.weight": "model-00027-of-00037.safetensors",
656
+ "model.layers.58.input_layernorm.weight": "model-00027-of-00037.safetensors",
657
+ "model.layers.58.mlp.down_proj.weight": "model-00027-of-00037.safetensors",
658
+ "model.layers.58.mlp.gate_proj.weight": "model-00027-of-00037.safetensors",
659
+ "model.layers.58.mlp.up_proj.weight": "model-00027-of-00037.safetensors",
660
+ "model.layers.58.post_attention_layernorm.weight": "model-00027-of-00037.safetensors",
661
+ "model.layers.58.self_attn.k_proj.bias": "model-00027-of-00037.safetensors",
662
+ "model.layers.58.self_attn.k_proj.weight": "model-00027-of-00037.safetensors",
663
+ "model.layers.58.self_attn.o_proj.weight": "model-00027-of-00037.safetensors",
664
+ "model.layers.58.self_attn.q_proj.bias": "model-00027-of-00037.safetensors",
665
+ "model.layers.58.self_attn.q_proj.weight": "model-00027-of-00037.safetensors",
666
+ "model.layers.58.self_attn.v_proj.bias": "model-00027-of-00037.safetensors",
667
+ "model.layers.58.self_attn.v_proj.weight": "model-00027-of-00037.safetensors",
668
+ "model.layers.59.input_layernorm.weight": "model-00027-of-00037.safetensors",
669
+ "model.layers.59.mlp.down_proj.weight": "model-00028-of-00037.safetensors",
670
+ "model.layers.59.mlp.gate_proj.weight": "model-00028-of-00037.safetensors",
671
+ "model.layers.59.mlp.up_proj.weight": "model-00028-of-00037.safetensors",
672
+ "model.layers.59.post_attention_layernorm.weight": "model-00027-of-00037.safetensors",
673
+ "model.layers.59.self_attn.k_proj.bias": "model-00027-of-00037.safetensors",
674
+ "model.layers.59.self_attn.k_proj.weight": "model-00027-of-00037.safetensors",
675
+ "model.layers.59.self_attn.o_proj.weight": "model-00027-of-00037.safetensors",
676
+ "model.layers.59.self_attn.q_proj.bias": "model-00027-of-00037.safetensors",
677
+ "model.layers.59.self_attn.q_proj.weight": "model-00027-of-00037.safetensors",
678
+ "model.layers.59.self_attn.v_proj.bias": "model-00027-of-00037.safetensors",
679
+ "model.layers.59.self_attn.v_proj.weight": "model-00027-of-00037.safetensors",
680
+ "model.layers.6.input_layernorm.weight": "model-00004-of-00037.safetensors",
681
+ "model.layers.6.mlp.down_proj.weight": "model-00004-of-00037.safetensors",
682
+ "model.layers.6.mlp.gate_proj.weight": "model-00004-of-00037.safetensors",
683
+ "model.layers.6.mlp.up_proj.weight": "model-00004-of-00037.safetensors",
684
+ "model.layers.6.post_attention_layernorm.weight": "model-00004-of-00037.safetensors",
685
+ "model.layers.6.self_attn.k_proj.bias": "model-00004-of-00037.safetensors",
686
+ "model.layers.6.self_attn.k_proj.weight": "model-00004-of-00037.safetensors",
687
+ "model.layers.6.self_attn.o_proj.weight": "model-00004-of-00037.safetensors",
688
+ "model.layers.6.self_attn.q_proj.bias": "model-00004-of-00037.safetensors",
689
+ "model.layers.6.self_attn.q_proj.weight": "model-00004-of-00037.safetensors",
690
+ "model.layers.6.self_attn.v_proj.bias": "model-00004-of-00037.safetensors",
691
+ "model.layers.6.self_attn.v_proj.weight": "model-00004-of-00037.safetensors",
692
+ "model.layers.60.input_layernorm.weight": "model-00028-of-00037.safetensors",
693
+ "model.layers.60.mlp.down_proj.weight": "model-00028-of-00037.safetensors",
694
+ "model.layers.60.mlp.gate_proj.weight": "model-00028-of-00037.safetensors",
695
+ "model.layers.60.mlp.up_proj.weight": "model-00028-of-00037.safetensors",
696
+ "model.layers.60.post_attention_layernorm.weight": "model-00028-of-00037.safetensors",
697
+ "model.layers.60.self_attn.k_proj.bias": "model-00028-of-00037.safetensors",
698
+ "model.layers.60.self_attn.k_proj.weight": "model-00028-of-00037.safetensors",
699
+ "model.layers.60.self_attn.o_proj.weight": "model-00028-of-00037.safetensors",
700
+ "model.layers.60.self_attn.q_proj.bias": "model-00028-of-00037.safetensors",
701
+ "model.layers.60.self_attn.q_proj.weight": "model-00028-of-00037.safetensors",
702
+ "model.layers.60.self_attn.v_proj.bias": "model-00028-of-00037.safetensors",
703
+ "model.layers.60.self_attn.v_proj.weight": "model-00028-of-00037.safetensors",
704
+ "model.layers.61.input_layernorm.weight": "model-00028-of-00037.safetensors",
705
+ "model.layers.61.mlp.down_proj.weight": "model-00029-of-00037.safetensors",
706
+ "model.layers.61.mlp.gate_proj.weight": "model-00029-of-00037.safetensors",
707
+ "model.layers.61.mlp.up_proj.weight": "model-00028-of-00037.safetensors",
708
+ "model.layers.61.post_attention_layernorm.weight": "model-00028-of-00037.safetensors",
709
+ "model.layers.61.self_attn.k_proj.bias": "model-00028-of-00037.safetensors",
710
+ "model.layers.61.self_attn.k_proj.weight": "model-00028-of-00037.safetensors",
711
+ "model.layers.61.self_attn.o_proj.weight": "model-00028-of-00037.safetensors",
712
+ "model.layers.61.self_attn.q_proj.bias": "model-00028-of-00037.safetensors",
713
+ "model.layers.61.self_attn.q_proj.weight": "model-00028-of-00037.safetensors",
714
+ "model.layers.61.self_attn.v_proj.bias": "model-00028-of-00037.safetensors",
715
+ "model.layers.61.self_attn.v_proj.weight": "model-00028-of-00037.safetensors",
716
+ "model.layers.62.input_layernorm.weight": "model-00029-of-00037.safetensors",
717
+ "model.layers.62.mlp.down_proj.weight": "model-00029-of-00037.safetensors",
718
+ "model.layers.62.mlp.gate_proj.weight": "model-00029-of-00037.safetensors",
719
+ "model.layers.62.mlp.up_proj.weight": "model-00029-of-00037.safetensors",
720
+ "model.layers.62.post_attention_layernorm.weight": "model-00029-of-00037.safetensors",
721
+ "model.layers.62.self_attn.k_proj.bias": "model-00029-of-00037.safetensors",
722
+ "model.layers.62.self_attn.k_proj.weight": "model-00029-of-00037.safetensors",
723
+ "model.layers.62.self_attn.o_proj.weight": "model-00029-of-00037.safetensors",
724
+ "model.layers.62.self_attn.q_proj.bias": "model-00029-of-00037.safetensors",
725
+ "model.layers.62.self_attn.q_proj.weight": "model-00029-of-00037.safetensors",
726
+ "model.layers.62.self_attn.v_proj.bias": "model-00029-of-00037.safetensors",
727
+ "model.layers.62.self_attn.v_proj.weight": "model-00029-of-00037.safetensors",
728
+ "model.layers.63.input_layernorm.weight": "model-00029-of-00037.safetensors",
729
+ "model.layers.63.mlp.down_proj.weight": "model-00030-of-00037.safetensors",
730
+ "model.layers.63.mlp.gate_proj.weight": "model-00029-of-00037.safetensors",
731
+ "model.layers.63.mlp.up_proj.weight": "model-00029-of-00037.safetensors",
732
+ "model.layers.63.post_attention_layernorm.weight": "model-00029-of-00037.safetensors",
733
+ "model.layers.63.self_attn.k_proj.bias": "model-00029-of-00037.safetensors",
734
+ "model.layers.63.self_attn.k_proj.weight": "model-00029-of-00037.safetensors",
735
+ "model.layers.63.self_attn.o_proj.weight": "model-00029-of-00037.safetensors",
736
+ "model.layers.63.self_attn.q_proj.bias": "model-00029-of-00037.safetensors",
737
+ "model.layers.63.self_attn.q_proj.weight": "model-00029-of-00037.safetensors",
738
+ "model.layers.63.self_attn.v_proj.bias": "model-00029-of-00037.safetensors",
739
+ "model.layers.63.self_attn.v_proj.weight": "model-00029-of-00037.safetensors",
740
+ "model.layers.64.input_layernorm.weight": "model-00030-of-00037.safetensors",
741
+ "model.layers.64.mlp.down_proj.weight": "model-00030-of-00037.safetensors",
742
+ "model.layers.64.mlp.gate_proj.weight": "model-00030-of-00037.safetensors",
743
+ "model.layers.64.mlp.up_proj.weight": "model-00030-of-00037.safetensors",
744
+ "model.layers.64.post_attention_layernorm.weight": "model-00030-of-00037.safetensors",
745
+ "model.layers.64.self_attn.k_proj.bias": "model-00030-of-00037.safetensors",
746
+ "model.layers.64.self_attn.k_proj.weight": "model-00030-of-00037.safetensors",
747
+ "model.layers.64.self_attn.o_proj.weight": "model-00030-of-00037.safetensors",
748
+ "model.layers.64.self_attn.q_proj.bias": "model-00030-of-00037.safetensors",
749
+ "model.layers.64.self_attn.q_proj.weight": "model-00030-of-00037.safetensors",
750
+ "model.layers.64.self_attn.v_proj.bias": "model-00030-of-00037.safetensors",
751
+ "model.layers.64.self_attn.v_proj.weight": "model-00030-of-00037.safetensors",
752
+ "model.layers.65.input_layernorm.weight": "model-00030-of-00037.safetensors",
753
+ "model.layers.65.mlp.down_proj.weight": "model-00030-of-00037.safetensors",
754
+ "model.layers.65.mlp.gate_proj.weight": "model-00030-of-00037.safetensors",
755
+ "model.layers.65.mlp.up_proj.weight": "model-00030-of-00037.safetensors",
756
+ "model.layers.65.post_attention_layernorm.weight": "model-00030-of-00037.safetensors",
757
+ "model.layers.65.self_attn.k_proj.bias": "model-00030-of-00037.safetensors",
758
+ "model.layers.65.self_attn.k_proj.weight": "model-00030-of-00037.safetensors",
759
+ "model.layers.65.self_attn.o_proj.weight": "model-00030-of-00037.safetensors",
760
+ "model.layers.65.self_attn.q_proj.bias": "model-00030-of-00037.safetensors",
761
+ "model.layers.65.self_attn.q_proj.weight": "model-00030-of-00037.safetensors",
762
+ "model.layers.65.self_attn.v_proj.bias": "model-00030-of-00037.safetensors",
763
+ "model.layers.65.self_attn.v_proj.weight": "model-00030-of-00037.safetensors",
764
+ "model.layers.66.input_layernorm.weight": "model-00030-of-00037.safetensors",
765
+ "model.layers.66.mlp.down_proj.weight": "model-00031-of-00037.safetensors",
766
+ "model.layers.66.mlp.gate_proj.weight": "model-00031-of-00037.safetensors",
767
+ "model.layers.66.mlp.up_proj.weight": "model-00031-of-00037.safetensors",
768
+ "model.layers.66.post_attention_layernorm.weight": "model-00031-of-00037.safetensors",
769
+ "model.layers.66.self_attn.k_proj.bias": "model-00031-of-00037.safetensors",
770
+ "model.layers.66.self_attn.k_proj.weight": "model-00031-of-00037.safetensors",
771
+ "model.layers.66.self_attn.o_proj.weight": "model-00031-of-00037.safetensors",
772
+ "model.layers.66.self_attn.q_proj.bias": "model-00031-of-00037.safetensors",
773
+ "model.layers.66.self_attn.q_proj.weight": "model-00031-of-00037.safetensors",
774
+ "model.layers.66.self_attn.v_proj.bias": "model-00031-of-00037.safetensors",
775
+ "model.layers.66.self_attn.v_proj.weight": "model-00031-of-00037.safetensors",
776
+ "model.layers.67.input_layernorm.weight": "model-00031-of-00037.safetensors",
777
+ "model.layers.67.mlp.down_proj.weight": "model-00031-of-00037.safetensors",
778
+ "model.layers.67.mlp.gate_proj.weight": "model-00031-of-00037.safetensors",
779
+ "model.layers.67.mlp.up_proj.weight": "model-00031-of-00037.safetensors",
780
+ "model.layers.67.post_attention_layernorm.weight": "model-00031-of-00037.safetensors",
781
+ "model.layers.67.self_attn.k_proj.bias": "model-00031-of-00037.safetensors",
782
+ "model.layers.67.self_attn.k_proj.weight": "model-00031-of-00037.safetensors",
783
+ "model.layers.67.self_attn.o_proj.weight": "model-00031-of-00037.safetensors",
784
+ "model.layers.67.self_attn.q_proj.bias": "model-00031-of-00037.safetensors",
785
+ "model.layers.67.self_attn.q_proj.weight": "model-00031-of-00037.safetensors",
786
+ "model.layers.67.self_attn.v_proj.bias": "model-00031-of-00037.safetensors",
787
+ "model.layers.67.self_attn.v_proj.weight": "model-00031-of-00037.safetensors",
788
+ "model.layers.68.input_layernorm.weight": "model-00031-of-00037.safetensors",
789
+ "model.layers.68.mlp.down_proj.weight": "model-00032-of-00037.safetensors",
790
+ "model.layers.68.mlp.gate_proj.weight": "model-00032-of-00037.safetensors",
791
+ "model.layers.68.mlp.up_proj.weight": "model-00032-of-00037.safetensors",
792
+ "model.layers.68.post_attention_layernorm.weight": "model-00031-of-00037.safetensors",
793
+ "model.layers.68.self_attn.k_proj.bias": "model-00031-of-00037.safetensors",
794
+ "model.layers.68.self_attn.k_proj.weight": "model-00031-of-00037.safetensors",
795
+ "model.layers.68.self_attn.o_proj.weight": "model-00031-of-00037.safetensors",
796
+ "model.layers.68.self_attn.q_proj.bias": "model-00031-of-00037.safetensors",
797
+ "model.layers.68.self_attn.q_proj.weight": "model-00031-of-00037.safetensors",
798
+ "model.layers.68.self_attn.v_proj.bias": "model-00031-of-00037.safetensors",
799
+ "model.layers.68.self_attn.v_proj.weight": "model-00031-of-00037.safetensors",
800
+ "model.layers.69.input_layernorm.weight": "model-00032-of-00037.safetensors",
801
+ "model.layers.69.mlp.down_proj.weight": "model-00032-of-00037.safetensors",
802
+ "model.layers.69.mlp.gate_proj.weight": "model-00032-of-00037.safetensors",
803
+ "model.layers.69.mlp.up_proj.weight": "model-00032-of-00037.safetensors",
804
+ "model.layers.69.post_attention_layernorm.weight": "model-00032-of-00037.safetensors",
805
+ "model.layers.69.self_attn.k_proj.bias": "model-00032-of-00037.safetensors",
806
+ "model.layers.69.self_attn.k_proj.weight": "model-00032-of-00037.safetensors",
807
+ "model.layers.69.self_attn.o_proj.weight": "model-00032-of-00037.safetensors",
808
+ "model.layers.69.self_attn.q_proj.bias": "model-00032-of-00037.safetensors",
809
+ "model.layers.69.self_attn.q_proj.weight": "model-00032-of-00037.safetensors",
810
+ "model.layers.69.self_attn.v_proj.bias": "model-00032-of-00037.safetensors",
811
+ "model.layers.69.self_attn.v_proj.weight": "model-00032-of-00037.safetensors",
812
+ "model.layers.7.input_layernorm.weight": "model-00004-of-00037.safetensors",
813
+ "model.layers.7.mlp.down_proj.weight": "model-00005-of-00037.safetensors",
814
+ "model.layers.7.mlp.gate_proj.weight": "model-00005-of-00037.safetensors",
815
+ "model.layers.7.mlp.up_proj.weight": "model-00004-of-00037.safetensors",
816
+ "model.layers.7.post_attention_layernorm.weight": "model-00004-of-00037.safetensors",
817
+ "model.layers.7.self_attn.k_proj.bias": "model-00004-of-00037.safetensors",
818
+ "model.layers.7.self_attn.k_proj.weight": "model-00004-of-00037.safetensors",
819
+ "model.layers.7.self_attn.o_proj.weight": "model-00004-of-00037.safetensors",
820
+ "model.layers.7.self_attn.q_proj.bias": "model-00004-of-00037.safetensors",
821
+ "model.layers.7.self_attn.q_proj.weight": "model-00004-of-00037.safetensors",
822
+ "model.layers.7.self_attn.v_proj.bias": "model-00004-of-00037.safetensors",
823
+ "model.layers.7.self_attn.v_proj.weight": "model-00004-of-00037.safetensors",
824
+ "model.layers.70.input_layernorm.weight": "model-00032-of-00037.safetensors",
825
+ "model.layers.70.mlp.down_proj.weight": "model-00033-of-00037.safetensors",
826
+ "model.layers.70.mlp.gate_proj.weight": "model-00033-of-00037.safetensors",
827
+ "model.layers.70.mlp.up_proj.weight": "model-00032-of-00037.safetensors",
828
+ "model.layers.70.post_attention_layernorm.weight": "model-00032-of-00037.safetensors",
829
+ "model.layers.70.self_attn.k_proj.bias": "model-00032-of-00037.safetensors",
830
+ "model.layers.70.self_attn.k_proj.weight": "model-00032-of-00037.safetensors",
831
+ "model.layers.70.self_attn.o_proj.weight": "model-00032-of-00037.safetensors",
832
+ "model.layers.70.self_attn.q_proj.bias": "model-00032-of-00037.safetensors",
833
+ "model.layers.70.self_attn.q_proj.weight": "model-00032-of-00037.safetensors",
834
+ "model.layers.70.self_attn.v_proj.bias": "model-00032-of-00037.safetensors",
835
+ "model.layers.70.self_attn.v_proj.weight": "model-00032-of-00037.safetensors",
836
+ "model.layers.71.input_layernorm.weight": "model-00033-of-00037.safetensors",
837
+ "model.layers.71.mlp.down_proj.weight": "model-00033-of-00037.safetensors",
838
+ "model.layers.71.mlp.gate_proj.weight": "model-00033-of-00037.safetensors",
839
+ "model.layers.71.mlp.up_proj.weight": "model-00033-of-00037.safetensors",
840
+ "model.layers.71.post_attention_layernorm.weight": "model-00033-of-00037.safetensors",
841
+ "model.layers.71.self_attn.k_proj.bias": "model-00033-of-00037.safetensors",
842
+ "model.layers.71.self_attn.k_proj.weight": "model-00033-of-00037.safetensors",
843
+ "model.layers.71.self_attn.o_proj.weight": "model-00033-of-00037.safetensors",
844
+ "model.layers.71.self_attn.q_proj.bias": "model-00033-of-00037.safetensors",
845
+ "model.layers.71.self_attn.q_proj.weight": "model-00033-of-00037.safetensors",
846
+ "model.layers.71.self_attn.v_proj.bias": "model-00033-of-00037.safetensors",
847
+ "model.layers.71.self_attn.v_proj.weight": "model-00033-of-00037.safetensors",
848
+ "model.layers.72.input_layernorm.weight": "model-00033-of-00037.safetensors",
849
+ "model.layers.72.mlp.down_proj.weight": "model-00034-of-00037.safetensors",
850
+ "model.layers.72.mlp.gate_proj.weight": "model-00033-of-00037.safetensors",
851
+ "model.layers.72.mlp.up_proj.weight": "model-00033-of-00037.safetensors",
852
+ "model.layers.72.post_attention_layernorm.weight": "model-00033-of-00037.safetensors",
853
+ "model.layers.72.self_attn.k_proj.bias": "model-00033-of-00037.safetensors",
854
+ "model.layers.72.self_attn.k_proj.weight": "model-00033-of-00037.safetensors",
855
+ "model.layers.72.self_attn.o_proj.weight": "model-00033-of-00037.safetensors",
856
+ "model.layers.72.self_attn.q_proj.bias": "model-00033-of-00037.safetensors",
857
+ "model.layers.72.self_attn.q_proj.weight": "model-00033-of-00037.safetensors",
858
+ "model.layers.72.self_attn.v_proj.bias": "model-00033-of-00037.safetensors",
859
+ "model.layers.72.self_attn.v_proj.weight": "model-00033-of-00037.safetensors",
860
+ "model.layers.73.input_layernorm.weight": "model-00034-of-00037.safetensors",
861
+ "model.layers.73.mlp.down_proj.weight": "model-00034-of-00037.safetensors",
862
+ "model.layers.73.mlp.gate_proj.weight": "model-00034-of-00037.safetensors",
863
+ "model.layers.73.mlp.up_proj.weight": "model-00034-of-00037.safetensors",
864
+ "model.layers.73.post_attention_layernorm.weight": "model-00034-of-00037.safetensors",
865
+ "model.layers.73.self_attn.k_proj.bias": "model-00034-of-00037.safetensors",
866
+ "model.layers.73.self_attn.k_proj.weight": "model-00034-of-00037.safetensors",
867
+ "model.layers.73.self_attn.o_proj.weight": "model-00034-of-00037.safetensors",
868
+ "model.layers.73.self_attn.q_proj.bias": "model-00034-of-00037.safetensors",
869
+ "model.layers.73.self_attn.q_proj.weight": "model-00034-of-00037.safetensors",
870
+ "model.layers.73.self_attn.v_proj.bias": "model-00034-of-00037.safetensors",
871
+ "model.layers.73.self_attn.v_proj.weight": "model-00034-of-00037.safetensors",
872
+ "model.layers.74.input_layernorm.weight": "model-00034-of-00037.safetensors",
873
+ "model.layers.74.mlp.down_proj.weight": "model-00034-of-00037.safetensors",
874
+ "model.layers.74.mlp.gate_proj.weight": "model-00034-of-00037.safetensors",
875
+ "model.layers.74.mlp.up_proj.weight": "model-00034-of-00037.safetensors",
876
+ "model.layers.74.post_attention_layernorm.weight": "model-00034-of-00037.safetensors",
877
+ "model.layers.74.self_attn.k_proj.bias": "model-00034-of-00037.safetensors",
878
+ "model.layers.74.self_attn.k_proj.weight": "model-00034-of-00037.safetensors",
879
+ "model.layers.74.self_attn.o_proj.weight": "model-00034-of-00037.safetensors",
880
+ "model.layers.74.self_attn.q_proj.bias": "model-00034-of-00037.safetensors",
881
+ "model.layers.74.self_attn.q_proj.weight": "model-00034-of-00037.safetensors",
882
+ "model.layers.74.self_attn.v_proj.bias": "model-00034-of-00037.safetensors",
883
+ "model.layers.74.self_attn.v_proj.weight": "model-00034-of-00037.safetensors",
884
+ "model.layers.75.input_layernorm.weight": "model-00034-of-00037.safetensors",
885
+ "model.layers.75.mlp.down_proj.weight": "model-00035-of-00037.safetensors",
886
+ "model.layers.75.mlp.gate_proj.weight": "model-00035-of-00037.safetensors",
887
+ "model.layers.75.mlp.up_proj.weight": "model-00035-of-00037.safetensors",
888
+ "model.layers.75.post_attention_layernorm.weight": "model-00035-of-00037.safetensors",
889
+ "model.layers.75.self_attn.k_proj.bias": "model-00035-of-00037.safetensors",
890
+ "model.layers.75.self_attn.k_proj.weight": "model-00035-of-00037.safetensors",
891
+ "model.layers.75.self_attn.o_proj.weight": "model-00035-of-00037.safetensors",
892
+ "model.layers.75.self_attn.q_proj.bias": "model-00035-of-00037.safetensors",
893
+ "model.layers.75.self_attn.q_proj.weight": "model-00035-of-00037.safetensors",
894
+ "model.layers.75.self_attn.v_proj.bias": "model-00035-of-00037.safetensors",
895
+ "model.layers.75.self_attn.v_proj.weight": "model-00035-of-00037.safetensors",
896
+ "model.layers.76.input_layernorm.weight": "model-00035-of-00037.safetensors",
897
+ "model.layers.76.mlp.down_proj.weight": "model-00035-of-00037.safetensors",
898
+ "model.layers.76.mlp.gate_proj.weight": "model-00035-of-00037.safetensors",
899
+ "model.layers.76.mlp.up_proj.weight": "model-00035-of-00037.safetensors",
900
+ "model.layers.76.post_attention_layernorm.weight": "model-00035-of-00037.safetensors",
901
+ "model.layers.76.self_attn.k_proj.bias": "model-00035-of-00037.safetensors",
902
+ "model.layers.76.self_attn.k_proj.weight": "model-00035-of-00037.safetensors",
903
+ "model.layers.76.self_attn.o_proj.weight": "model-00035-of-00037.safetensors",
904
+ "model.layers.76.self_attn.q_proj.bias": "model-00035-of-00037.safetensors",
905
+ "model.layers.76.self_attn.q_proj.weight": "model-00035-of-00037.safetensors",
906
+ "model.layers.76.self_attn.v_proj.bias": "model-00035-of-00037.safetensors",
907
+ "model.layers.76.self_attn.v_proj.weight": "model-00035-of-00037.safetensors",
908
+ "model.layers.77.input_layernorm.weight": "model-00035-of-00037.safetensors",
909
+ "model.layers.77.mlp.down_proj.weight": "model-00036-of-00037.safetensors",
910
+ "model.layers.77.mlp.gate_proj.weight": "model-00036-of-00037.safetensors",
911
+ "model.layers.77.mlp.up_proj.weight": "model-00036-of-00037.safetensors",
912
+ "model.layers.77.post_attention_layernorm.weight": "model-00035-of-00037.safetensors",
913
+ "model.layers.77.self_attn.k_proj.bias": "model-00035-of-00037.safetensors",
914
+ "model.layers.77.self_attn.k_proj.weight": "model-00035-of-00037.safetensors",
915
+ "model.layers.77.self_attn.o_proj.weight": "model-00035-of-00037.safetensors",
916
+ "model.layers.77.self_attn.q_proj.bias": "model-00035-of-00037.safetensors",
917
+ "model.layers.77.self_attn.q_proj.weight": "model-00035-of-00037.safetensors",
918
+ "model.layers.77.self_attn.v_proj.bias": "model-00035-of-00037.safetensors",
919
+ "model.layers.77.self_attn.v_proj.weight": "model-00035-of-00037.safetensors",
920
+ "model.layers.78.input_layernorm.weight": "model-00036-of-00037.safetensors",
921
+ "model.layers.78.mlp.down_proj.weight": "model-00036-of-00037.safetensors",
922
+ "model.layers.78.mlp.gate_proj.weight": "model-00036-of-00037.safetensors",
923
+ "model.layers.78.mlp.up_proj.weight": "model-00036-of-00037.safetensors",
924
+ "model.layers.78.post_attention_layernorm.weight": "model-00036-of-00037.safetensors",
925
+ "model.layers.78.self_attn.k_proj.bias": "model-00036-of-00037.safetensors",
926
+ "model.layers.78.self_attn.k_proj.weight": "model-00036-of-00037.safetensors",
927
+ "model.layers.78.self_attn.o_proj.weight": "model-00036-of-00037.safetensors",
928
+ "model.layers.78.self_attn.q_proj.bias": "model-00036-of-00037.safetensors",
929
+ "model.layers.78.self_attn.q_proj.weight": "model-00036-of-00037.safetensors",
930
+ "model.layers.78.self_attn.v_proj.bias": "model-00036-of-00037.safetensors",
931
+ "model.layers.78.self_attn.v_proj.weight": "model-00036-of-00037.safetensors",
932
+ "model.layers.79.input_layernorm.weight": "model-00036-of-00037.safetensors",
933
+ "model.layers.79.mlp.down_proj.weight": "model-00037-of-00037.safetensors",
934
+ "model.layers.79.mlp.gate_proj.weight": "model-00037-of-00037.safetensors",
935
+ "model.layers.79.mlp.up_proj.weight": "model-00036-of-00037.safetensors",
936
+ "model.layers.79.post_attention_layernorm.weight": "model-00036-of-00037.safetensors",
937
+ "model.layers.79.self_attn.k_proj.bias": "model-00036-of-00037.safetensors",
938
+ "model.layers.79.self_attn.k_proj.weight": "model-00036-of-00037.safetensors",
939
+ "model.layers.79.self_attn.o_proj.weight": "model-00036-of-00037.safetensors",
940
+ "model.layers.79.self_attn.q_proj.bias": "model-00036-of-00037.safetensors",
941
+ "model.layers.79.self_attn.q_proj.weight": "model-00036-of-00037.safetensors",
942
+ "model.layers.79.self_attn.v_proj.bias": "model-00036-of-00037.safetensors",
943
+ "model.layers.79.self_attn.v_proj.weight": "model-00036-of-00037.safetensors",
944
+ "model.layers.8.input_layernorm.weight": "model-00005-of-00037.safetensors",
945
+ "model.layers.8.mlp.down_proj.weight": "model-00005-of-00037.safetensors",
946
+ "model.layers.8.mlp.gate_proj.weight": "model-00005-of-00037.safetensors",
947
+ "model.layers.8.mlp.up_proj.weight": "model-00005-of-00037.safetensors",
948
+ "model.layers.8.post_attention_layernorm.weight": "model-00005-of-00037.safetensors",
949
+ "model.layers.8.self_attn.k_proj.bias": "model-00005-of-00037.safetensors",
950
+ "model.layers.8.self_attn.k_proj.weight": "model-00005-of-00037.safetensors",
951
+ "model.layers.8.self_attn.o_proj.weight": "model-00005-of-00037.safetensors",
952
+ "model.layers.8.self_attn.q_proj.bias": "model-00005-of-00037.safetensors",
953
+ "model.layers.8.self_attn.q_proj.weight": "model-00005-of-00037.safetensors",
954
+ "model.layers.8.self_attn.v_proj.bias": "model-00005-of-00037.safetensors",
955
+ "model.layers.8.self_attn.v_proj.weight": "model-00005-of-00037.safetensors",
956
+ "model.layers.9.input_layernorm.weight": "model-00005-of-00037.safetensors",
957
+ "model.layers.9.mlp.down_proj.weight": "model-00006-of-00037.safetensors",
958
+ "model.layers.9.mlp.gate_proj.weight": "model-00005-of-00037.safetensors",
959
+ "model.layers.9.mlp.up_proj.weight": "model-00005-of-00037.safetensors",
960
+ "model.layers.9.post_attention_layernorm.weight": "model-00005-of-00037.safetensors",
961
+ "model.layers.9.self_attn.k_proj.bias": "model-00005-of-00037.safetensors",
962
+ "model.layers.9.self_attn.k_proj.weight": "model-00005-of-00037.safetensors",
963
+ "model.layers.9.self_attn.o_proj.weight": "model-00005-of-00037.safetensors",
964
+ "model.layers.9.self_attn.q_proj.bias": "model-00005-of-00037.safetensors",
965
+ "model.layers.9.self_attn.q_proj.weight": "model-00005-of-00037.safetensors",
966
+ "model.layers.9.self_attn.v_proj.bias": "model-00005-of-00037.safetensors",
967
+ "model.layers.9.self_attn.v_proj.weight": "model-00005-of-00037.safetensors",
968
+ "model.norm.weight": "model-00037-of-00037.safetensors",
969
+ "score.0.bias": "model-00001-of-00037.safetensors",
970
+ "score.0.weight": "model-00001-of-00037.safetensors",
971
+ "score.2.bias": "model-00001-of-00037.safetensors",
972
+ "score.2.weight": "model-00001-of-00037.safetensors"
973
+ }
974
+ }
modeling_qwen2_rm.py ADDED
@@ -0,0 +1,1549 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 The Qwen team, Alibaba Group 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 Qwen2 model."""
21
+
22
+ import math
23
+ from typing import List, Optional, Tuple, Union
24
+
25
+ import torch
26
+ import torch.utils.checkpoint
27
+ from torch import nn
28
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
29
+
30
+ from transformers.activations import ACT2FN
31
+ from transformers.cache_utils import Cache, DynamicCache#, StaticCache
32
+ from transformers.modeling_attn_mask_utils import AttentionMaskConverter
33
+ from transformers.modeling_outputs import (
34
+ BaseModelOutputWithPast,
35
+ CausalLMOutputWithPast,
36
+ SequenceClassifierOutputWithPast,
37
+ TokenClassifierOutput,
38
+ )
39
+ from transformers.modeling_utils import PreTrainedModel
40
+ from transformers.utils import (
41
+ add_start_docstrings,
42
+ add_start_docstrings_to_model_forward,
43
+ is_flash_attn_2_available,
44
+ is_flash_attn_greater_or_equal_2_10,
45
+ logging,
46
+ replace_return_docstrings,
47
+ )
48
+ from .configuration_qwen2_rm import Qwen2RMConfig as Qwen2Config
49
+
50
+
51
+ # if is_flash_attn_2_available():
52
+ # from transformers.modeling_flash_attention_utils import _flash_attention_forward
53
+
54
+
55
+ logger = logging.get_logger(__name__)
56
+
57
+
58
+ _CHECKPOINT_FOR_DOC = "Qwen/Qwen2-7B-beta"
59
+ _CONFIG_FOR_DOC = "Qwen2Config"
60
+
61
+
62
+ # Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
63
+ def _prepare_4d_causal_attention_mask_with_cache_position(
64
+ attention_mask: torch.Tensor,
65
+ sequence_length: int,
66
+ target_length: int,
67
+ dtype: torch.dtype,
68
+ device: torch.device,
69
+ min_dtype: float,
70
+ cache_position: torch.Tensor,
71
+ batch_size: int,
72
+ ):
73
+ """
74
+ Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
75
+ `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
76
+
77
+ Args:
78
+ attention_mask (`torch.Tensor`):
79
+ A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
80
+ sequence_length (`int`):
81
+ The sequence length being processed.
82
+ target_length (`int`):
83
+ The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
84
+ dtype (`torch.dtype`):
85
+ The dtype to use for the 4D attention mask.
86
+ device (`torch.device`):
87
+ The device to plcae the 4D attention mask on.
88
+ min_dtype (`float`):
89
+ The minimum value representable with the dtype `dtype`.
90
+ cache_position (`torch.Tensor`):
91
+ Indices depicting the position of the input sequence tokens in the sequence.
92
+ batch_size (`torch.Tensor`):
93
+ Batch size.
94
+ """
95
+ if attention_mask is not None and attention_mask.dim() == 4:
96
+ # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
97
+ causal_mask = attention_mask
98
+ else:
99
+ causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
100
+ if sequence_length != 1:
101
+ causal_mask = torch.triu(causal_mask, diagonal=1)
102
+ causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
103
+ causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
104
+ if attention_mask is not None:
105
+ causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
106
+ mask_length = attention_mask.shape[-1]
107
+ padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
108
+ padding_mask = padding_mask == 0
109
+ causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
110
+ padding_mask, min_dtype
111
+ )
112
+
113
+ return causal_mask
114
+
115
+
116
+ # Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Qwen2
117
+ class Qwen2RMSNorm(nn.Module):
118
+ def __init__(self, hidden_size, eps=1e-6):
119
+ """
120
+ Qwen2RMSNorm is equivalent to T5LayerNorm
121
+ """
122
+ super().__init__()
123
+ self.weight = nn.Parameter(torch.ones(hidden_size))
124
+ self.variance_epsilon = eps
125
+
126
+ def forward(self, hidden_states):
127
+ input_dtype = hidden_states.dtype
128
+ hidden_states = hidden_states.to(torch.float32)
129
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
130
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
131
+ return self.weight * hidden_states.to(input_dtype)
132
+
133
+ def extra_repr(self):
134
+ return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
135
+
136
+
137
+ # Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->Qwen2
138
+ class Qwen2RotaryEmbedding(nn.Module):
139
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
140
+ super().__init__()
141
+
142
+ self.dim = dim
143
+ self.max_position_embeddings = max_position_embeddings
144
+ self.base = base
145
+ inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
146
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
147
+
148
+ # Build here to make `torch.jit.trace` work.
149
+ self._set_cos_sin_cache(
150
+ seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
151
+ )
152
+
153
+ def _set_cos_sin_cache(self, seq_len, device, dtype):
154
+ self.max_seq_len_cached = seq_len
155
+ t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
156
+
157
+ freqs = torch.outer(t, self.inv_freq)
158
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
159
+ emb = torch.cat((freqs, freqs), dim=-1)
160
+ self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
161
+ self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
162
+
163
+ def forward(self, x, seq_len=None):
164
+ # x: [bs, num_attention_heads, seq_len, head_size]
165
+ if seq_len > self.max_seq_len_cached:
166
+ self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
167
+
168
+ return (
169
+ self.cos_cached[:seq_len].to(dtype=x.dtype),
170
+ self.sin_cached[:seq_len].to(dtype=x.dtype),
171
+ )
172
+
173
+
174
+ # Copied from transformers.models.llama.modeling_llama.rotate_half
175
+ def rotate_half(x):
176
+ """Rotates half the hidden dims of the input."""
177
+ x1 = x[..., : x.shape[-1] // 2]
178
+ x2 = x[..., x.shape[-1] // 2 :]
179
+ return torch.cat((-x2, x1), dim=-1)
180
+
181
+
182
+ # Copied from transformers.models.mixtral.modeling_mixtral.apply_rotary_pos_emb
183
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
184
+ """Applies Rotary Position Embedding to the query and key tensors.
185
+
186
+ Args:
187
+ q (`torch.Tensor`): The query tensor.
188
+ k (`torch.Tensor`): The key tensor.
189
+ cos (`torch.Tensor`): The cosine part of the rotary embedding.
190
+ sin (`torch.Tensor`): The sine part of the rotary embedding.
191
+ position_ids (`torch.Tensor`):
192
+ The position indices of the tokens corresponding to the query and key tensors. For example, this can be
193
+ used to pass offsetted position ids when working with a KV-cache.
194
+ unsqueeze_dim (`int`, *optional*, defaults to 1):
195
+ The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
196
+ sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
197
+ that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
198
+ k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
199
+ cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
200
+ the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
201
+ Returns:
202
+ `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
203
+ """
204
+ cos = cos[position_ids].unsqueeze(unsqueeze_dim)
205
+ sin = sin[position_ids].unsqueeze(unsqueeze_dim)
206
+ q_embed = (q * cos) + (rotate_half(q) * sin)
207
+ k_embed = (k * cos) + (rotate_half(k) * sin)
208
+ return q_embed, k_embed
209
+
210
+
211
+ # Copied from transformers.models.mistral.modeling_mistral.MistralMLP with Mistral->Qwen2
212
+ class Qwen2MLP(nn.Module):
213
+ def __init__(self, config):
214
+ super().__init__()
215
+ self.hidden_size = config.hidden_size
216
+ self.intermediate_size = config.intermediate_size
217
+ self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
218
+ self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
219
+ self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
220
+ self.act_fn = ACT2FN[config.hidden_act]
221
+
222
+ def forward(self, hidden_state):
223
+ return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
224
+
225
+
226
+ # Copied from transformers.models.llama.modeling_llama.repeat_kv
227
+ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
228
+ """
229
+ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
230
+ num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
231
+ """
232
+ batch, num_key_value_heads, slen, head_dim = hidden_states.shape
233
+ if n_rep == 1:
234
+ return hidden_states
235
+ hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
236
+ return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
237
+
238
+
239
+ class Qwen2Attention(nn.Module):
240
+ """
241
+ Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
242
+ and "Generating Long Sequences with Sparse Transformers".
243
+ """
244
+
245
+ def __init__(self, config: Qwen2Config, layer_idx: Optional[int] = None):
246
+ super().__init__()
247
+ self.config = config
248
+ self.layer_idx = layer_idx
249
+ if layer_idx is None:
250
+ logger.warning_once(
251
+ f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
252
+ "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
253
+ "when creating this class."
254
+ )
255
+
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.num_key_value_heads = config.num_key_value_heads
260
+ self.num_key_value_groups = self.num_heads // self.num_key_value_heads
261
+ self.max_position_embeddings = config.max_position_embeddings
262
+ self.rope_theta = config.rope_theta
263
+ self.is_causal = True
264
+ self.attention_dropout = config.attention_dropout
265
+
266
+ if (self.head_dim * self.num_heads) != self.hidden_size:
267
+ raise ValueError(
268
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
269
+ f" and `num_heads`: {self.num_heads})."
270
+ )
271
+ self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
272
+ self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
273
+ self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
274
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
275
+
276
+ self.rotary_emb = Qwen2RotaryEmbedding(
277
+ self.head_dim,
278
+ max_position_embeddings=self.max_position_embeddings,
279
+ base=self.rope_theta,
280
+ )
281
+
282
+ def forward(
283
+ self,
284
+ hidden_states: torch.Tensor,
285
+ attention_mask: Optional[torch.Tensor] = None,
286
+ position_ids: Optional[torch.LongTensor] = None,
287
+ past_key_value: Optional[Cache] = None,
288
+ output_attentions: bool = False,
289
+ use_cache: bool = False,
290
+ cache_position: Optional[torch.LongTensor] = None,
291
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
292
+ bsz, q_len, _ = hidden_states.size()
293
+
294
+ query_states = self.q_proj(hidden_states)
295
+ key_states = self.k_proj(hidden_states)
296
+ value_states = self.v_proj(hidden_states)
297
+
298
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
299
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
300
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
301
+
302
+ kv_seq_len = key_states.shape[-2]
303
+ if past_key_value is not None:
304
+ if self.layer_idx is None:
305
+ raise ValueError(
306
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
307
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
308
+ "with a layer index."
309
+ )
310
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
311
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
312
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
313
+
314
+ if past_key_value is not None:
315
+ cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
316
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
317
+
318
+ # repeat k/v heads if n_kv_heads < n_heads
319
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
320
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
321
+
322
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
323
+
324
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
325
+ raise ValueError(
326
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
327
+ f" {attn_weights.size()}"
328
+ )
329
+
330
+ if attention_mask is not None: # no matter the length, we just slice it
331
+ causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
332
+ attn_weights = attn_weights + causal_mask
333
+
334
+ # upcast attention to fp32
335
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
336
+ attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
337
+ attn_output = torch.matmul(attn_weights, value_states)
338
+
339
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
340
+ raise ValueError(
341
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
342
+ f" {attn_output.size()}"
343
+ )
344
+
345
+ attn_output = attn_output.transpose(1, 2).contiguous()
346
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
347
+
348
+ attn_output = self.o_proj(attn_output)
349
+
350
+ if not output_attentions:
351
+ attn_weights = None
352
+
353
+ return attn_output, attn_weights, past_key_value
354
+
355
+
356
+ class Qwen2FlashAttention2(Qwen2Attention):
357
+ """
358
+ Qwen2 flash attention module, following Qwen2 attention module. This module inherits from `Qwen2Attention`
359
+ as the weights of the module stays untouched. The only required change would be on the forward pass
360
+ where it needs to correctly call the public API of flash attention and deal with padding tokens
361
+ in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom
362
+ config.max_window_layers layers.
363
+ """
364
+
365
+ # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
366
+ def __init__(self, *args, **kwargs):
367
+ super().__init__(*args, **kwargs)
368
+
369
+ # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
370
+ # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
371
+ # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
372
+ self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
373
+
374
+ def forward(
375
+ self,
376
+ hidden_states: torch.Tensor,
377
+ attention_mask: Optional[torch.Tensor] = None,
378
+ position_ids: Optional[torch.LongTensor] = None,
379
+ past_key_value: Optional[Cache] = None,
380
+ output_attentions: bool = False,
381
+ use_cache: bool = False,
382
+ cache_position: Optional[torch.LongTensor] = None,
383
+ ):
384
+ bsz, q_len, _ = hidden_states.size()
385
+
386
+ query_states = self.q_proj(hidden_states)
387
+ key_states = self.k_proj(hidden_states)
388
+ value_states = self.v_proj(hidden_states)
389
+
390
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
391
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
392
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
393
+
394
+ kv_seq_len = key_states.shape[-2]
395
+ if past_key_value is not None:
396
+ if self.layer_idx is None:
397
+ raise ValueError(
398
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
399
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
400
+ "with a layer index."
401
+ )
402
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
403
+
404
+ # Because the input can be padded, the absolute sequence length depends on the max position id.
405
+ rotary_seq_len = (
406
+ max(kv_seq_len, position_ids[:, -1].max().item() + 1) if position_ids is not None else kv_seq_len
407
+ )
408
+
409
+ cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)
410
+
411
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
412
+
413
+ if past_key_value is not None:
414
+ # Activate slicing cache only if the config has a value `sliding_windows` attribute
415
+ cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
416
+ if (
417
+ getattr(self.config, "sliding_window", None) is not None
418
+ and kv_seq_len > self.config.sliding_window
419
+ and cache_has_contents
420
+ ):
421
+ slicing_tokens = 1 - self.config.sliding_window
422
+
423
+ past_key = past_key_value[self.layer_idx][0]
424
+ past_value = past_key_value[self.layer_idx][1]
425
+
426
+ past_key = past_key[:, :, slicing_tokens:, :].contiguous()
427
+ past_value = past_value[:, :, slicing_tokens:, :].contiguous()
428
+
429
+ if past_key.shape[-2] != self.config.sliding_window - 1:
430
+ raise ValueError(
431
+ f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
432
+ f" {past_key.shape}"
433
+ )
434
+
435
+ if attention_mask is not None:
436
+ attention_mask = attention_mask[:, slicing_tokens:]
437
+ attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
438
+
439
+ cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
440
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
441
+
442
+ # repeat k/v heads if n_kv_heads < n_heads
443
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
444
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
445
+ dropout_rate = 0.0 if not self.training else self.attention_dropout
446
+
447
+ # In PEFT, usually we cast the layer norms in float32 for training stability reasons
448
+ # therefore the input hidden states gets silently casted in float32. Hence, we need
449
+ # cast them back in float16 just to be sure everything works as expected.
450
+ input_dtype = query_states.dtype
451
+ if input_dtype == torch.float32:
452
+ if torch.is_autocast_enabled():
453
+ target_dtype = torch.get_autocast_gpu_dtype()
454
+ # Handle the case where the model is quantized
455
+ elif hasattr(self.config, "_pre_quantization_dtype"):
456
+ target_dtype = self.config._pre_quantization_dtype
457
+ else:
458
+ target_dtype = self.q_proj.weight.dtype
459
+
460
+ logger.warning_once(
461
+ f"The input hidden states seems to be silently casted in float32, this might be related to"
462
+ f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
463
+ f" {target_dtype}."
464
+ )
465
+
466
+ query_states = query_states.to(target_dtype)
467
+ key_states = key_states.to(target_dtype)
468
+ value_states = value_states.to(target_dtype)
469
+
470
+ # Reashape to the expected shape for Flash Attention
471
+ query_states = query_states.transpose(1, 2)
472
+ key_states = key_states.transpose(1, 2)
473
+ value_states = value_states.transpose(1, 2)
474
+
475
+ if (
476
+ self.config.use_sliding_window
477
+ and getattr(self.config, "sliding_window", None) is not None
478
+ and self.layer_idx >= self.config.max_window_layers
479
+ ):
480
+ sliding_window = self.config.sliding_window
481
+ else:
482
+ sliding_window = None
483
+
484
+ attn_output = _flash_attention_forward(
485
+ query_states,
486
+ key_states,
487
+ value_states,
488
+ attention_mask,
489
+ q_len,
490
+ position_ids=position_ids,
491
+ dropout=dropout_rate,
492
+ sliding_window=sliding_window,
493
+ is_causal=self.is_causal,
494
+ use_top_left_mask=self._flash_attn_uses_top_left_mask,
495
+ )
496
+
497
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
498
+ attn_output = self.o_proj(attn_output)
499
+
500
+ if not output_attentions:
501
+ attn_weights = None
502
+
503
+ return attn_output, attn_weights, past_key_value
504
+
505
+
506
+ # Copied from transformers.models.mixtral.modeling_mixtral.MixtralSdpaAttention with Mixtral->Qwen2
507
+ class Qwen2SdpaAttention(Qwen2Attention):
508
+ """
509
+ Qwen2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
510
+ `Qwen2Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
511
+ SDPA API.
512
+ """
513
+
514
+ # Adapted from Qwen2Attention.forward
515
+ def forward(
516
+ self,
517
+ hidden_states: torch.Tensor,
518
+ attention_mask: Optional[torch.Tensor] = None,
519
+ position_ids: Optional[torch.LongTensor] = None,
520
+ past_key_value: Optional[Cache] = None,
521
+ output_attentions: bool = False,
522
+ use_cache: bool = False,
523
+ cache_position: Optional[torch.LongTensor] = None,
524
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
525
+ if output_attentions:
526
+ # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
527
+ logger.warning_once(
528
+ "Qwen2Model is using Qwen2SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
529
+ 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
530
+ )
531
+ return super().forward(
532
+ hidden_states=hidden_states,
533
+ attention_mask=attention_mask,
534
+ position_ids=position_ids,
535
+ past_key_value=past_key_value,
536
+ output_attentions=output_attentions,
537
+ use_cache=use_cache,
538
+ )
539
+
540
+ bsz, q_len, _ = hidden_states.size()
541
+
542
+ query_states = self.q_proj(hidden_states)
543
+ key_states = self.k_proj(hidden_states)
544
+ value_states = self.v_proj(hidden_states)
545
+
546
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
547
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
548
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
549
+
550
+ kv_seq_len = key_states.shape[-2]
551
+ if past_key_value is not None:
552
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
553
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
554
+
555
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
556
+
557
+ if past_key_value is not None:
558
+ cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
559
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
560
+
561
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
562
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
563
+
564
+ causal_mask = attention_mask
565
+ if attention_mask is not None: # no matter the length, we just slice it
566
+ causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
567
+
568
+ # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
569
+ # Reference: https://github.com/pytorch/pytorch/issues/112577.
570
+ if query_states.device.type == "cuda" and attention_mask is not None:
571
+ query_states = query_states.contiguous()
572
+ key_states = key_states.contiguous()
573
+ value_states = value_states.contiguous()
574
+
575
+ # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
576
+ # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
577
+ # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
578
+ is_causal = True if causal_mask is None and q_len > 1 else False
579
+
580
+ attn_output = torch.nn.functional.scaled_dot_product_attention(
581
+ query_states,
582
+ key_states,
583
+ value_states,
584
+ attn_mask=causal_mask,
585
+ dropout_p=self.attention_dropout if self.training else 0.0,
586
+ is_causal=is_causal,
587
+ )
588
+
589
+ attn_output = attn_output.transpose(1, 2).contiguous()
590
+ attn_output = attn_output.view(bsz, q_len, self.hidden_size)
591
+
592
+ attn_output = self.o_proj(attn_output)
593
+
594
+ return attn_output, None, past_key_value
595
+
596
+
597
+ QWEN2_ATTENTION_CLASSES = {
598
+ "eager": Qwen2Attention,
599
+ "flash_attention_2": Qwen2FlashAttention2,
600
+ "sdpa": Qwen2SdpaAttention,
601
+ }
602
+
603
+
604
+ class Qwen2DecoderLayer(nn.Module):
605
+ def __init__(self, config: Qwen2Config, layer_idx: int):
606
+ super().__init__()
607
+ self.hidden_size = config.hidden_size
608
+
609
+ if config.sliding_window and config._attn_implementation != "flash_attention_2":
610
+ logger.warning_once(
611
+ f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
612
+ "unexpected results may be encountered."
613
+ )
614
+ self.self_attn = QWEN2_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
615
+
616
+ self.mlp = Qwen2MLP(config)
617
+ self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
618
+ self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
619
+
620
+ def forward(
621
+ self,
622
+ hidden_states: torch.Tensor,
623
+ attention_mask: Optional[torch.Tensor] = None,
624
+ position_ids: Optional[torch.LongTensor] = None,
625
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
626
+ output_attentions: Optional[bool] = False,
627
+ use_cache: Optional[bool] = False,
628
+ cache_position: Optional[torch.LongTensor] = None,
629
+ **kwargs,
630
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
631
+ """
632
+ Args:
633
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
634
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
635
+ `(batch, sequence_length)` where padding elements are indicated by 0.
636
+ output_attentions (`bool`, *optional*):
637
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
638
+ returned tensors for more detail.
639
+ use_cache (`bool`, *optional*):
640
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
641
+ (see `past_key_values`).
642
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
643
+ cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
644
+ Indices depicting the position of the input sequence tokens in the sequence.
645
+ kwargs (`dict`, *optional*):
646
+ Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
647
+ into the model
648
+ """
649
+
650
+ residual = hidden_states
651
+
652
+ hidden_states = self.input_layernorm(hidden_states)
653
+
654
+ # Self Attention
655
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
656
+ hidden_states=hidden_states,
657
+ attention_mask=attention_mask,
658
+ position_ids=position_ids,
659
+ past_key_value=past_key_value,
660
+ output_attentions=output_attentions,
661
+ use_cache=use_cache,
662
+ cache_position=cache_position,
663
+ )
664
+ hidden_states = residual + hidden_states
665
+
666
+ # Fully Connected
667
+ residual = hidden_states
668
+ hidden_states = self.post_attention_layernorm(hidden_states)
669
+ hidden_states = self.mlp(hidden_states)
670
+ hidden_states = residual + hidden_states
671
+
672
+ outputs = (hidden_states,)
673
+
674
+ if output_attentions:
675
+ outputs += (self_attn_weights,)
676
+
677
+ if use_cache:
678
+ outputs += (present_key_value,)
679
+
680
+ return outputs
681
+
682
+
683
+ QWEN2_START_DOCSTRING = r"""
684
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
685
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
686
+ etc.)
687
+
688
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
689
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
690
+ and behavior.
691
+
692
+ Parameters:
693
+ config ([`Qwen2Config`]):
694
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
695
+ load the weights associated with the model, only the configuration. Check out the
696
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
697
+ """
698
+
699
+
700
+ @add_start_docstrings(
701
+ "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
702
+ QWEN2_START_DOCSTRING,
703
+ )
704
+ class Qwen2PreTrainedModel(PreTrainedModel):
705
+ config_class = Qwen2Config
706
+ base_model_prefix = "model"
707
+ supports_gradient_checkpointing = True
708
+ _no_split_modules = ["Qwen2DecoderLayer"]
709
+ _skip_keys_device_placement = "past_key_values"
710
+ _supports_flash_attn_2 = True
711
+ _supports_sdpa = True
712
+ _supports_cache_class = True
713
+
714
+ def _init_weights(self, module):
715
+ std = self.config.initializer_range
716
+ if isinstance(module, nn.Linear):
717
+ module.weight.data.normal_(mean=0.0, std=std)
718
+ if module.bias is not None:
719
+ module.bias.data.zero_()
720
+ elif isinstance(module, nn.Embedding):
721
+ module.weight.data.normal_(mean=0.0, std=std)
722
+ if module.padding_idx is not None:
723
+ module.weight.data[module.padding_idx].zero_()
724
+
725
+
726
+ QWEN2_INPUTS_DOCSTRING = r"""
727
+ Args:
728
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
729
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
730
+ it.
731
+
732
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
733
+ [`PreTrainedTokenizer.__call__`] for details.
734
+
735
+ [What are input IDs?](../glossary#input-ids)
736
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
737
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
738
+
739
+ - 1 for tokens that are **not masked**,
740
+ - 0 for tokens that are **masked**.
741
+
742
+ [What are attention masks?](../glossary#attention-mask)
743
+
744
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
745
+ [`PreTrainedTokenizer.__call__`] for details.
746
+
747
+ If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
748
+ `past_key_values`).
749
+
750
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
751
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
752
+ information on the default strategy.
753
+
754
+ - 1 indicates the head is **not masked**,
755
+ - 0 indicates the head is **masked**.
756
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
757
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
758
+ config.n_positions - 1]`.
759
+
760
+ [What are position IDs?](../glossary#position-ids)
761
+ past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
762
+ Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
763
+ blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
764
+ returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
765
+
766
+ Two formats are allowed:
767
+ - a [`~cache_utils.Cache`] instance;
768
+ - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
769
+ shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
770
+ cache format.
771
+
772
+ The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
773
+ legacy cache format will be returned.
774
+
775
+ If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
776
+ have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
777
+ of shape `(batch_size, sequence_length)`.
778
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
779
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
780
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
781
+ model's internal embedding lookup matrix.
782
+ use_cache (`bool`, *optional*):
783
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
784
+ `past_key_values`).
785
+ output_attentions (`bool`, *optional*):
786
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
787
+ tensors for more detail.
788
+ output_hidden_states (`bool`, *optional*):
789
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
790
+ more detail.
791
+ return_dict (`bool`, *optional*):
792
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
793
+ cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
794
+ Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
795
+ this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
796
+ the complete sequence length.
797
+ """
798
+
799
+
800
+ @add_start_docstrings(
801
+ "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
802
+ QWEN2_START_DOCSTRING,
803
+ )
804
+ class Qwen2Model(Qwen2PreTrainedModel):
805
+ """
806
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2DecoderLayer`]
807
+
808
+ Args:
809
+ config: Qwen2Config
810
+ """
811
+
812
+ def __init__(self, config: Qwen2Config):
813
+ super().__init__(config)
814
+ self.padding_idx = config.pad_token_id
815
+ self.vocab_size = config.vocab_size
816
+
817
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
818
+ self.layers = nn.ModuleList(
819
+ [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
820
+ )
821
+ self._attn_implementation = config._attn_implementation
822
+ self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
823
+
824
+ self.gradient_checkpointing = False
825
+ # Initialize weights and apply final processing
826
+ self.post_init()
827
+
828
+ def get_input_embeddings(self):
829
+ return self.embed_tokens
830
+
831
+ def set_input_embeddings(self, value):
832
+ self.embed_tokens = value
833
+
834
+ @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
835
+ def forward(
836
+ self,
837
+ input_ids: torch.LongTensor = None,
838
+ attention_mask: Optional[torch.Tensor] = None,
839
+ position_ids: Optional[torch.LongTensor] = None,
840
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
841
+ inputs_embeds: Optional[torch.FloatTensor] = None,
842
+ use_cache: Optional[bool] = None,
843
+ output_attentions: Optional[bool] = None,
844
+ output_hidden_states: Optional[bool] = None,
845
+ return_dict: Optional[bool] = None,
846
+ cache_position: Optional[torch.LongTensor] = None,
847
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
848
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
849
+ output_hidden_states = (
850
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
851
+ )
852
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
853
+
854
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
855
+
856
+ if (input_ids is None) ^ (inputs_embeds is not None):
857
+ raise ValueError(
858
+ "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
859
+ )
860
+
861
+ if self.gradient_checkpointing and self.training:
862
+ if use_cache:
863
+ logger.warning_once(
864
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
865
+ )
866
+ use_cache = False
867
+
868
+ use_legacy_cache = False
869
+ if use_cache and not isinstance(past_key_values, Cache) and not self.training:
870
+ use_legacy_cache = True
871
+ past_key_values = DynamicCache.from_legacy_cache(past_key_values)
872
+ logger.warning_once(
873
+ "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. "
874
+ "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)"
875
+ )
876
+
877
+ if inputs_embeds is None:
878
+ inputs_embeds = self.embed_tokens(input_ids)
879
+
880
+ if cache_position is None:
881
+ past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
882
+ cache_position = torch.arange(
883
+ past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
884
+ )
885
+ if position_ids is None:
886
+ position_ids = cache_position.unsqueeze(0)
887
+
888
+ causal_mask = self._update_causal_mask(
889
+ attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
890
+ )
891
+
892
+ hidden_states = inputs_embeds
893
+
894
+ # decoder layers
895
+ all_hidden_states = () if output_hidden_states else None
896
+ all_self_attns = () if output_attentions else None
897
+ next_decoder_cache = None
898
+
899
+ for decoder_layer in self.layers:
900
+ if output_hidden_states:
901
+ all_hidden_states += (hidden_states,)
902
+
903
+ if self.gradient_checkpointing and self.training:
904
+ layer_outputs = self._gradient_checkpointing_func(
905
+ decoder_layer.__call__,
906
+ hidden_states,
907
+ causal_mask,
908
+ position_ids,
909
+ past_key_values,
910
+ output_attentions,
911
+ use_cache,
912
+ cache_position,
913
+ )
914
+ else:
915
+ layer_outputs = decoder_layer(
916
+ hidden_states,
917
+ attention_mask=causal_mask,
918
+ position_ids=position_ids,
919
+ past_key_value=past_key_values,
920
+ output_attentions=output_attentions,
921
+ use_cache=use_cache,
922
+ cache_position=cache_position,
923
+ )
924
+
925
+ hidden_states = layer_outputs[0]
926
+
927
+ if use_cache:
928
+ next_decoder_cache = layer_outputs[2 if output_attentions else 1]
929
+
930
+ if output_attentions:
931
+ all_self_attns += (layer_outputs[1],)
932
+
933
+ hidden_states = self.norm(hidden_states)
934
+
935
+ # add hidden states from the last decoder layer
936
+ if output_hidden_states:
937
+ all_hidden_states += (hidden_states,)
938
+
939
+ next_cache = None
940
+ if use_cache:
941
+ next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
942
+
943
+ if not return_dict:
944
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
945
+ return BaseModelOutputWithPast(
946
+ last_hidden_state=hidden_states,
947
+ past_key_values=next_cache,
948
+ hidden_states=all_hidden_states,
949
+ attentions=all_self_attns,
950
+ )
951
+
952
+ # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
953
+ def _update_causal_mask(
954
+ self,
955
+ attention_mask: torch.Tensor,
956
+ input_tensor: torch.Tensor,
957
+ cache_position: torch.Tensor,
958
+ past_key_values: Cache,
959
+ output_attentions: bool,
960
+ ):
961
+ # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
962
+ # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
963
+ # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
964
+ # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
965
+
966
+ if self.config._attn_implementation == "flash_attention_2":
967
+ if attention_mask is not None and 0.0 in attention_mask:
968
+ return attention_mask
969
+ return None
970
+
971
+ # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
972
+ # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
973
+ # to infer the attention mask.
974
+ past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
975
+ using_static_cache = False#isinstance(past_key_values, StaticCache)
976
+
977
+ # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
978
+ if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
979
+ if AttentionMaskConverter._ignore_causal_mask_sdpa(
980
+ attention_mask,
981
+ inputs_embeds=input_tensor,
982
+ past_key_values_length=past_seen_tokens,
983
+ is_training=self.training,
984
+ ):
985
+ return None
986
+
987
+ dtype, device = input_tensor.dtype, input_tensor.device
988
+ min_dtype = torch.finfo(dtype).min
989
+ sequence_length = input_tensor.shape[1]
990
+ if using_static_cache:
991
+ target_length = past_key_values.get_max_length()
992
+ else:
993
+ target_length = (
994
+ attention_mask.shape[-1]
995
+ if isinstance(attention_mask, torch.Tensor)
996
+ else past_seen_tokens + sequence_length + 1
997
+ )
998
+
999
+ # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
1000
+ causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
1001
+ attention_mask,
1002
+ sequence_length=sequence_length,
1003
+ target_length=target_length,
1004
+ dtype=dtype,
1005
+ device=device,
1006
+ min_dtype=min_dtype,
1007
+ cache_position=cache_position,
1008
+ batch_size=input_tensor.shape[0],
1009
+ )
1010
+
1011
+ if (
1012
+ self.config._attn_implementation == "sdpa"
1013
+ and attention_mask is not None
1014
+ and attention_mask.device.type == "cuda"
1015
+ and not output_attentions
1016
+ ):
1017
+ # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
1018
+ # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
1019
+ # Details: https://github.com/pytorch/pytorch/issues/110213
1020
+ causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
1021
+
1022
+ return causal_mask
1023
+
1024
+
1025
+ class Qwen2ForCausalLM(Qwen2PreTrainedModel):
1026
+ _tied_weights_keys = ["lm_head.weight"]
1027
+
1028
+ def __init__(self, config):
1029
+ super().__init__(config)
1030
+ self.model = Qwen2Model(config)
1031
+ self.vocab_size = config.vocab_size
1032
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
1033
+
1034
+ # Initialize weights and apply final processing
1035
+ self.post_init()
1036
+
1037
+ def get_input_embeddings(self):
1038
+ return self.model.embed_tokens
1039
+
1040
+ def set_input_embeddings(self, value):
1041
+ self.model.embed_tokens = value
1042
+
1043
+ def get_output_embeddings(self):
1044
+ return self.lm_head
1045
+
1046
+ def set_output_embeddings(self, new_embeddings):
1047
+ self.lm_head = new_embeddings
1048
+
1049
+ def set_decoder(self, decoder):
1050
+ self.model = decoder
1051
+
1052
+ def get_decoder(self):
1053
+ return self.model
1054
+
1055
+ @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
1056
+ @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
1057
+ def forward(
1058
+ self,
1059
+ input_ids: torch.LongTensor = None,
1060
+ attention_mask: Optional[torch.Tensor] = None,
1061
+ position_ids: Optional[torch.LongTensor] = None,
1062
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1063
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1064
+ labels: Optional[torch.LongTensor] = None,
1065
+ use_cache: Optional[bool] = None,
1066
+ output_attentions: Optional[bool] = None,
1067
+ output_hidden_states: Optional[bool] = None,
1068
+ return_dict: Optional[bool] = None,
1069
+ cache_position: Optional[torch.LongTensor] = None,
1070
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
1071
+ r"""
1072
+ Args:
1073
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1074
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
1075
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
1076
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
1077
+
1078
+ Returns:
1079
+
1080
+ Example:
1081
+
1082
+ ```python
1083
+ >>> from transformers import AutoTokenizer, Qwen2ForCausalLM
1084
+
1085
+ >>> model = Qwen2ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
1086
+ >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
1087
+
1088
+ >>> prompt = "Hey, are you conscious? Can you talk to me?"
1089
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
1090
+
1091
+ >>> # Generate
1092
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
1093
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
1094
+ "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
1095
+ ```"""
1096
+
1097
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
1098
+ output_hidden_states = (
1099
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
1100
+ )
1101
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1102
+
1103
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
1104
+ outputs = self.model(
1105
+ input_ids=input_ids,
1106
+ attention_mask=attention_mask,
1107
+ position_ids=position_ids,
1108
+ past_key_values=past_key_values,
1109
+ inputs_embeds=inputs_embeds,
1110
+ use_cache=use_cache,
1111
+ output_attentions=output_attentions,
1112
+ output_hidden_states=output_hidden_states,
1113
+ return_dict=return_dict,
1114
+ cache_position=cache_position,
1115
+ )
1116
+
1117
+ hidden_states = outputs[0]
1118
+ logits = self.lm_head(hidden_states)
1119
+ logits = logits.float()
1120
+
1121
+ loss = None
1122
+ if labels is not None:
1123
+ # Shift so that tokens < n predict n
1124
+ shift_logits = logits[..., :-1, :].contiguous()
1125
+ shift_labels = labels[..., 1:].contiguous()
1126
+ # Flatten the tokens
1127
+ loss_fct = CrossEntropyLoss()
1128
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
1129
+ shift_labels = shift_labels.view(-1)
1130
+ # Enable model parallelism
1131
+ shift_labels = shift_labels.to(shift_logits.device)
1132
+ loss = loss_fct(shift_logits, shift_labels)
1133
+
1134
+ if not return_dict:
1135
+ output = (logits,) + outputs[1:]
1136
+ return (loss,) + output if loss is not None else output
1137
+
1138
+ return CausalLMOutputWithPast(
1139
+ loss=loss,
1140
+ logits=logits,
1141
+ past_key_values=outputs.past_key_values,
1142
+ hidden_states=outputs.hidden_states,
1143
+ attentions=outputs.attentions,
1144
+ )
1145
+
1146
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.prepare_inputs_for_generation
1147
+ def prepare_inputs_for_generation(
1148
+ self,
1149
+ input_ids,
1150
+ past_key_values=None,
1151
+ attention_mask=None,
1152
+ inputs_embeds=None,
1153
+ cache_position=None,
1154
+ position_ids=None,
1155
+ use_cache=True,
1156
+ **kwargs,
1157
+ ):
1158
+ # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
1159
+ # Exception 1: when passing input_embeds, input_ids may be missing entries
1160
+ # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
1161
+ if past_key_values is not None:
1162
+ if inputs_embeds is not None: # Exception 1
1163
+ input_ids = input_ids[:, -cache_position.shape[0] :]
1164
+ elif input_ids.shape[1] != cache_position.shape[0]: # Default case (the "else", a no op, is Exception 2)
1165
+ input_ids = input_ids[:, cache_position]
1166
+
1167
+ if attention_mask is not None and position_ids is None:
1168
+ # create position_ids on the fly for batch generation
1169
+ position_ids = attention_mask.long().cumsum(-1) - 1
1170
+ position_ids.masked_fill_(attention_mask == 0, 1)
1171
+ if past_key_values:
1172
+ position_ids = position_ids[:, -input_ids.shape[1] :]
1173
+
1174
+ # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
1175
+ position_ids = position_ids.clone(memory_format=torch.contiguous_format)
1176
+
1177
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1178
+ if inputs_embeds is not None and cache_position[0] == 0:
1179
+ model_inputs = {"inputs_embeds": inputs_embeds}
1180
+ else:
1181
+ model_inputs = {"input_ids": input_ids}
1182
+
1183
+ if False and isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
1184
+ if inputs_embeds is not None:
1185
+ batch_size, sequence_length = inputs_embeds.shape
1186
+ device = inputs_embeds.device
1187
+ else:
1188
+ batch_size, sequence_length = input_ids.shape
1189
+ device = input_ids.device
1190
+
1191
+ dtype = self.lm_head.weight.dtype
1192
+ min_dtype = torch.finfo(dtype).min
1193
+
1194
+ attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
1195
+ attention_mask,
1196
+ sequence_length=sequence_length,
1197
+ target_length=past_key_values.get_max_length(),
1198
+ dtype=dtype,
1199
+ device=device,
1200
+ min_dtype=min_dtype,
1201
+ cache_position=cache_position,
1202
+ batch_size=batch_size,
1203
+ )
1204
+
1205
+ model_inputs.update(
1206
+ {
1207
+ "position_ids": position_ids,
1208
+ "cache_position": cache_position,
1209
+ "past_key_values": past_key_values,
1210
+ "use_cache": use_cache,
1211
+ "attention_mask": attention_mask,
1212
+ }
1213
+ )
1214
+ return model_inputs
1215
+
1216
+
1217
+ @add_start_docstrings(
1218
+ """
1219
+ The Qwen2 Model transformer with a sequence classification head on top (linear layer).
1220
+
1221
+ [`Qwen2ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
1222
+ (e.g. GPT-2) do.
1223
+
1224
+ Since it does classification on the last token, it requires to know the position of the last token. If a
1225
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
1226
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
1227
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
1228
+ each row of the batch).
1229
+ """,
1230
+ QWEN2_START_DOCSTRING,
1231
+ )
1232
+ class Qwen2ForSequenceClassification(Qwen2PreTrainedModel):
1233
+ def __init__(self, config):
1234
+ super().__init__(config)
1235
+ self.num_labels = config.num_labels
1236
+ self.model = Qwen2Model(config)
1237
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
1238
+
1239
+ # Initialize weights and apply final processing
1240
+ self.post_init()
1241
+
1242
+ def get_input_embeddings(self):
1243
+ return self.model.embed_tokens
1244
+
1245
+ def set_input_embeddings(self, value):
1246
+ self.model.embed_tokens = value
1247
+
1248
+ @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
1249
+ def forward(
1250
+ self,
1251
+ input_ids: torch.LongTensor = None,
1252
+ attention_mask: Optional[torch.Tensor] = None,
1253
+ position_ids: Optional[torch.LongTensor] = None,
1254
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1255
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1256
+ labels: Optional[torch.LongTensor] = None,
1257
+ use_cache: Optional[bool] = None,
1258
+ output_attentions: Optional[bool] = None,
1259
+ output_hidden_states: Optional[bool] = None,
1260
+ return_dict: Optional[bool] = None,
1261
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1262
+ r"""
1263
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1264
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1265
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1266
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1267
+ """
1268
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1269
+
1270
+ transformer_outputs = self.model(
1271
+ input_ids,
1272
+ attention_mask=attention_mask,
1273
+ position_ids=position_ids,
1274
+ past_key_values=past_key_values,
1275
+ inputs_embeds=inputs_embeds,
1276
+ use_cache=use_cache,
1277
+ output_attentions=output_attentions,
1278
+ output_hidden_states=output_hidden_states,
1279
+ return_dict=return_dict,
1280
+ )
1281
+ hidden_states = transformer_outputs[0]
1282
+ logits = self.score(hidden_states)
1283
+
1284
+ if input_ids is not None:
1285
+ batch_size = input_ids.shape[0]
1286
+ else:
1287
+ batch_size = inputs_embeds.shape[0]
1288
+
1289
+ if self.config.pad_token_id is None and batch_size != 1:
1290
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
1291
+ if self.config.pad_token_id is None:
1292
+ sequence_lengths = -1
1293
+ else:
1294
+ if input_ids is not None:
1295
+ # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
1296
+ sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
1297
+ sequence_lengths = sequence_lengths % input_ids.shape[-1]
1298
+ sequence_lengths = sequence_lengths.to(logits.device)
1299
+ else:
1300
+ sequence_lengths = -1
1301
+
1302
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1303
+
1304
+ loss = None
1305
+ if labels is not None:
1306
+ labels = labels.to(logits.device)
1307
+ if self.config.problem_type is None:
1308
+ if self.num_labels == 1:
1309
+ self.config.problem_type = "regression"
1310
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1311
+ self.config.problem_type = "single_label_classification"
1312
+ else:
1313
+ self.config.problem_type = "multi_label_classification"
1314
+
1315
+ if self.config.problem_type == "regression":
1316
+ loss_fct = MSELoss()
1317
+ if self.num_labels == 1:
1318
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1319
+ else:
1320
+ loss = loss_fct(pooled_logits, labels)
1321
+ elif self.config.problem_type == "single_label_classification":
1322
+ loss_fct = CrossEntropyLoss()
1323
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1324
+ elif self.config.problem_type == "multi_label_classification":
1325
+ loss_fct = BCEWithLogitsLoss()
1326
+ loss = loss_fct(pooled_logits, labels)
1327
+ if not return_dict:
1328
+ output = (pooled_logits,) + transformer_outputs[1:]
1329
+ return ((loss,) + output) if loss is not None else output
1330
+
1331
+ return SequenceClassifierOutputWithPast(
1332
+ loss=loss,
1333
+ logits=pooled_logits,
1334
+ past_key_values=transformer_outputs.past_key_values,
1335
+ hidden_states=transformer_outputs.hidden_states,
1336
+ attentions=transformer_outputs.attentions,
1337
+ )
1338
+
1339
+
1340
+ @add_start_docstrings(
1341
+ """
1342
+ The Qwen2 Model transformer with a token classification head on top (a linear layer on top of the hidden-states
1343
+ output) e.g. for Named-Entity-Recognition (NER) tasks.
1344
+ """,
1345
+ QWEN2_START_DOCSTRING,
1346
+ )
1347
+ # Copied from transformers.models.llama.modeling_llama.LlamaForTokenClassification with Llama->Qwen2, LLAMA->QWEN2
1348
+ class Qwen2ForTokenClassification(Qwen2PreTrainedModel):
1349
+ def __init__(self, config):
1350
+ super().__init__(config)
1351
+ self.num_labels = config.num_labels
1352
+ self.model = Qwen2Model(config)
1353
+ if getattr(config, "classifier_dropout", None) is not None:
1354
+ classifier_dropout = config.classifier_dropout
1355
+ elif getattr(config, "hidden_dropout", None) is not None:
1356
+ classifier_dropout = config.hidden_dropout
1357
+ else:
1358
+ classifier_dropout = 0.1
1359
+ self.dropout = nn.Dropout(classifier_dropout)
1360
+ self.score = nn.Linear(config.hidden_size, config.num_labels)
1361
+
1362
+ # Initialize weights and apply final processing
1363
+ self.post_init()
1364
+
1365
+ def get_input_embeddings(self):
1366
+ return self.model.embed_tokens
1367
+
1368
+ def set_input_embeddings(self, value):
1369
+ self.model.embed_tokens = value
1370
+
1371
+ @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
1372
+ def forward(
1373
+ self,
1374
+ input_ids: Optional[torch.LongTensor] = None,
1375
+ attention_mask: Optional[torch.Tensor] = None,
1376
+ position_ids: Optional[torch.LongTensor] = None,
1377
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1378
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1379
+ labels: Optional[torch.LongTensor] = None,
1380
+ use_cache: Optional[bool] = None,
1381
+ output_attentions: Optional[bool] = None,
1382
+ output_hidden_states: Optional[bool] = None,
1383
+ return_dict: Optional[bool] = None,
1384
+ ) -> Union[Tuple, TokenClassifierOutput]:
1385
+ r"""
1386
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1387
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1388
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1389
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1390
+ """
1391
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1392
+
1393
+ outputs = self.model(
1394
+ input_ids,
1395
+ attention_mask=attention_mask,
1396
+ position_ids=position_ids,
1397
+ past_key_values=past_key_values,
1398
+ inputs_embeds=inputs_embeds,
1399
+ use_cache=use_cache,
1400
+ output_attentions=output_attentions,
1401
+ output_hidden_states=output_hidden_states,
1402
+ return_dict=return_dict,
1403
+ )
1404
+ sequence_output = outputs[0]
1405
+ sequence_output = self.dropout(sequence_output)
1406
+ logits = self.score(sequence_output)
1407
+
1408
+ loss = None
1409
+ if labels is not None:
1410
+ loss_fct = CrossEntropyLoss()
1411
+ loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
1412
+
1413
+ if not return_dict:
1414
+ output = (logits,) + outputs[2:]
1415
+ return ((loss,) + output) if loss is not None else output
1416
+
1417
+ return TokenClassifierOutput(
1418
+ loss=loss,
1419
+ logits=logits,
1420
+ hidden_states=outputs.hidden_states,
1421
+ attentions=outputs.attentions,
1422
+ )
1423
+
1424
+
1425
+ @add_start_docstrings(
1426
+ """
1427
+ The Qwen2 Model transformer with a sequence classification head on top (linear layer).
1428
+
1429
+ [`Qwen2ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
1430
+ (e.g. GPT-2) do.
1431
+
1432
+ Since it does classification on the last token, it requires to know the position of the last token. If a
1433
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
1434
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
1435
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
1436
+ each row of the batch).
1437
+ """,
1438
+ QWEN2_START_DOCSTRING,
1439
+ )
1440
+ class Qwen2ForRewardModel(Qwen2PreTrainedModel):
1441
+ def __init__(self, config):
1442
+ super().__init__(config)
1443
+ self.num_labels = 1#config.num_labels
1444
+ self.model = Qwen2Model(config)
1445
+ self.score = nn.Sequential(
1446
+ nn.Linear(config.hidden_size, config.hidden_size),
1447
+ nn.ReLU(),
1448
+ nn.Linear(config.hidden_size, self.num_labels)
1449
+ )
1450
+
1451
+ # Initialize weights and apply final processing
1452
+ self.post_init()
1453
+
1454
+ def get_input_embeddings(self):
1455
+ return self.model.embed_tokens
1456
+
1457
+ def set_input_embeddings(self, value):
1458
+ self.model.embed_tokens = value
1459
+
1460
+ @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
1461
+ def forward(
1462
+ self,
1463
+ input_ids: torch.LongTensor = None,
1464
+ attention_mask: Optional[torch.Tensor] = None,
1465
+ position_ids: Optional[torch.LongTensor] = None,
1466
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1467
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1468
+ labels: Optional[torch.LongTensor] = None,
1469
+ use_cache: Optional[bool] = None,
1470
+ output_attentions: Optional[bool] = None,
1471
+ output_hidden_states: Optional[bool] = None,
1472
+ return_dict: Optional[bool] = None,
1473
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1474
+ r"""
1475
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1476
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1477
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1478
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1479
+ """
1480
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1481
+
1482
+ transformer_outputs = self.model(
1483
+ input_ids,
1484
+ attention_mask=attention_mask,
1485
+ position_ids=position_ids,
1486
+ past_key_values=past_key_values,
1487
+ inputs_embeds=inputs_embeds,
1488
+ use_cache=use_cache,
1489
+ output_attentions=output_attentions,
1490
+ output_hidden_states=output_hidden_states,
1491
+ return_dict=return_dict,
1492
+ )
1493
+ hidden_states = transformer_outputs[0]
1494
+ logits = self.score(hidden_states)
1495
+
1496
+ if input_ids is not None:
1497
+ batch_size = input_ids.shape[0]
1498
+ else:
1499
+ batch_size = inputs_embeds.shape[0]
1500
+
1501
+ if self.config.pad_token_id is None and batch_size != 1:
1502
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
1503
+ if self.config.pad_token_id is None:
1504
+ sequence_lengths = -1
1505
+ else:
1506
+ if input_ids is not None:
1507
+ # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
1508
+ sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
1509
+ sequence_lengths = sequence_lengths % input_ids.shape[-1]
1510
+ sequence_lengths = sequence_lengths.to(logits.device)
1511
+ else:
1512
+ sequence_lengths = -1
1513
+
1514
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1515
+
1516
+ loss = None
1517
+ if labels is not None:
1518
+ labels = labels.to(logits.device)
1519
+ if self.config.problem_type is None:
1520
+ if self.num_labels == 1:
1521
+ self.config.problem_type = "regression"
1522
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1523
+ self.config.problem_type = "single_label_classification"
1524
+ else:
1525
+ self.config.problem_type = "multi_label_classification"
1526
+
1527
+ if self.config.problem_type == "regression":
1528
+ loss_fct = MSELoss()
1529
+ if self.num_labels == 1:
1530
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1531
+ else:
1532
+ loss = loss_fct(pooled_logits, labels)
1533
+ elif self.config.problem_type == "single_label_classification":
1534
+ loss_fct = CrossEntropyLoss()
1535
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1536
+ elif self.config.problem_type == "multi_label_classification":
1537
+ loss_fct = BCEWithLogitsLoss()
1538
+ loss = loss_fct(pooled_logits, labels)
1539
+ if not return_dict:
1540
+ output = (pooled_logits,) + transformer_outputs[1:]
1541
+ return ((loss,) + output) if loss is not None else output
1542
+
1543
+ return SequenceClassifierOutputWithPast(
1544
+ loss=loss,
1545
+ logits=pooled_logits,
1546
+ past_key_values=transformer_outputs.past_key_values,
1547
+ hidden_states=transformer_outputs.hidden_states,
1548
+ attentions=transformer_outputs.attentions,
1549
+ )
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,207 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "added_tokens_decoder": {
4
+ "151643": {
5
+ "content": "<|endoftext|>",
6
+ "lstrip": false,
7
+ "normalized": false,
8
+ "rstrip": false,
9
+ "single_word": false,
10
+ "special": true
11
+ },
12
+ "151644": {
13
+ "content": "<|im_start|>",
14
+ "lstrip": false,
15
+ "normalized": false,
16
+ "rstrip": false,
17
+ "single_word": false,
18
+ "special": true
19
+ },
20
+ "151645": {
21
+ "content": "<|im_end|>",
22
+ "lstrip": false,
23
+ "normalized": false,
24
+ "rstrip": false,
25
+ "single_word": false,
26
+ "special": true
27
+ },
28
+ "151646": {
29
+ "content": "<|object_ref_start|>",
30
+ "lstrip": false,
31
+ "normalized": false,
32
+ "rstrip": false,
33
+ "single_word": false,
34
+ "special": true
35
+ },
36
+ "151647": {
37
+ "content": "<|object_ref_end|>",
38
+ "lstrip": false,
39
+ "normalized": false,
40
+ "rstrip": false,
41
+ "single_word": false,
42
+ "special": true
43
+ },
44
+ "151648": {
45
+ "content": "<|box_start|>",
46
+ "lstrip": false,
47
+ "normalized": false,
48
+ "rstrip": false,
49
+ "single_word": false,
50
+ "special": true
51
+ },
52
+ "151649": {
53
+ "content": "<|box_end|>",
54
+ "lstrip": false,
55
+ "normalized": false,
56
+ "rstrip": false,
57
+ "single_word": false,
58
+ "special": true
59
+ },
60
+ "151650": {
61
+ "content": "<|quad_start|>",
62
+ "lstrip": false,
63
+ "normalized": false,
64
+ "rstrip": false,
65
+ "single_word": false,
66
+ "special": true
67
+ },
68
+ "151651": {
69
+ "content": "<|quad_end|>",
70
+ "lstrip": false,
71
+ "normalized": false,
72
+ "rstrip": false,
73
+ "single_word": false,
74
+ "special": true
75
+ },
76
+ "151652": {
77
+ "content": "<|vision_start|>",
78
+ "lstrip": false,
79
+ "normalized": false,
80
+ "rstrip": false,
81
+ "single_word": false,
82
+ "special": true
83
+ },
84
+ "151653": {
85
+ "content": "<|vision_end|>",
86
+ "lstrip": false,
87
+ "normalized": false,
88
+ "rstrip": false,
89
+ "single_word": false,
90
+ "special": true
91
+ },
92
+ "151654": {
93
+ "content": "<|vision_pad|>",
94
+ "lstrip": false,
95
+ "normalized": false,
96
+ "rstrip": false,
97
+ "single_word": false,
98
+ "special": true
99
+ },
100
+ "151655": {
101
+ "content": "<|image_pad|>",
102
+ "lstrip": false,
103
+ "normalized": false,
104
+ "rstrip": false,
105
+ "single_word": false,
106
+ "special": true
107
+ },
108
+ "151656": {
109
+ "content": "<|video_pad|>",
110
+ "lstrip": false,
111
+ "normalized": false,
112
+ "rstrip": false,
113
+ "single_word": false,
114
+ "special": true
115
+ },
116
+ "151657": {
117
+ "content": "<tool_call>",
118
+ "lstrip": false,
119
+ "normalized": false,
120
+ "rstrip": false,
121
+ "single_word": false,
122
+ "special": false
123
+ },
124
+ "151658": {
125
+ "content": "</tool_call>",
126
+ "lstrip": false,
127
+ "normalized": false,
128
+ "rstrip": false,
129
+ "single_word": false,
130
+ "special": false
131
+ },
132
+ "151659": {
133
+ "content": "<|fim_prefix|>",
134
+ "lstrip": false,
135
+ "normalized": false,
136
+ "rstrip": false,
137
+ "single_word": false,
138
+ "special": false
139
+ },
140
+ "151660": {
141
+ "content": "<|fim_middle|>",
142
+ "lstrip": false,
143
+ "normalized": false,
144
+ "rstrip": false,
145
+ "single_word": false,
146
+ "special": false
147
+ },
148
+ "151661": {
149
+ "content": "<|fim_suffix|>",
150
+ "lstrip": false,
151
+ "normalized": false,
152
+ "rstrip": false,
153
+ "single_word": false,
154
+ "special": false
155
+ },
156
+ "151662": {
157
+ "content": "<|fim_pad|>",
158
+ "lstrip": false,
159
+ "normalized": false,
160
+ "rstrip": false,
161
+ "single_word": false,
162
+ "special": false
163
+ },
164
+ "151663": {
165
+ "content": "<|repo_name|>",
166
+ "lstrip": false,
167
+ "normalized": false,
168
+ "rstrip": false,
169
+ "single_word": false,
170
+ "special": false
171
+ },
172
+ "151664": {
173
+ "content": "<|file_sep|>",
174
+ "lstrip": false,
175
+ "normalized": false,
176
+ "rstrip": false,
177
+ "single_word": false,
178
+ "special": false
179
+ }
180
+ },
181
+ "additional_special_tokens": [
182
+ "<|im_start|>",
183
+ "<|im_end|>",
184
+ "<|object_ref_start|>",
185
+ "<|object_ref_end|>",
186
+ "<|box_start|>",
187
+ "<|box_end|>",
188
+ "<|quad_start|>",
189
+ "<|quad_end|>",
190
+ "<|vision_start|>",
191
+ "<|vision_end|>",
192
+ "<|vision_pad|>",
193
+ "<|image_pad|>",
194
+ "<|video_pad|>"
195
+ ],
196
+ "bos_token": null,
197
+ "chat_template": "{%- if tools %}\n {{- '<|im_start|>system\\n' }}\n {%- if messages[0]['role'] == 'system' %}\n {{- messages[0]['content'] }}\n {%- else %}\n {{- 'Please reason step by step, and put your final answer within \\\\boxed{}.' }}\n {%- endif %}\n {{- \"\\n\\n# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n {%- for tool in tools %}\n {{- \"\\n\" }}\n {{- tool | tojson }}\n {%- endfor %}\n {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n {%- if messages[0]['role'] == 'system' %}\n {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n {%- else %}\n {{- '<|im_start|>system\\nPlease reason step by step, and put your final answer within \\\\boxed{}.<|im_end|>\\n' }}\n {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n {%- elif message.role == \"assistant\" %}\n {{- '<|im_start|>' + message.role }}\n {%- if message.content %}\n {{- '\\n' + message.content }}\n {%- endif %}\n {%- for tool_call in message.tool_calls %}\n {%- if tool_call.function is defined %}\n {%- set tool_call = tool_call.function %}\n {%- endif %}\n {{- '\\n<tool_call>\\n{\"name\": \"' }}\n {{- tool_call.name }}\n {{- '\", \"arguments\": ' }}\n {{- tool_call.arguments | tojson }}\n {{- '}\\n</tool_call>' }}\n {%- endfor %}\n {{- '<|im_end|>\\n' }}\n {%- elif message.role == \"tool\" %}\n {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n {{- '<|im_start|>user' }}\n {%- endif %}\n {{- '\\n<tool_response>\\n' }}\n {{- message.content }}\n {{- '\\n</tool_response>' }}\n {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n {{- '<|im_end|>\\n' }}\n {%- endif %}\n {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
198
+ "clean_up_tokenization_spaces": false,
199
+ "eos_token": "<|im_end|>",
200
+ "errors": "replace",
201
+ "model_max_length": 131072,
202
+ "pad_token": "<|endoftext|>",
203
+ "split_special_tokens": false,
204
+ "tokenizer_class": "Qwen2Tokenizer",
205
+ "unk_token": null,
206
+ "add_bos_token": false
207
+ }
vocab.json ADDED
The diff for this file is too large to render. See raw diff