Text Generation
Transformers
Safetensors
Chinese
English
qwen
conversational
custom_code
yuyijiong commited on
Commit
46aeb47
1 Parent(s): 42389c8

Upload 15 files

Browse files
config.json ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "/data1/yuyijiong/\u81ea\u5df1\u8bad\u7ec3\u7684\u6a21\u578b-merged/qwen-chat-14b-32k-pre-lora-yarn-100step/",
3
+ "architectures": [
4
+ "QWenLMHeadModel"
5
+ ],
6
+ "attn_dropout_prob": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_qwen.QWenConfig",
9
+ "AutoModelForCausalLM": "modeling_qwen_yarn.QWenLMHeadModel"
10
+ },
11
+ "bf16": true,
12
+ "emb_dropout_prob": 0.0,
13
+ "fp16": false,
14
+ "fp32": false,
15
+ "hidden_size": 5120,
16
+ "initializer_range": 0.02,
17
+ "intermediate_size": 27392,
18
+ "kv_channels": 128,
19
+ "layer_norm_epsilon": 1e-06,
20
+ "max_position_embeddings": 8192,
21
+ "model_type": "qwen",
22
+ "no_bias": true,
23
+ "num_attention_heads": 40,
24
+ "num_hidden_layers": 40,
25
+ "onnx_safe": null,
26
+ "rotary_emb_base": 10000,
27
+ "rotary_pct": 1.0,
28
+ "scale_attn_weights": true,
29
+ "seq_length": 2048,
30
+ "softmax_in_fp32": false,
31
+ "tie_word_embeddings": false,
32
+ "tokenizer_class": "QWenTokenizer",
33
+ "torch_dtype": "bfloat16",
34
+ "transformers_version": "4.36.0.dev0",
35
+ "use_cache": true,
36
+ "use_cache_kernel": false,
37
+ "use_cache_quantization": false,
38
+ "use_dynamic_ntk": true,
39
+ "use_flash_attn": true,
40
+ "use_logn_attn": false,
41
+ "vocab_size": 152064
42
+ }
configuration_qwen.py ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ from transformers import PretrainedConfig
7
+
8
+
9
+ class QWenConfig(PretrainedConfig):
10
+ model_type = "qwen"
11
+ keys_to_ignore_at_inference = ["past_key_values"]
12
+
13
+ def __init__(
14
+ self,
15
+ vocab_size=151936,
16
+ hidden_size=4096,
17
+ num_hidden_layers=32,
18
+ num_attention_heads=32,
19
+ emb_dropout_prob=0.0,
20
+ attn_dropout_prob=0.0,
21
+ layer_norm_epsilon=1e-6,
22
+ initializer_range=0.02,
23
+ max_position_embeddings=8192,
24
+ scale_attn_weights=True,
25
+ use_cache=True,
26
+ bf16=False,
27
+ fp16=False,
28
+ fp32=False,
29
+ kv_channels=128,
30
+ rotary_pct=1.0,
31
+ rotary_emb_base=10000,
32
+ use_dynamic_ntk=True,
33
+ use_logn_attn=True,
34
+ use_flash_attn="auto",
35
+ intermediate_size=22016,
36
+ no_bias=True,
37
+ tie_word_embeddings=False,
38
+ use_cache_quantization=False,
39
+ use_cache_kernel=False,
40
+ softmax_in_fp32=False,
41
+ **kwargs,
42
+ ):
43
+ self.vocab_size = vocab_size
44
+ self.hidden_size = hidden_size
45
+ self.intermediate_size = intermediate_size
46
+ self.num_hidden_layers = num_hidden_layers
47
+ self.num_attention_heads = num_attention_heads
48
+ self.emb_dropout_prob = emb_dropout_prob
49
+ self.attn_dropout_prob = attn_dropout_prob
50
+ self.layer_norm_epsilon = layer_norm_epsilon
51
+ self.initializer_range = initializer_range
52
+ self.scale_attn_weights = scale_attn_weights
53
+ self.use_cache = use_cache
54
+ self.max_position_embeddings = max_position_embeddings
55
+ self.bf16 = bf16
56
+ self.fp16 = fp16
57
+ self.fp32 = fp32
58
+ self.kv_channels = kv_channels
59
+ self.rotary_pct = rotary_pct
60
+ self.rotary_emb_base = rotary_emb_base
61
+ self.use_dynamic_ntk = use_dynamic_ntk
62
+ self.use_logn_attn = use_logn_attn
63
+ self.use_flash_attn = use_flash_attn
64
+ self.no_bias = no_bias
65
+ self.use_cache_quantization = use_cache_quantization
66
+ self.use_cache_kernel = use_cache_kernel
67
+ self.softmax_in_fp32 = softmax_in_fp32
68
+ super().__init__(
69
+ tie_word_embeddings=tie_word_embeddings,
70
+ **kwargs
71
+ )
generation_config.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "chat_format": "chatml",
3
+ "do_sample": true,
4
+ "eos_token_id": 151643,
5
+ "max_new_tokens": 512,
6
+ "max_window_size": 6144,
7
+ "pad_token_id": 151643,
8
+ "repetition_penalty": 1.1,
9
+ "top_k": 0,
10
+ "top_p": 0.8,
11
+ "transformers_version": "4.36.0.dev0"
12
+ }
model-00001-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:63cd6cfa0024f2dcb74d7dca7fbf2b2de10f64a14ea309a306a8af448a1f5c29
3
+ size 4919444336
model-00002-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d8a90c3f13c1fd0efdfb524e94d5742373597046dce0dc0002be9195c880747f
3
+ size 4991627864
model-00003-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4b4998828b82dc2cdd79e61cd858b4f5658e285ca78a5b892c1d68a599ec8058
3
+ size 4886749824
model-00004-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3854691b32a2e504b1bad9d6d03f0e5e7bd9872b67409a01017c34b63cdbe5af
3
+ size 4903809664
model-00005-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1d474c18f994ecece6efa52b9dbcbe9556235014d26f91315ccd5147b0c1bade
3
+ size 4903820016
model-00006-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:58f6a718acbe8101a52b575c4af12c2dfb6414ea6657c2b779903d018f37f7e2
3
+ size 3729165312
model.safetensors.index.json ADDED
@@ -0,0 +1,330 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 28334581760
4
+ },
5
+ "weight_map": {
6
+ "lm_head.weight": "model-00006-of-00006.safetensors",
7
+ "transformer.h.0.attn.c_attn.bias": "model-00001-of-00006.safetensors",
8
+ "transformer.h.0.attn.c_attn.weight": "model-00001-of-00006.safetensors",
9
+ "transformer.h.0.attn.c_proj.weight": "model-00001-of-00006.safetensors",
10
+ "transformer.h.0.ln_1.weight": "model-00001-of-00006.safetensors",
11
+ "transformer.h.0.ln_2.weight": "model-00001-of-00006.safetensors",
12
+ "transformer.h.0.mlp.c_proj.weight": "model-00001-of-00006.safetensors",
13
+ "transformer.h.0.mlp.w1.weight": "model-00001-of-00006.safetensors",
14
+ "transformer.h.0.mlp.w2.weight": "model-00001-of-00006.safetensors",
15
+ "transformer.h.1.attn.c_attn.bias": "model-00001-of-00006.safetensors",
16
+ "transformer.h.1.attn.c_attn.weight": "model-00001-of-00006.safetensors",
17
+ "transformer.h.1.attn.c_proj.weight": "model-00001-of-00006.safetensors",
18
+ "transformer.h.1.ln_1.weight": "model-00001-of-00006.safetensors",
19
+ "transformer.h.1.ln_2.weight": "model-00001-of-00006.safetensors",
20
+ "transformer.h.1.mlp.c_proj.weight": "model-00001-of-00006.safetensors",
21
+ "transformer.h.1.mlp.w1.weight": "model-00001-of-00006.safetensors",
22
+ "transformer.h.1.mlp.w2.weight": "model-00001-of-00006.safetensors",
23
+ "transformer.h.10.attn.c_attn.bias": "model-00002-of-00006.safetensors",
24
+ "transformer.h.10.attn.c_attn.weight": "model-00002-of-00006.safetensors",
25
+ "transformer.h.10.attn.c_proj.weight": "model-00002-of-00006.safetensors",
26
+ "transformer.h.10.ln_1.weight": "model-00002-of-00006.safetensors",
27
+ "transformer.h.10.ln_2.weight": "model-00002-of-00006.safetensors",
28
+ "transformer.h.10.mlp.c_proj.weight": "model-00002-of-00006.safetensors",
29
+ "transformer.h.10.mlp.w1.weight": "model-00002-of-00006.safetensors",
30
+ "transformer.h.10.mlp.w2.weight": "model-00002-of-00006.safetensors",
31
+ "transformer.h.11.attn.c_attn.bias": "model-00002-of-00006.safetensors",
32
+ "transformer.h.11.attn.c_attn.weight": "model-00002-of-00006.safetensors",
33
+ "transformer.h.11.attn.c_proj.weight": "model-00002-of-00006.safetensors",
34
+ "transformer.h.11.ln_1.weight": "model-00002-of-00006.safetensors",
35
+ "transformer.h.11.ln_2.weight": "model-00002-of-00006.safetensors",
36
+ "transformer.h.11.mlp.c_proj.weight": "model-00002-of-00006.safetensors",
37
+ "transformer.h.11.mlp.w1.weight": "model-00002-of-00006.safetensors",
38
+ "transformer.h.11.mlp.w2.weight": "model-00002-of-00006.safetensors",
39
+ "transformer.h.12.attn.c_attn.bias": "model-00002-of-00006.safetensors",
40
+ "transformer.h.12.attn.c_attn.weight": "model-00002-of-00006.safetensors",
41
+ "transformer.h.12.attn.c_proj.weight": "model-00002-of-00006.safetensors",
42
+ "transformer.h.12.ln_1.weight": "model-00002-of-00006.safetensors",
43
+ "transformer.h.12.ln_2.weight": "model-00002-of-00006.safetensors",
44
+ "transformer.h.12.mlp.c_proj.weight": "model-00002-of-00006.safetensors",
45
+ "transformer.h.12.mlp.w1.weight": "model-00002-of-00006.safetensors",
46
+ "transformer.h.12.mlp.w2.weight": "model-00002-of-00006.safetensors",
47
+ "transformer.h.13.attn.c_attn.bias": "model-00002-of-00006.safetensors",
48
+ "transformer.h.13.attn.c_attn.weight": "model-00002-of-00006.safetensors",
49
+ "transformer.h.13.attn.c_proj.weight": "model-00003-of-00006.safetensors",
50
+ "transformer.h.13.ln_1.weight": "model-00002-of-00006.safetensors",
51
+ "transformer.h.13.ln_2.weight": "model-00003-of-00006.safetensors",
52
+ "transformer.h.13.mlp.c_proj.weight": "model-00003-of-00006.safetensors",
53
+ "transformer.h.13.mlp.w1.weight": "model-00003-of-00006.safetensors",
54
+ "transformer.h.13.mlp.w2.weight": "model-00003-of-00006.safetensors",
55
+ "transformer.h.14.attn.c_attn.bias": "model-00003-of-00006.safetensors",
56
+ "transformer.h.14.attn.c_attn.weight": "model-00003-of-00006.safetensors",
57
+ "transformer.h.14.attn.c_proj.weight": "model-00003-of-00006.safetensors",
58
+ "transformer.h.14.ln_1.weight": "model-00003-of-00006.safetensors",
59
+ "transformer.h.14.ln_2.weight": "model-00003-of-00006.safetensors",
60
+ "transformer.h.14.mlp.c_proj.weight": "model-00003-of-00006.safetensors",
61
+ "transformer.h.14.mlp.w1.weight": "model-00003-of-00006.safetensors",
62
+ "transformer.h.14.mlp.w2.weight": "model-00003-of-00006.safetensors",
63
+ "transformer.h.15.attn.c_attn.bias": "model-00003-of-00006.safetensors",
64
+ "transformer.h.15.attn.c_attn.weight": "model-00003-of-00006.safetensors",
65
+ "transformer.h.15.attn.c_proj.weight": "model-00003-of-00006.safetensors",
66
+ "transformer.h.15.ln_1.weight": "model-00003-of-00006.safetensors",
67
+ "transformer.h.15.ln_2.weight": "model-00003-of-00006.safetensors",
68
+ "transformer.h.15.mlp.c_proj.weight": "model-00003-of-00006.safetensors",
69
+ "transformer.h.15.mlp.w1.weight": "model-00003-of-00006.safetensors",
70
+ "transformer.h.15.mlp.w2.weight": "model-00003-of-00006.safetensors",
71
+ "transformer.h.16.attn.c_attn.bias": "model-00003-of-00006.safetensors",
72
+ "transformer.h.16.attn.c_attn.weight": "model-00003-of-00006.safetensors",
73
+ "transformer.h.16.attn.c_proj.weight": "model-00003-of-00006.safetensors",
74
+ "transformer.h.16.ln_1.weight": "model-00003-of-00006.safetensors",
75
+ "transformer.h.16.ln_2.weight": "model-00003-of-00006.safetensors",
76
+ "transformer.h.16.mlp.c_proj.weight": "model-00003-of-00006.safetensors",
77
+ "transformer.h.16.mlp.w1.weight": "model-00003-of-00006.safetensors",
78
+ "transformer.h.16.mlp.w2.weight": "model-00003-of-00006.safetensors",
79
+ "transformer.h.17.attn.c_attn.bias": "model-00003-of-00006.safetensors",
80
+ "transformer.h.17.attn.c_attn.weight": "model-00003-of-00006.safetensors",
81
+ "transformer.h.17.attn.c_proj.weight": "model-00003-of-00006.safetensors",
82
+ "transformer.h.17.ln_1.weight": "model-00003-of-00006.safetensors",
83
+ "transformer.h.17.ln_2.weight": "model-00003-of-00006.safetensors",
84
+ "transformer.h.17.mlp.c_proj.weight": "model-00003-of-00006.safetensors",
85
+ "transformer.h.17.mlp.w1.weight": "model-00003-of-00006.safetensors",
86
+ "transformer.h.17.mlp.w2.weight": "model-00003-of-00006.safetensors",
87
+ "transformer.h.18.attn.c_attn.bias": "model-00003-of-00006.safetensors",
88
+ "transformer.h.18.attn.c_attn.weight": "model-00003-of-00006.safetensors",
89
+ "transformer.h.18.attn.c_proj.weight": "model-00003-of-00006.safetensors",
90
+ "transformer.h.18.ln_1.weight": "model-00003-of-00006.safetensors",
91
+ "transformer.h.18.ln_2.weight": "model-00003-of-00006.safetensors",
92
+ "transformer.h.18.mlp.c_proj.weight": "model-00003-of-00006.safetensors",
93
+ "transformer.h.18.mlp.w1.weight": "model-00003-of-00006.safetensors",
94
+ "transformer.h.18.mlp.w2.weight": "model-00003-of-00006.safetensors",
95
+ "transformer.h.19.attn.c_attn.bias": "model-00003-of-00006.safetensors",
96
+ "transformer.h.19.attn.c_attn.weight": "model-00003-of-00006.safetensors",
97
+ "transformer.h.19.attn.c_proj.weight": "model-00003-of-00006.safetensors",
98
+ "transformer.h.19.ln_1.weight": "model-00003-of-00006.safetensors",
99
+ "transformer.h.19.ln_2.weight": "model-00003-of-00006.safetensors",
100
+ "transformer.h.19.mlp.c_proj.weight": "model-00003-of-00006.safetensors",
101
+ "transformer.h.19.mlp.w1.weight": "model-00003-of-00006.safetensors",
102
+ "transformer.h.19.mlp.w2.weight": "model-00003-of-00006.safetensors",
103
+ "transformer.h.2.attn.c_attn.bias": "model-00001-of-00006.safetensors",
104
+ "transformer.h.2.attn.c_attn.weight": "model-00001-of-00006.safetensors",
105
+ "transformer.h.2.attn.c_proj.weight": "model-00001-of-00006.safetensors",
106
+ "transformer.h.2.ln_1.weight": "model-00001-of-00006.safetensors",
107
+ "transformer.h.2.ln_2.weight": "model-00001-of-00006.safetensors",
108
+ "transformer.h.2.mlp.c_proj.weight": "model-00001-of-00006.safetensors",
109
+ "transformer.h.2.mlp.w1.weight": "model-00001-of-00006.safetensors",
110
+ "transformer.h.2.mlp.w2.weight": "model-00001-of-00006.safetensors",
111
+ "transformer.h.20.attn.c_attn.bias": "model-00003-of-00006.safetensors",
112
+ "transformer.h.20.attn.c_attn.weight": "model-00003-of-00006.safetensors",
113
+ "transformer.h.20.attn.c_proj.weight": "model-00003-of-00006.safetensors",
114
+ "transformer.h.20.ln_1.weight": "model-00003-of-00006.safetensors",
115
+ "transformer.h.20.ln_2.weight": "model-00003-of-00006.safetensors",
116
+ "transformer.h.20.mlp.c_proj.weight": "model-00003-of-00006.safetensors",
117
+ "transformer.h.20.mlp.w1.weight": "model-00003-of-00006.safetensors",
118
+ "transformer.h.20.mlp.w2.weight": "model-00003-of-00006.safetensors",
119
+ "transformer.h.21.attn.c_attn.bias": "model-00004-of-00006.safetensors",
120
+ "transformer.h.21.attn.c_attn.weight": "model-00004-of-00006.safetensors",
121
+ "transformer.h.21.attn.c_proj.weight": "model-00004-of-00006.safetensors",
122
+ "transformer.h.21.ln_1.weight": "model-00003-of-00006.safetensors",
123
+ "transformer.h.21.ln_2.weight": "model-00004-of-00006.safetensors",
124
+ "transformer.h.21.mlp.c_proj.weight": "model-00004-of-00006.safetensors",
125
+ "transformer.h.21.mlp.w1.weight": "model-00004-of-00006.safetensors",
126
+ "transformer.h.21.mlp.w2.weight": "model-00004-of-00006.safetensors",
127
+ "transformer.h.22.attn.c_attn.bias": "model-00004-of-00006.safetensors",
128
+ "transformer.h.22.attn.c_attn.weight": "model-00004-of-00006.safetensors",
129
+ "transformer.h.22.attn.c_proj.weight": "model-00004-of-00006.safetensors",
130
+ "transformer.h.22.ln_1.weight": "model-00004-of-00006.safetensors",
131
+ "transformer.h.22.ln_2.weight": "model-00004-of-00006.safetensors",
132
+ "transformer.h.22.mlp.c_proj.weight": "model-00004-of-00006.safetensors",
133
+ "transformer.h.22.mlp.w1.weight": "model-00004-of-00006.safetensors",
134
+ "transformer.h.22.mlp.w2.weight": "model-00004-of-00006.safetensors",
135
+ "transformer.h.23.attn.c_attn.bias": "model-00004-of-00006.safetensors",
136
+ "transformer.h.23.attn.c_attn.weight": "model-00004-of-00006.safetensors",
137
+ "transformer.h.23.attn.c_proj.weight": "model-00004-of-00006.safetensors",
138
+ "transformer.h.23.ln_1.weight": "model-00004-of-00006.safetensors",
139
+ "transformer.h.23.ln_2.weight": "model-00004-of-00006.safetensors",
140
+ "transformer.h.23.mlp.c_proj.weight": "model-00004-of-00006.safetensors",
141
+ "transformer.h.23.mlp.w1.weight": "model-00004-of-00006.safetensors",
142
+ "transformer.h.23.mlp.w2.weight": "model-00004-of-00006.safetensors",
143
+ "transformer.h.24.attn.c_attn.bias": "model-00004-of-00006.safetensors",
144
+ "transformer.h.24.attn.c_attn.weight": "model-00004-of-00006.safetensors",
145
+ "transformer.h.24.attn.c_proj.weight": "model-00004-of-00006.safetensors",
146
+ "transformer.h.24.ln_1.weight": "model-00004-of-00006.safetensors",
147
+ "transformer.h.24.ln_2.weight": "model-00004-of-00006.safetensors",
148
+ "transformer.h.24.mlp.c_proj.weight": "model-00004-of-00006.safetensors",
149
+ "transformer.h.24.mlp.w1.weight": "model-00004-of-00006.safetensors",
150
+ "transformer.h.24.mlp.w2.weight": "model-00004-of-00006.safetensors",
151
+ "transformer.h.25.attn.c_attn.bias": "model-00004-of-00006.safetensors",
152
+ "transformer.h.25.attn.c_attn.weight": "model-00004-of-00006.safetensors",
153
+ "transformer.h.25.attn.c_proj.weight": "model-00004-of-00006.safetensors",
154
+ "transformer.h.25.ln_1.weight": "model-00004-of-00006.safetensors",
155
+ "transformer.h.25.ln_2.weight": "model-00004-of-00006.safetensors",
156
+ "transformer.h.25.mlp.c_proj.weight": "model-00004-of-00006.safetensors",
157
+ "transformer.h.25.mlp.w1.weight": "model-00004-of-00006.safetensors",
158
+ "transformer.h.25.mlp.w2.weight": "model-00004-of-00006.safetensors",
159
+ "transformer.h.26.attn.c_attn.bias": "model-00004-of-00006.safetensors",
160
+ "transformer.h.26.attn.c_attn.weight": "model-00004-of-00006.safetensors",
161
+ "transformer.h.26.attn.c_proj.weight": "model-00004-of-00006.safetensors",
162
+ "transformer.h.26.ln_1.weight": "model-00004-of-00006.safetensors",
163
+ "transformer.h.26.ln_2.weight": "model-00004-of-00006.safetensors",
164
+ "transformer.h.26.mlp.c_proj.weight": "model-00004-of-00006.safetensors",
165
+ "transformer.h.26.mlp.w1.weight": "model-00004-of-00006.safetensors",
166
+ "transformer.h.26.mlp.w2.weight": "model-00004-of-00006.safetensors",
167
+ "transformer.h.27.attn.c_attn.bias": "model-00004-of-00006.safetensors",
168
+ "transformer.h.27.attn.c_attn.weight": "model-00004-of-00006.safetensors",
169
+ "transformer.h.27.attn.c_proj.weight": "model-00004-of-00006.safetensors",
170
+ "transformer.h.27.ln_1.weight": "model-00004-of-00006.safetensors",
171
+ "transformer.h.27.ln_2.weight": "model-00004-of-00006.safetensors",
172
+ "transformer.h.27.mlp.c_proj.weight": "model-00004-of-00006.safetensors",
173
+ "transformer.h.27.mlp.w1.weight": "model-00004-of-00006.safetensors",
174
+ "transformer.h.27.mlp.w2.weight": "model-00004-of-00006.safetensors",
175
+ "transformer.h.28.attn.c_attn.bias": "model-00004-of-00006.safetensors",
176
+ "transformer.h.28.attn.c_attn.weight": "model-00004-of-00006.safetensors",
177
+ "transformer.h.28.attn.c_proj.weight": "model-00004-of-00006.safetensors",
178
+ "transformer.h.28.ln_1.weight": "model-00004-of-00006.safetensors",
179
+ "transformer.h.28.ln_2.weight": "model-00004-of-00006.safetensors",
180
+ "transformer.h.28.mlp.c_proj.weight": "model-00005-of-00006.safetensors",
181
+ "transformer.h.28.mlp.w1.weight": "model-00004-of-00006.safetensors",
182
+ "transformer.h.28.mlp.w2.weight": "model-00004-of-00006.safetensors",
183
+ "transformer.h.29.attn.c_attn.bias": "model-00005-of-00006.safetensors",
184
+ "transformer.h.29.attn.c_attn.weight": "model-00005-of-00006.safetensors",
185
+ "transformer.h.29.attn.c_proj.weight": "model-00005-of-00006.safetensors",
186
+ "transformer.h.29.ln_1.weight": "model-00005-of-00006.safetensors",
187
+ "transformer.h.29.ln_2.weight": "model-00005-of-00006.safetensors",
188
+ "transformer.h.29.mlp.c_proj.weight": "model-00005-of-00006.safetensors",
189
+ "transformer.h.29.mlp.w1.weight": "model-00005-of-00006.safetensors",
190
+ "transformer.h.29.mlp.w2.weight": "model-00005-of-00006.safetensors",
191
+ "transformer.h.3.attn.c_attn.bias": "model-00001-of-00006.safetensors",
192
+ "transformer.h.3.attn.c_attn.weight": "model-00001-of-00006.safetensors",
193
+ "transformer.h.3.attn.c_proj.weight": "model-00001-of-00006.safetensors",
194
+ "transformer.h.3.ln_1.weight": "model-00001-of-00006.safetensors",
195
+ "transformer.h.3.ln_2.weight": "model-00001-of-00006.safetensors",
196
+ "transformer.h.3.mlp.c_proj.weight": "model-00001-of-00006.safetensors",
197
+ "transformer.h.3.mlp.w1.weight": "model-00001-of-00006.safetensors",
198
+ "transformer.h.3.mlp.w2.weight": "model-00001-of-00006.safetensors",
199
+ "transformer.h.30.attn.c_attn.bias": "model-00005-of-00006.safetensors",
200
+ "transformer.h.30.attn.c_attn.weight": "model-00005-of-00006.safetensors",
201
+ "transformer.h.30.attn.c_proj.weight": "model-00005-of-00006.safetensors",
202
+ "transformer.h.30.ln_1.weight": "model-00005-of-00006.safetensors",
203
+ "transformer.h.30.ln_2.weight": "model-00005-of-00006.safetensors",
204
+ "transformer.h.30.mlp.c_proj.weight": "model-00005-of-00006.safetensors",
205
+ "transformer.h.30.mlp.w1.weight": "model-00005-of-00006.safetensors",
206
+ "transformer.h.30.mlp.w2.weight": "model-00005-of-00006.safetensors",
207
+ "transformer.h.31.attn.c_attn.bias": "model-00005-of-00006.safetensors",
208
+ "transformer.h.31.attn.c_attn.weight": "model-00005-of-00006.safetensors",
209
+ "transformer.h.31.attn.c_proj.weight": "model-00005-of-00006.safetensors",
210
+ "transformer.h.31.ln_1.weight": "model-00005-of-00006.safetensors",
211
+ "transformer.h.31.ln_2.weight": "model-00005-of-00006.safetensors",
212
+ "transformer.h.31.mlp.c_proj.weight": "model-00005-of-00006.safetensors",
213
+ "transformer.h.31.mlp.w1.weight": "model-00005-of-00006.safetensors",
214
+ "transformer.h.31.mlp.w2.weight": "model-00005-of-00006.safetensors",
215
+ "transformer.h.32.attn.c_attn.bias": "model-00005-of-00006.safetensors",
216
+ "transformer.h.32.attn.c_attn.weight": "model-00005-of-00006.safetensors",
217
+ "transformer.h.32.attn.c_proj.weight": "model-00005-of-00006.safetensors",
218
+ "transformer.h.32.ln_1.weight": "model-00005-of-00006.safetensors",
219
+ "transformer.h.32.ln_2.weight": "model-00005-of-00006.safetensors",
220
+ "transformer.h.32.mlp.c_proj.weight": "model-00005-of-00006.safetensors",
221
+ "transformer.h.32.mlp.w1.weight": "model-00005-of-00006.safetensors",
222
+ "transformer.h.32.mlp.w2.weight": "model-00005-of-00006.safetensors",
223
+ "transformer.h.33.attn.c_attn.bias": "model-00005-of-00006.safetensors",
224
+ "transformer.h.33.attn.c_attn.weight": "model-00005-of-00006.safetensors",
225
+ "transformer.h.33.attn.c_proj.weight": "model-00005-of-00006.safetensors",
226
+ "transformer.h.33.ln_1.weight": "model-00005-of-00006.safetensors",
227
+ "transformer.h.33.ln_2.weight": "model-00005-of-00006.safetensors",
228
+ "transformer.h.33.mlp.c_proj.weight": "model-00005-of-00006.safetensors",
229
+ "transformer.h.33.mlp.w1.weight": "model-00005-of-00006.safetensors",
230
+ "transformer.h.33.mlp.w2.weight": "model-00005-of-00006.safetensors",
231
+ "transformer.h.34.attn.c_attn.bias": "model-00005-of-00006.safetensors",
232
+ "transformer.h.34.attn.c_attn.weight": "model-00005-of-00006.safetensors",
233
+ "transformer.h.34.attn.c_proj.weight": "model-00005-of-00006.safetensors",
234
+ "transformer.h.34.ln_1.weight": "model-00005-of-00006.safetensors",
235
+ "transformer.h.34.ln_2.weight": "model-00005-of-00006.safetensors",
236
+ "transformer.h.34.mlp.c_proj.weight": "model-00005-of-00006.safetensors",
237
+ "transformer.h.34.mlp.w1.weight": "model-00005-of-00006.safetensors",
238
+ "transformer.h.34.mlp.w2.weight": "model-00005-of-00006.safetensors",
239
+ "transformer.h.35.attn.c_attn.bias": "model-00005-of-00006.safetensors",
240
+ "transformer.h.35.attn.c_attn.weight": "model-00005-of-00006.safetensors",
241
+ "transformer.h.35.attn.c_proj.weight": "model-00005-of-00006.safetensors",
242
+ "transformer.h.35.ln_1.weight": "model-00005-of-00006.safetensors",
243
+ "transformer.h.35.ln_2.weight": "model-00005-of-00006.safetensors",
244
+ "transformer.h.35.mlp.c_proj.weight": "model-00005-of-00006.safetensors",
245
+ "transformer.h.35.mlp.w1.weight": "model-00005-of-00006.safetensors",
246
+ "transformer.h.35.mlp.w2.weight": "model-00005-of-00006.safetensors",
247
+ "transformer.h.36.attn.c_attn.bias": "model-00005-of-00006.safetensors",
248
+ "transformer.h.36.attn.c_attn.weight": "model-00005-of-00006.safetensors",
249
+ "transformer.h.36.attn.c_proj.weight": "model-00005-of-00006.safetensors",
250
+ "transformer.h.36.ln_1.weight": "model-00005-of-00006.safetensors",
251
+ "transformer.h.36.ln_2.weight": "model-00005-of-00006.safetensors",
252
+ "transformer.h.36.mlp.c_proj.weight": "model-00006-of-00006.safetensors",
253
+ "transformer.h.36.mlp.w1.weight": "model-00005-of-00006.safetensors",
254
+ "transformer.h.36.mlp.w2.weight": "model-00006-of-00006.safetensors",
255
+ "transformer.h.37.attn.c_attn.bias": "model-00006-of-00006.safetensors",
256
+ "transformer.h.37.attn.c_attn.weight": "model-00006-of-00006.safetensors",
257
+ "transformer.h.37.attn.c_proj.weight": "model-00006-of-00006.safetensors",
258
+ "transformer.h.37.ln_1.weight": "model-00006-of-00006.safetensors",
259
+ "transformer.h.37.ln_2.weight": "model-00006-of-00006.safetensors",
260
+ "transformer.h.37.mlp.c_proj.weight": "model-00006-of-00006.safetensors",
261
+ "transformer.h.37.mlp.w1.weight": "model-00006-of-00006.safetensors",
262
+ "transformer.h.37.mlp.w2.weight": "model-00006-of-00006.safetensors",
263
+ "transformer.h.38.attn.c_attn.bias": "model-00006-of-00006.safetensors",
264
+ "transformer.h.38.attn.c_attn.weight": "model-00006-of-00006.safetensors",
265
+ "transformer.h.38.attn.c_proj.weight": "model-00006-of-00006.safetensors",
266
+ "transformer.h.38.ln_1.weight": "model-00006-of-00006.safetensors",
267
+ "transformer.h.38.ln_2.weight": "model-00006-of-00006.safetensors",
268
+ "transformer.h.38.mlp.c_proj.weight": "model-00006-of-00006.safetensors",
269
+ "transformer.h.38.mlp.w1.weight": "model-00006-of-00006.safetensors",
270
+ "transformer.h.38.mlp.w2.weight": "model-00006-of-00006.safetensors",
271
+ "transformer.h.39.attn.c_attn.bias": "model-00006-of-00006.safetensors",
272
+ "transformer.h.39.attn.c_attn.weight": "model-00006-of-00006.safetensors",
273
+ "transformer.h.39.attn.c_proj.weight": "model-00006-of-00006.safetensors",
274
+ "transformer.h.39.ln_1.weight": "model-00006-of-00006.safetensors",
275
+ "transformer.h.39.ln_2.weight": "model-00006-of-00006.safetensors",
276
+ "transformer.h.39.mlp.c_proj.weight": "model-00006-of-00006.safetensors",
277
+ "transformer.h.39.mlp.w1.weight": "model-00006-of-00006.safetensors",
278
+ "transformer.h.39.mlp.w2.weight": "model-00006-of-00006.safetensors",
279
+ "transformer.h.4.attn.c_attn.bias": "model-00001-of-00006.safetensors",
280
+ "transformer.h.4.attn.c_attn.weight": "model-00001-of-00006.safetensors",
281
+ "transformer.h.4.attn.c_proj.weight": "model-00001-of-00006.safetensors",
282
+ "transformer.h.4.ln_1.weight": "model-00001-of-00006.safetensors",
283
+ "transformer.h.4.ln_2.weight": "model-00001-of-00006.safetensors",
284
+ "transformer.h.4.mlp.c_proj.weight": "model-00001-of-00006.safetensors",
285
+ "transformer.h.4.mlp.w1.weight": "model-00001-of-00006.safetensors",
286
+ "transformer.h.4.mlp.w2.weight": "model-00001-of-00006.safetensors",
287
+ "transformer.h.5.attn.c_attn.bias": "model-00001-of-00006.safetensors",
288
+ "transformer.h.5.attn.c_attn.weight": "model-00001-of-00006.safetensors",
289
+ "transformer.h.5.attn.c_proj.weight": "model-00001-of-00006.safetensors",
290
+ "transformer.h.5.ln_1.weight": "model-00001-of-00006.safetensors",
291
+ "transformer.h.5.ln_2.weight": "model-00001-of-00006.safetensors",
292
+ "transformer.h.5.mlp.c_proj.weight": "model-00002-of-00006.safetensors",
293
+ "transformer.h.5.mlp.w1.weight": "model-00002-of-00006.safetensors",
294
+ "transformer.h.5.mlp.w2.weight": "model-00002-of-00006.safetensors",
295
+ "transformer.h.6.attn.c_attn.bias": "model-00002-of-00006.safetensors",
296
+ "transformer.h.6.attn.c_attn.weight": "model-00002-of-00006.safetensors",
297
+ "transformer.h.6.attn.c_proj.weight": "model-00002-of-00006.safetensors",
298
+ "transformer.h.6.ln_1.weight": "model-00002-of-00006.safetensors",
299
+ "transformer.h.6.ln_2.weight": "model-00002-of-00006.safetensors",
300
+ "transformer.h.6.mlp.c_proj.weight": "model-00002-of-00006.safetensors",
301
+ "transformer.h.6.mlp.w1.weight": "model-00002-of-00006.safetensors",
302
+ "transformer.h.6.mlp.w2.weight": "model-00002-of-00006.safetensors",
303
+ "transformer.h.7.attn.c_attn.bias": "model-00002-of-00006.safetensors",
304
+ "transformer.h.7.attn.c_attn.weight": "model-00002-of-00006.safetensors",
305
+ "transformer.h.7.attn.c_proj.weight": "model-00002-of-00006.safetensors",
306
+ "transformer.h.7.ln_1.weight": "model-00002-of-00006.safetensors",
307
+ "transformer.h.7.ln_2.weight": "model-00002-of-00006.safetensors",
308
+ "transformer.h.7.mlp.c_proj.weight": "model-00002-of-00006.safetensors",
309
+ "transformer.h.7.mlp.w1.weight": "model-00002-of-00006.safetensors",
310
+ "transformer.h.7.mlp.w2.weight": "model-00002-of-00006.safetensors",
311
+ "transformer.h.8.attn.c_attn.bias": "model-00002-of-00006.safetensors",
312
+ "transformer.h.8.attn.c_attn.weight": "model-00002-of-00006.safetensors",
313
+ "transformer.h.8.attn.c_proj.weight": "model-00002-of-00006.safetensors",
314
+ "transformer.h.8.ln_1.weight": "model-00002-of-00006.safetensors",
315
+ "transformer.h.8.ln_2.weight": "model-00002-of-00006.safetensors",
316
+ "transformer.h.8.mlp.c_proj.weight": "model-00002-of-00006.safetensors",
317
+ "transformer.h.8.mlp.w1.weight": "model-00002-of-00006.safetensors",
318
+ "transformer.h.8.mlp.w2.weight": "model-00002-of-00006.safetensors",
319
+ "transformer.h.9.attn.c_attn.bias": "model-00002-of-00006.safetensors",
320
+ "transformer.h.9.attn.c_attn.weight": "model-00002-of-00006.safetensors",
321
+ "transformer.h.9.attn.c_proj.weight": "model-00002-of-00006.safetensors",
322
+ "transformer.h.9.ln_1.weight": "model-00002-of-00006.safetensors",
323
+ "transformer.h.9.ln_2.weight": "model-00002-of-00006.safetensors",
324
+ "transformer.h.9.mlp.c_proj.weight": "model-00002-of-00006.safetensors",
325
+ "transformer.h.9.mlp.w1.weight": "model-00002-of-00006.safetensors",
326
+ "transformer.h.9.mlp.w2.weight": "model-00002-of-00006.safetensors",
327
+ "transformer.ln_f.weight": "model-00006-of-00006.safetensors",
328
+ "transformer.wte.weight": "model-00001-of-00006.safetensors"
329
+ }
330
+ }
modeling_qwen_yarn.py ADDED
@@ -0,0 +1,1522 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ import importlib
7
+ import math
8
+ from typing import TYPE_CHECKING, Optional, Tuple, Union, Callable, List, Any, Generator
9
+
10
+ import torch
11
+ import torch.nn.functional as F
12
+ import torch.utils.checkpoint
13
+ from torch.cuda.amp import autocast
14
+
15
+ from torch.nn import CrossEntropyLoss
16
+ from transformers import PreTrainedTokenizer, GenerationConfig, StoppingCriteriaList
17
+ from transformers.generation.logits_process import LogitsProcessorList
18
+
19
+ if TYPE_CHECKING:
20
+ from transformers.generation.streamers import BaseStreamer
21
+ from transformers.generation.utils import GenerateOutput
22
+ from transformers.modeling_outputs import (
23
+ BaseModelOutputWithPast,
24
+ CausalLMOutputWithPast,
25
+ )
26
+ from transformers.modeling_utils import PreTrainedModel
27
+ from transformers.utils import logging
28
+
29
+ try:
30
+ from einops import rearrange
31
+ except ImportError:
32
+ rearrange = None
33
+ from torch import nn
34
+
35
+ try:
36
+ from kernels.cpp_kernels import cache_autogptq_cuda_256
37
+ except ImportError:
38
+ cache_autogptq_cuda_256 = None
39
+
40
+ SUPPORT_CUDA = torch.cuda.is_available()
41
+ SUPPORT_BF16 = SUPPORT_CUDA and torch.cuda.is_bf16_supported()
42
+ SUPPORT_FP16 = SUPPORT_CUDA and torch.cuda.get_device_capability(0)[0] >= 7
43
+
44
+ from .configuration_qwen import QWenConfig
45
+ from .qwen_generation_utils import (
46
+ HistoryType,
47
+ make_context,
48
+ decode_tokens,
49
+ get_stop_words_ids,
50
+ StopWordsLogitsProcessor,
51
+ )
52
+ from flash_attn.bert_padding import unpad_input, pad_input
53
+
54
+ logger = logging.get_logger(__name__)
55
+
56
+ _CHECKPOINT_FOR_DOC = "qwen"
57
+ _CONFIG_FOR_DOC = "QWenConfig"
58
+
59
+ QWen_PRETRAINED_MODEL_ARCHIVE_LIST = ["qwen-7b"]
60
+
61
+ _ERROR_BAD_CHAT_FORMAT = """\
62
+ We detect you are probably using the pretrained model (rather than chat model) for chatting, since the chat_format in generation_config is not "chatml".
63
+ If you are directly using the model downloaded from Huggingface, please make sure you are using our "Qwen/Qwen-7B-Chat" Huggingface model (rather than "Qwen/Qwen-7B") when you call model.chat().
64
+ 我们检测到您可能在使用预训练模型(而非chat模型)进行多轮chat,因为您当前在generation_config指定的chat_format,并未设置为我们在对话中所支持的"chatml"格式。
65
+ 如果您在直接使用我们从Huggingface提供的模型,请确保您在调用model.chat()时,使用的是"Qwen/Qwen-7B-Chat"模型(而非"Qwen/Qwen-7B"预训练模型)。
66
+ """
67
+
68
+ _SENTINEL = object()
69
+ _ERROR_STREAM_IN_CHAT = """\
70
+ Pass argument `stream` to model.chat() is buggy, deprecated, and marked for removal. Please use model.chat_stream(...) instead of model.chat(..., stream=True).
71
+ 向model.chat()传入参数stream的用法可能存在Bug,该用法已被废弃,将在未来被移除。请使用model.chat_stream(...)代替model.chat(..., stream=True)。
72
+ """
73
+
74
+ _ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED = """\
75
+ We detect you have activated flash attention support, but running model computation on CPU. Please make sure that your input data has been placed on GPU. If you actually want to run CPU computation, please following the readme and set device_map="cpu" to disable flash attention when loading the model (calling AutoModelForCausalLM.from_pretrained).
76
+ 检测到您的模型已激活了flash attention支持,但正在执行CPU运算任务。如使用flash attention,请您确认模型输入已经传到GPU上。如果您确认要执行CPU运算,请您在载入模型(调用AutoModelForCausalLM.from_pretrained)时,按照readme说法,指定device_map="cpu"以禁用flash attention。
77
+ """
78
+
79
+ apply_rotary_emb_func = None
80
+ rms_norm = None
81
+ flash_attn_unpadded_func = None
82
+
83
+ def _import_flash_attn():
84
+ global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func
85
+ try:
86
+ from flash_attn.layers.rotary import apply_rotary_emb_func as __apply_rotary_emb_func
87
+ apply_rotary_emb_func = __apply_rotary_emb_func
88
+ except ImportError:
89
+ logger.warn(
90
+ "Warning: import flash_attn rotary fail, please install FlashAttention rotary to get higher efficiency "
91
+ "https://github.com/Dao-AILab/flash-attention/tree/main/csrc/rotary"
92
+ )
93
+
94
+ try:
95
+ from flash_attn.ops.rms_norm import rms_norm as __rms_norm
96
+ rms_norm = __rms_norm
97
+ except ImportError:
98
+ logger.warn(
99
+ "Warning: import flash_attn rms_norm fail, please install FlashAttention layer_norm to get higher efficiency "
100
+ "https://github.com/Dao-AILab/flash-attention/tree/main/csrc/layer_norm"
101
+ )
102
+
103
+ try:
104
+ import flash_attn
105
+ if not hasattr(flash_attn, '__version__'):
106
+ from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
107
+ else:
108
+ if int(flash_attn.__version__.split(".")[0]) >= 2:
109
+ from flash_attn.flash_attn_interface import flash_attn_varlen_func as __flash_attn_unpadded_func
110
+ else:
111
+ from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
112
+ flash_attn_unpadded_func = __flash_attn_unpadded_func
113
+ except ImportError:
114
+ logger.warn(
115
+ "Warning: import flash_attn fail, please install FlashAttention to get higher efficiency "
116
+ "https://github.com/Dao-AILab/flash-attention"
117
+ )
118
+
119
+ def quantize_cache_v(fdata, bits, qmax, qmin):
120
+ # b, s, head, h-dim->b, head, s, h-dim
121
+ qtype = torch.uint8
122
+ device = fdata.device
123
+ shape = fdata.shape
124
+
125
+ fdata_cal = torch.flatten(fdata, 2)
126
+ fmax = torch.amax(fdata_cal, dim=-1, keepdim=True)
127
+ fmin = torch.amin(fdata_cal, dim=-1, keepdim=True)
128
+ # Compute params
129
+ if qmax.device != fmax.device:
130
+ qmax = qmax.to(device)
131
+ qmin = qmin.to(device)
132
+ scale = (fmax - fmin) / (qmax - qmin)
133
+ zero = qmin - fmin / scale
134
+ scale = scale.unsqueeze(-1).repeat(1,1,shape[2],1).contiguous()
135
+ zero = zero.unsqueeze(-1).repeat(1,1,shape[2],1).contiguous()
136
+ # Quantize
137
+ res_data = fdata / scale + zero
138
+ qdata = torch.clamp(res_data, qmin, qmax).to(qtype)
139
+ return qdata.contiguous(), scale, zero
140
+
141
+ def dequantize_cache_torch(qdata, scale, zero):
142
+ data = scale * (qdata - zero)
143
+ return data
144
+
145
+ class FlashSelfAttention(torch.nn.Module):
146
+ def __init__(
147
+ self,
148
+ causal=False,
149
+ softmax_scale=None,
150
+ attention_dropout=0.0,
151
+ ):
152
+ super().__init__()
153
+ assert flash_attn_unpadded_func is not None, (
154
+ "Please install FlashAttention first, " "e.g., with pip install flash-attn"
155
+ )
156
+ assert (
157
+ rearrange is not None
158
+ ), "Please install einops first, e.g., with pip install einops"
159
+ self.causal = causal
160
+ self.softmax_scale = softmax_scale
161
+ self.dropout_p = attention_dropout
162
+
163
+ def unpad_input(self, hidden_states, attention_mask):
164
+ valid_mask = attention_mask.squeeze(1).squeeze(1).eq(0)
165
+ seqlens_in_batch = valid_mask.sum(dim=-1, dtype=torch.int32)
166
+ indices = torch.nonzero(valid_mask.flatten(), as_tuple=False).flatten()
167
+ max_seqlen_in_batch = seqlens_in_batch.max().item()
168
+ cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
169
+ hidden_states = hidden_states[indices]
170
+ return hidden_states, indices, cu_seqlens, max_seqlen_in_batch
171
+
172
+ def pad_input(self, hidden_states, indices, batch, seqlen):
173
+ #torch.cuda.empty_cache()
174
+ output = torch.zeros(batch * seqlen, *hidden_states.shape[1:], device=hidden_states.device,
175
+ dtype=hidden_states.dtype)
176
+ output[indices] = hidden_states
177
+ return rearrange(output, '(b s) ... -> b s ...', b=batch)
178
+
179
+ def forward(self, q, k, v, attention_mask=None):
180
+ assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
181
+ assert all((i.is_cuda for i in (q, k, v)))
182
+ q_len_origin=q.shape[1]
183
+ batch_size, seqlen_q = q.shape[0], q.shape[1]
184
+ seqlen_k = k.shape[1]
185
+
186
+ q, k, v = [rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v]]
187
+ cu_seqlens_q = torch.arange(
188
+ 0,
189
+ (batch_size + 1) * seqlen_q,
190
+ step=seqlen_q,
191
+ dtype=torch.int32,
192
+ device=q.device,
193
+ )
194
+
195
+ if attention_mask is not None:
196
+ k, indices_k, cu_seqlens_k, seqlen_k = self.unpad_input(k, attention_mask)
197
+ v = v[indices_k]
198
+ if self.training or q.size(0) == k.size(0):
199
+ q = q[indices_k]
200
+ cu_seqlens_q = cu_seqlens_k
201
+ seqlen_q = seqlen_k
202
+ else:
203
+ cu_seqlens_k = torch.arange(
204
+ 0,
205
+ (batch_size + 1) * seqlen_k,
206
+ step=seqlen_k,
207
+ dtype=torch.int32,
208
+ device=q.device,
209
+ )
210
+
211
+ if self.training:
212
+ assert seqlen_k == seqlen_q
213
+ is_causal = self.causal
214
+ dropout_p = self.dropout_p
215
+ else:
216
+ is_causal = seqlen_q == seqlen_k
217
+ dropout_p = 0
218
+
219
+ output = flash_attn_unpadded_func(
220
+ q,
221
+ k,
222
+ v,
223
+ cu_seqlens_q,
224
+ cu_seqlens_k,
225
+ seqlen_q,
226
+ seqlen_k,
227
+ dropout_p,
228
+ softmax_scale=self.softmax_scale,
229
+ causal=is_causal,
230
+ )
231
+ if attention_mask is not None and seqlen_q == seqlen_k:
232
+ output = self.pad_input(output, indices_k, batch_size, seqlen_q)
233
+
234
+
235
+ else:
236
+ new_shape = (batch_size, output.shape[0] // batch_size) + output.shape[1:]
237
+ output = output.view(new_shape)
238
+ return output
239
+
240
+
241
+ class QWenAttention(nn.Module):
242
+ def __init__(self, config):
243
+ super().__init__()
244
+
245
+ self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)
246
+ self.seq_length = config.seq_length
247
+
248
+ self.hidden_size = config.hidden_size
249
+ self.split_size = config.hidden_size
250
+ self.num_heads = config.num_attention_heads
251
+ self.head_dim = self.hidden_size // self.num_heads
252
+
253
+ self.use_flash_attn = config.use_flash_attn
254
+ self.scale_attn_weights = True
255
+
256
+ self.projection_size = config.kv_channels * config.num_attention_heads
257
+
258
+ assert self.projection_size % config.num_attention_heads == 0
259
+ self.hidden_size_per_attention_head = (
260
+ self.projection_size // config.num_attention_heads
261
+ )
262
+
263
+ self.c_attn = nn.Linear(config.hidden_size, 3 * self.projection_size)
264
+
265
+ self.c_proj = nn.Linear(
266
+ config.hidden_size, self.projection_size, bias=not config.no_bias
267
+ )
268
+
269
+ self.is_fp32 = not (config.bf16 or config.fp16)
270
+ if (
271
+ self.use_flash_attn
272
+ and flash_attn_unpadded_func is not None
273
+ and not self.is_fp32
274
+ ):
275
+ self.core_attention_flash = FlashSelfAttention(
276
+ causal=True, attention_dropout=config.attn_dropout_prob
277
+ )
278
+ self.bf16 = config.bf16
279
+
280
+ self.use_dynamic_ntk = config.use_dynamic_ntk
281
+ self.use_logn_attn = config.use_logn_attn
282
+
283
+ logn_list = [
284
+ math.log(i, self.seq_length) if i > self.seq_length else 1
285
+ for i in range(1, 32768)
286
+ ]
287
+ logn_tensor = torch.tensor(logn_list)[None, :, None, None]
288
+ self.register_buffer("logn_tensor", logn_tensor, persistent=False)
289
+
290
+ self.attn_dropout = nn.Dropout(config.attn_dropout_prob)
291
+ self.use_cache_quantization = config.use_cache_quantization if hasattr(config, 'use_cache_quantization') else False
292
+ self.use_cache_kernel = config.use_cache_kernel if hasattr(config,'use_cache_kernel') else False
293
+ cache_dtype = torch.float
294
+ if self.bf16:
295
+ cache_dtype=torch.bfloat16
296
+ elif config.fp16:
297
+ cache_dtype = torch.float16
298
+ self.cache_qmax = torch.tensor(torch.iinfo(torch.uint8).max, dtype=cache_dtype)
299
+ self.cache_qmin = torch.tensor(torch.iinfo(torch.uint8).min, dtype=cache_dtype)
300
+
301
+ def _attn(self, query, key, value, registered_causal_mask, attention_mask=None, head_mask=None):
302
+ device = query.device
303
+ if self.use_cache_quantization:
304
+ qk, qk_scale, qk_zero = key
305
+ if self.use_cache_kernel and cache_autogptq_cuda_256 is not None:
306
+ shape = query.shape[:-1] + (qk.shape[-2],)
307
+ attn_weights = torch.zeros(shape, dtype=torch.float16, device=device)
308
+ cache_autogptq_cuda_256.vecquant8matmul_batched_faster_old(
309
+ query.contiguous() if query.dtype == torch.float16 else query.to(torch.float16).contiguous(),
310
+ qk.transpose(-1, -2).contiguous(),
311
+ attn_weights,
312
+ qk_scale.contiguous() if qk_scale.dtype == torch.float16 else qk_scale.to(torch.float16).contiguous(),
313
+ qk_zero.contiguous()if qk_zero.dtype == torch.float16 else qk_zero.to(torch.float16).contiguous())
314
+ # attn_weights = attn_weights.to(query.dtype).contiguous()
315
+ else:
316
+ key = dequantize_cache_torch(qk, qk_scale, qk_zero)
317
+ attn_weights = torch.matmul(query, key.transpose(-1, -2))
318
+ else:
319
+ attn_weights = torch.matmul(query, key.transpose(-1, -2))
320
+
321
+ if self.scale_attn_weights:
322
+ if self.use_cache_quantization:
323
+ size_temp = value[0].size(-1)
324
+ else:
325
+ size_temp = value.size(-1)
326
+ attn_weights = attn_weights / torch.full(
327
+ [],
328
+ size_temp ** 0.5,
329
+ dtype=attn_weights.dtype,
330
+ device=attn_weights.device,
331
+ )
332
+ if self.use_cache_quantization:
333
+ query_length, key_length = query.size(-2), key[0].size(-2)
334
+ else:
335
+ query_length, key_length = query.size(-2), key.size(-2)
336
+ causal_mask = registered_causal_mask[
337
+ :, :, key_length - query_length : key_length, :key_length
338
+ ]
339
+ mask_value = torch.finfo(attn_weights.dtype).min
340
+ mask_value = torch.full([], mask_value, dtype=attn_weights.dtype).to(
341
+ attn_weights.device
342
+ )
343
+ attn_weights = torch.where(
344
+ causal_mask, attn_weights.to(attn_weights.dtype), mask_value
345
+ )
346
+
347
+ if attention_mask is not None:
348
+ attn_weights = attn_weights + attention_mask
349
+
350
+ attn_weights = nn.functional.softmax(attn_weights.float(), dim=-1)
351
+
352
+ attn_weights = attn_weights.type(query.dtype)
353
+ attn_weights = self.attn_dropout(attn_weights)
354
+
355
+ if head_mask is not None:
356
+ attn_weights = attn_weights * head_mask
357
+
358
+ if self.use_cache_quantization:
359
+ qv, qv_scale, qv_zero = value
360
+ if self.use_cache_kernel and cache_autogptq_cuda_256 is not None:
361
+ shape = attn_weights.shape[:-1] + (query.shape[-1],)
362
+ attn_output = torch.zeros(shape, dtype=torch.float16, device=device)
363
+ cache_autogptq_cuda_256.vecquant8matmul_batched_column_compression_faster_old(
364
+ attn_weights.contiguous() if attn_weights.dtype == torch.float16 else attn_weights.to(torch.float16).contiguous(),
365
+ qv.contiguous(), # dtype: int32
366
+ attn_output,
367
+ qv_scale.contiguous() if qv_scale.dtype == torch.float16 else qv_scale.to(torch.float16).contiguous(),
368
+ qv_zero.contiguous() if qv_zero.dtype == torch.float16 else qv_zero.to(torch.float16).contiguous())
369
+ if attn_output.dtype != query.dtype:
370
+ attn_output = attn_output.to(query.dtype)
371
+ attn_weights = attn_weights.to(query.dtype)
372
+ else:
373
+ value = dequantize_cache_torch(qv, qv_scale, qv_zero)
374
+ attn_output = torch.matmul(attn_weights, value)
375
+ else:
376
+ attn_output = torch.matmul(attn_weights, value)
377
+
378
+ attn_output = attn_output.transpose(1, 2)
379
+
380
+ return attn_output, attn_weights
381
+
382
+ def _upcast_and_reordered_attn(
383
+ self, query, key, value, registered_causal_mask, attention_mask=None, head_mask=None
384
+ ):
385
+ bsz, num_heads, q_seq_len, dk = query.size()
386
+ _, _, k_seq_len, _ = key.size()
387
+
388
+ attn_weights = torch.empty(
389
+ bsz * num_heads,
390
+ q_seq_len,
391
+ k_seq_len,
392
+ dtype=torch.float32,
393
+ device=query.device,
394
+ )
395
+
396
+ scale_factor = 1.0
397
+ if self.scale_attn_weights:
398
+ scale_factor /= float(value.size(-1)) ** 0.5
399
+
400
+ with autocast(enabled=False):
401
+ q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(
402
+ -1, dk, k_seq_len
403
+ )
404
+ attn_weights = torch.baddbmm(
405
+ attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor
406
+ )
407
+ attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
408
+
409
+ query_length, key_length = query.size(-2), key.size(-2)
410
+ causal_mask = registered_causal_mask[
411
+ :, :, key_length - query_length : key_length, :key_length
412
+ ] #registered_causal_mask是shape为(1, 1, max_positions, max_positions)的二值张量,其中max_positions为8192
413
+ mask_value = torch.finfo(attn_weights.dtype).min
414
+ mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(
415
+ attn_weights.device
416
+ )
417
+ attn_weights = torch.where(causal_mask, attn_weights, mask_value) #causal_mask中为1的位置,attn_weights中保留,否则用mask_value填充
418
+
419
+ if attention_mask is not None:
420
+ attn_weights = attn_weights + attention_mask
421
+
422
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1)
423
+
424
+ if attn_weights.dtype != torch.float32:
425
+ raise RuntimeError(
426
+ "Error with upcasting, attn_weights does not have dtype torch.float32"
427
+ )
428
+ attn_weights = attn_weights.type(value.dtype)
429
+ attn_weights = self.attn_dropout(attn_weights)
430
+
431
+ if head_mask is not None:
432
+ attn_weights = attn_weights * head_mask
433
+
434
+ attn_output = torch.matmul(attn_weights, value)
435
+
436
+ return attn_output, attn_weights
437
+
438
+ def _split_heads(self, tensor, num_heads, attn_head_size):
439
+ new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
440
+ tensor = tensor.view(new_shape)
441
+ return tensor
442
+
443
+ def _merge_heads(self, tensor, num_heads, attn_head_size):
444
+ tensor = tensor.contiguous()
445
+ new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
446
+ return tensor.view(new_shape)
447
+
448
+ def forward(
449
+ self,
450
+ hidden_states: Optional[Tuple[torch.FloatTensor]],
451
+ rotary_pos_emb_list: Optional[List[torch.Tensor]] = None,
452
+ registered_causal_mask: Optional[torch.Tensor] = None,
453
+ layer_past: Optional[Tuple[torch.Tensor]] = None,
454
+ attention_mask: Optional[torch.FloatTensor] = None,
455
+ head_mask: Optional[torch.FloatTensor] = None,
456
+ encoder_hidden_states: Optional[torch.Tensor] = None,
457
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
458
+ output_attentions: Optional[bool] = False,
459
+ use_cache: Optional[bool] = False,
460
+ ):
461
+ mixed_x_layer = self.c_attn(hidden_states)
462
+
463
+ query, key, value = mixed_x_layer.split(self.split_size, dim=2)
464
+
465
+ query = self._split_heads(query, self.num_heads, self.head_dim)
466
+ key = self._split_heads(key, self.num_heads, self.head_dim)
467
+ value = self._split_heads(value, self.num_heads, self.head_dim)
468
+
469
+ if rotary_pos_emb_list is not None:
470
+ cur_len = query.shape[1]
471
+ if len(rotary_pos_emb_list) == 1:
472
+ rotary_pos_emb = rotary_pos_emb_list[0]
473
+ rotary_pos_emb = [i[:, -cur_len:, :, :] for i in rotary_pos_emb]
474
+ rotary_pos_emb = (rotary_pos_emb,) * 2
475
+ q_pos_emb, k_pos_emb = rotary_pos_emb
476
+ # Slice the pos emb for current inference
477
+ query = apply_rotary_pos_emb(query, q_pos_emb)
478
+ key = apply_rotary_pos_emb(key, k_pos_emb)
479
+ else:
480
+ query_list = []
481
+ key_list = []
482
+ for i, rotary_pos_emb in enumerate(rotary_pos_emb_list):
483
+ rotary_pos_emb = [i[:, -cur_len:, :, :] for i in rotary_pos_emb]
484
+ rotary_pos_emb = (rotary_pos_emb,) * 2
485
+ q_pos_emb, k_pos_emb = rotary_pos_emb
486
+ # Slice the pos emb for current inference
487
+ query_list += [apply_rotary_pos_emb(query[i:i+1, :, :], q_pos_emb)]
488
+ key_list += [apply_rotary_pos_emb(key[i:i+1, :, :], k_pos_emb)]
489
+ query = torch.cat(query_list, dim=0)
490
+ key = torch.cat(key_list, dim=0)
491
+
492
+ if self.use_cache_quantization:
493
+ key = quantize_cache_v(key.permute(0, 2, 1, 3),
494
+ bits=8,
495
+ qmin=self.cache_qmin,
496
+ qmax=self.cache_qmax)
497
+ value = quantize_cache_v(value.permute(0, 2, 1, 3),
498
+ bits=8,
499
+ qmin=self.cache_qmin,
500
+ qmax=self.cache_qmax)
501
+
502
+
503
+ if layer_past is not None:
504
+ past_key, past_value = layer_past[0], layer_past[1]
505
+ if self.use_cache_quantization:
506
+ # use_cache_quantization:
507
+ # present=((q_key,key_scale,key_zero_point),
508
+ # (q_value,value_scale,value_zero_point))
509
+ key = (torch.cat((past_key[0], key[0]), dim=2),
510
+ torch.cat((past_key[1], key[1]), dim=2),
511
+ torch.cat((past_key[2], key[2]), dim=2))
512
+ value = (torch.cat((past_value[0], value[0]), dim=2),
513
+ torch.cat((past_value[1], value[1]), dim=2),
514
+ torch.cat((past_value[2], value[2]), dim=2))
515
+ else:
516
+ # not use_cache_quantization:
517
+ # present=(key,value)
518
+ key = torch.cat((past_key, key), dim=1)
519
+ value = torch.cat((past_value, value), dim=1)
520
+
521
+ if use_cache:
522
+ present = (key, value)
523
+ else:
524
+ present = None
525
+
526
+ if self.use_logn_attn and not self.training:
527
+ if self.use_cache_quantization:
528
+ seq_start = key[0].size(2) - query.size(1)
529
+ seq_end = key[0].size(2)
530
+ else:
531
+ seq_start = key.size(1) - query.size(1)
532
+ seq_end = key.size(1)
533
+ logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :]
534
+ query = query * logn_tensor.expand_as(query) #使用logn_attn时,query乘以logn_tensor,避免长文本时注意力不稳定
535
+
536
+ if (
537
+ self.use_flash_attn
538
+ and flash_attn_unpadded_func is not None
539
+ and not self.is_fp32
540
+ and query.is_cuda
541
+ ):
542
+ q, k, v = query, key, value
543
+ context_layer = self.core_attention_flash(q, k, v, attention_mask=attention_mask)
544
+
545
+ # b s h d -> b s (h d)
546
+ context_layer = context_layer.flatten(2,3).contiguous()
547
+
548
+ else:
549
+ query = query.permute(0, 2, 1, 3)
550
+ if not self.use_cache_quantization:
551
+ key = key.permute(0, 2, 1, 3)
552
+ value = value.permute(0, 2, 1, 3)
553
+ if (
554
+ registered_causal_mask is None
555
+ and self.use_flash_attn
556
+ and flash_attn_unpadded_func is not None
557
+ and not self.is_fp32
558
+ and not query.is_cuda
559
+ ):
560
+ raise Exception(_ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED)
561
+ attn_output, attn_weight = self._attn(
562
+ query, key, value, registered_causal_mask, attention_mask, head_mask
563
+ ) #如果没有使用flash attention,才会使用registered_causal_mask
564
+ context_layer = self._merge_heads(
565
+ attn_output, self.num_heads, self.head_dim
566
+ )
567
+
568
+ attn_output = self.c_proj(context_layer)
569
+
570
+ outputs = (attn_output, present)
571
+ if output_attentions:
572
+ if (
573
+ self.use_flash_attn
574
+ and flash_attn_unpadded_func is not None
575
+ and not self.is_fp32
576
+ ):
577
+ raise ValueError("Cannot output attentions while using flash-attn")
578
+ else:
579
+ outputs += (attn_weight,)
580
+
581
+ return outputs
582
+
583
+
584
+ class QWenMLP(nn.Module):
585
+ def __init__(self, config):
586
+ super().__init__()
587
+ self.w1 = nn.Linear(
588
+ config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
589
+ )
590
+ self.w2 = nn.Linear(
591
+ config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
592
+ )
593
+ ff_dim_in = config.intermediate_size // 2
594
+ self.c_proj = nn.Linear(ff_dim_in, config.hidden_size, bias=not config.no_bias)
595
+
596
+ def forward(self, hidden_states):
597
+ a1 = self.w1(hidden_states)
598
+ a2 = self.w2(hidden_states)
599
+ intermediate_parallel = a1 * F.silu(a2)
600
+ output = self.c_proj(intermediate_parallel)
601
+ return output
602
+
603
+ class QWenBlock(nn.Module):
604
+ def __init__(self, config):
605
+ super().__init__()
606
+ hidden_size = config.hidden_size
607
+ self.bf16 = config.bf16
608
+
609
+ self.ln_1 = RMSNorm(
610
+ hidden_size,
611
+ eps=config.layer_norm_epsilon,
612
+ )
613
+ self.attn = QWenAttention(config)
614
+ self.ln_2 = RMSNorm(
615
+ hidden_size,
616
+ eps=config.layer_norm_epsilon,
617
+ )
618
+
619
+ self.mlp = QWenMLP(config)
620
+
621
+ def forward(
622
+ self,
623
+ hidden_states: Optional[Tuple[torch.FloatTensor]],
624
+ rotary_pos_emb_list: Optional[List[torch.Tensor]] = None,
625
+ registered_causal_mask: Optional[torch.Tensor] = None,
626
+ layer_past: Optional[Tuple[torch.Tensor]] = None,
627
+ attention_mask: Optional[torch.FloatTensor] = None,
628
+ head_mask: Optional[torch.FloatTensor] = None,
629
+ encoder_hidden_states: Optional[torch.Tensor] = None,
630
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
631
+ use_cache: Optional[bool] = False,
632
+ output_attentions: Optional[bool] = False,
633
+ ):
634
+ layernorm_output = self.ln_1(hidden_states)
635
+
636
+ attn_outputs = self.attn(
637
+ layernorm_output,
638
+ rotary_pos_emb_list,
639
+ registered_causal_mask=registered_causal_mask,
640
+ layer_past=layer_past,
641
+ attention_mask=attention_mask,
642
+ head_mask=head_mask,
643
+ use_cache=use_cache,
644
+ output_attentions=output_attentions,
645
+ )
646
+ attn_output = attn_outputs[0]
647
+
648
+ outputs = attn_outputs[1:]
649
+
650
+ residual = hidden_states
651
+ layernorm_input = attn_output + residual
652
+
653
+ layernorm_output = self.ln_2(layernorm_input)
654
+
655
+ residual = layernorm_input
656
+ mlp_output = self.mlp(layernorm_output)
657
+ hidden_states = residual + mlp_output
658
+
659
+ if use_cache:
660
+ outputs = (hidden_states,) + outputs
661
+ else:
662
+ outputs = (hidden_states,) + outputs[1:]
663
+
664
+ return outputs
665
+
666
+
667
+ class QWenPreTrainedModel(PreTrainedModel):
668
+ config_class = QWenConfig
669
+ base_model_prefix = "transformer"
670
+ is_parallelizable = False
671
+ supports_gradient_checkpointing = True
672
+ _no_split_modules = ["QWenBlock"]
673
+
674
+ def __init__(self, *inputs, **kwargs):
675
+ super().__init__(*inputs, **kwargs)
676
+
677
+ def _init_weights(self, module):
678
+ """Initialize the weights."""
679
+ if isinstance(module, nn.Linear):
680
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
681
+ if module.bias is not None:
682
+ module.bias.data.zero_()
683
+ elif isinstance(module, nn.Embedding):
684
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
685
+ if module.padding_idx is not None:
686
+ module.weight.data[module.padding_idx].zero_()
687
+ elif isinstance(module, RMSNorm):
688
+ module.weight.data.fill_(1.0)
689
+
690
+ for name, p in module.named_parameters():
691
+ if name == "c_proj.weight":
692
+ p.data.normal_(
693
+ mean=0.0,
694
+ std=(
695
+ self.config.initializer_range
696
+ / math.sqrt(2 * self.config.num_hidden_layers)
697
+ ),
698
+ )
699
+
700
+ def _set_gradient_checkpointing(self, module, value=False):
701
+ if isinstance(module, QWenModel):
702
+ module.gradient_checkpointing = value
703
+
704
+
705
+ class QWenModel(QWenPreTrainedModel):
706
+ _keys_to_ignore_on_load_missing = ["attn.masked_bias"]
707
+
708
+ def __init__(self, config):
709
+ super().__init__(config)
710
+ self.vocab_size = config.vocab_size
711
+ self.num_hidden_layers = config.num_hidden_layers
712
+ self.embed_dim = config.hidden_size
713
+ self.use_cache_quantization = self.config.use_cache_quantization if hasattr(self.config, 'use_cache_quantization') else False
714
+
715
+ self.gradient_checkpointing = False
716
+ self.use_dynamic_ntk = config.use_dynamic_ntk
717
+ self.seq_length = config.seq_length
718
+
719
+ self.wte = nn.Embedding(self.vocab_size, self.embed_dim)
720
+
721
+ self.drop = nn.Dropout(config.emb_dropout_prob)
722
+
723
+ if config.rotary_pct == 1.0:
724
+ self.rotary_ndims = None
725
+ else:
726
+ assert config.rotary_pct < 1
727
+ self.rotary_ndims = int(
728
+ config.kv_channels * config.rotary_pct
729
+ )
730
+ dim = (
731
+ self.rotary_ndims
732
+ if self.rotary_ndims is not None
733
+ else config.kv_channels
734
+ )
735
+ self.rotary_emb = YaRNRotaryEmbedding(dim, base=config.rotary_emb_base,original_max_position_embeddings=config.seq_length)
736
+ #self.rotary_emb = RotaryEmbedding(dim, base=config.rotary_emb_base)
737
+ import warnings
738
+ warnings.warn("使用YarnRotaryEmbedding,强制设置config.use_logn_attn = False,config.use_dynamic_ntk = True")
739
+ config.use_logn_attn = False
740
+ config.use_dynamic_ntk = True
741
+
742
+ self.use_flash_attn = config.use_flash_attn
743
+ self.is_fp32 = not (config.bf16 or config.fp16)
744
+ if (
745
+ self.use_flash_attn
746
+ and flash_attn_unpadded_func is not None
747
+ and not self.is_fp32
748
+ ):
749
+ self.registered_causal_mask = None
750
+ else:
751
+ max_positions = config.max_position_embeddings
752
+ self.register_buffer(
753
+ "registered_causal_mask",
754
+ torch.tril(
755
+ torch.ones((max_positions, max_positions), dtype=torch.bool)
756
+ ).view(1, 1, max_positions, max_positions),
757
+ persistent=False,
758
+ )
759
+
760
+ self.h = nn.ModuleList(
761
+ [
762
+ QWenBlock(
763
+ config
764
+ )
765
+ for i in range(config.num_hidden_layers)
766
+ ]
767
+ )
768
+ self.ln_f = RMSNorm(
769
+ self.embed_dim,
770
+ eps=config.layer_norm_epsilon,
771
+ )
772
+
773
+ self.post_init()
774
+
775
+ def get_input_embeddings(self):
776
+ return self.wte
777
+
778
+ def set_input_embeddings(self, new_embeddings):
779
+ self.wte = new_embeddings
780
+
781
+ def get_ntk_alpha(self, true_seq_len):
782
+ ntk_alpha = true_seq_len / self.seq_length + 512/self.seq_length
783
+ ntk_alpha = max(ntk_alpha, 1)
784
+ # context_value = math.log(true_seq_len / self.seq_length, 2) + 1
785
+ # ntk_alpha = 2 ** math.ceil(context_value) - 1
786
+ # ntk_alpha = max(ntk_alpha, 1)
787
+
788
+ #假设seq_length=2k,true_seq_len=8k,context_value=3,ntk_alpha=7,相当于扩展到14k长度
789
+ return ntk_alpha
790
+
791
+ def forward(
792
+ self,
793
+ input_ids: Optional[torch.LongTensor] = None,
794
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
795
+ attention_mask: Optional[torch.FloatTensor] = None,
796
+ token_type_ids: Optional[torch.LongTensor] = None,
797
+ position_ids: Optional[torch.LongTensor] = None,
798
+ head_mask: Optional[torch.FloatTensor] = None,
799
+ inputs_embeds: Optional[torch.FloatTensor] = None,
800
+ encoder_hidden_states: Optional[torch.Tensor] = None,
801
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
802
+ use_cache: Optional[bool] = None,
803
+ output_attentions: Optional[bool] = None,
804
+ output_hidden_states: Optional[bool] = None,
805
+ return_dict: Optional[bool] = None,
806
+ ):
807
+ output_attentions = (
808
+ output_attentions
809
+ if output_attentions is not None
810
+ else self.config.output_attentions
811
+ )
812
+ output_hidden_states = (
813
+ output_hidden_states
814
+ if output_hidden_states is not None
815
+ else self.config.output_hidden_states
816
+ )
817
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
818
+ return_dict = (
819
+ return_dict if return_dict is not None else self.config.use_return_dict
820
+ )
821
+
822
+ if input_ids is not None and inputs_embeds is not None:
823
+ raise ValueError(
824
+ "You cannot specify both input_ids and inputs_embeds at the same time"
825
+ )
826
+ elif input_ids is not None:
827
+ input_shape = input_ids.size()
828
+ input_ids = input_ids.view(-1, input_shape[-1])
829
+ batch_size = input_ids.shape[0]
830
+ elif inputs_embeds is not None:
831
+ input_shape = inputs_embeds.size()[:-1]
832
+ batch_size = inputs_embeds.shape[0]
833
+ else:
834
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
835
+
836
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
837
+
838
+ if token_type_ids is not None:
839
+ token_type_ids = token_type_ids.view(-1, input_shape[-1])
840
+ if position_ids is not None:
841
+ position_ids = position_ids.view(-1, input_shape[-1])
842
+
843
+ if past_key_values is None:
844
+ past_length = 0
845
+ past_key_values = tuple([None] * len(self.h))
846
+ else:
847
+ if self.use_cache_quantization:
848
+ past_length = past_key_values[0][0][0].size(2)
849
+ else:
850
+ past_length = past_key_values[0][0].size(-2)
851
+ if position_ids is None:
852
+ position_ids = torch.arange(
853
+ past_length,
854
+ input_shape[-1] + past_length,
855
+ dtype=torch.long,
856
+ device=device,
857
+ )
858
+ position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
859
+
860
+ if attention_mask is not None:
861
+ if batch_size <= 0:
862
+ raise ValueError("batch_size has to be defined and > 0")
863
+ attention_mask = attention_mask.view(batch_size, -1)
864
+ attention_mask = attention_mask[:, None, None, :]
865
+ attention_mask = attention_mask.to(dtype=self.dtype)
866
+ attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
867
+
868
+ encoder_attention_mask = None
869
+ head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
870
+
871
+ if inputs_embeds is None:
872
+ inputs_embeds = self.wte(input_ids)
873
+ hidden_states = inputs_embeds
874
+
875
+ kv_seq_len = hidden_states.size()[1]
876
+ if past_key_values[0] is not None:
877
+ # past key values[0][0] shape: bs * seq_len * head_num * dim
878
+ if self.use_cache_quantization:
879
+ kv_seq_len += past_key_values[0][0][0].shape[2]
880
+ else:
881
+ kv_seq_len += past_key_values[0][0].shape[1]
882
+
883
+ #if self.training or not self.use_dynamic_ntk:
884
+ if not self.use_dynamic_ntk:
885
+ ntk_alpha_list = [1.0]
886
+ elif kv_seq_len != hidden_states.size()[1]:
887
+ ntk_alpha_list = self.rotary_emb._ntk_alpha_cached_list
888
+ else:
889
+ ntk_alpha_list = []
890
+ if attention_mask is not None and kv_seq_len > self.seq_length:
891
+ true_seq_lens = attention_mask.squeeze(1).squeeze(1).eq(0).sum(dim=-1, dtype=torch.int32)
892
+ for i in range(hidden_states.size()[0]):
893
+ #给batch中的每个样本计算ntk_alpha,计算方法是 true_seq_len / self.seq_length。qwen-7b中,self.seq_length=8192,qwen-14b中,self.seq_length=2048
894
+ true_seq_len = true_seq_lens[i].item()
895
+ ntk_alpha = self.get_ntk_alpha(true_seq_len)
896
+ ntk_alpha_list.append(ntk_alpha)
897
+ else:
898
+ ntk_alpha = self.get_ntk_alpha(kv_seq_len)
899
+ ntk_alpha_list.append(ntk_alpha)
900
+ self.rotary_emb._ntk_alpha_cached_list = ntk_alpha_list
901
+
902
+ rotary_pos_emb_list = []
903
+ for ntk_alpha in ntk_alpha_list:
904
+ rotary_pos_emb = self.rotary_emb(kv_seq_len, ntk_alpha=ntk_alpha) #训练时,ntk_alpha=1.0,rotary_emb根据kv_seq_len生成
905
+ rotary_pos_emb_list.append(rotary_pos_emb)
906
+
907
+ hidden_states = self.drop(hidden_states)
908
+ output_shape = input_shape + (hidden_states.size(-1),)
909
+
910
+ if self.gradient_checkpointing and self.training:
911
+ if use_cache:
912
+ logger.warning_once(
913
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
914
+ )
915
+ use_cache = False
916
+
917
+ presents = () if use_cache else None
918
+ all_self_attentions = () if output_attentions else None
919
+ all_hidden_states = () if output_hidden_states else None
920
+ for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
921
+
922
+ if output_hidden_states:
923
+ all_hidden_states = all_hidden_states + (hidden_states,)
924
+
925
+ if self.gradient_checkpointing and self.training:
926
+
927
+ def create_custom_forward(module):
928
+ def custom_forward(*inputs):
929
+ # None for past_key_value
930
+ return module(*inputs, use_cache, output_attentions)
931
+
932
+ return custom_forward
933
+
934
+ outputs = torch.utils.checkpoint.checkpoint(
935
+ create_custom_forward(block),
936
+ hidden_states,
937
+ rotary_pos_emb_list,
938
+ self.registered_causal_mask,
939
+ None,
940
+ attention_mask,
941
+ head_mask[i],
942
+ encoder_hidden_states,
943
+ encoder_attention_mask,
944
+ )
945
+ else:
946
+ outputs = block(
947
+ hidden_states,
948
+ layer_past=layer_past,
949
+ rotary_pos_emb_list=rotary_pos_emb_list,
950
+ registered_causal_mask=self.registered_causal_mask,
951
+ attention_mask=attention_mask,
952
+ head_mask=head_mask[i],
953
+ encoder_hidden_states=encoder_hidden_states,
954
+ encoder_attention_mask=encoder_attention_mask,
955
+ use_cache=use_cache,
956
+ output_attentions=output_attentions,
957
+ )
958
+
959
+ hidden_states = outputs[0]
960
+ if use_cache is True:
961
+ presents = presents + (outputs[1],)
962
+
963
+ if output_attentions:
964
+ all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
965
+
966
+ hidden_states = self.ln_f(hidden_states)
967
+ hidden_states = hidden_states.view(output_shape)
968
+ # Add last hidden state
969
+ if output_hidden_states:
970
+ all_hidden_states = all_hidden_states + (hidden_states,)
971
+
972
+ if not return_dict:
973
+ return tuple(
974
+ v for v in [hidden_states, presents, all_hidden_states] if v is not None
975
+ )
976
+
977
+ return BaseModelOutputWithPast(
978
+ last_hidden_state=hidden_states,
979
+ past_key_values=presents,
980
+ hidden_states=all_hidden_states,
981
+ attentions=all_self_attentions,
982
+ )
983
+
984
+
985
+ class QWenLMHeadModel(QWenPreTrainedModel):
986
+ _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.rotary_emb\.inv_freq"]
987
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.masked_bias"]
988
+
989
+ def __init__(self, config):
990
+ super().__init__(config)
991
+ assert (
992
+ config.bf16 + config.fp16 + config.fp32 <= 1
993
+ ), "Only one of \"bf16\", \"fp16\", \"fp32\" can be true"
994
+ logger.warn(
995
+ "Warning: please make sure that you are using the latest codes and checkpoints, "
996
+ "especially if you used Qwen-7B before 09.25.2023."
997
+ "请使用最新模型和代码,尤其如果你在9月25日前已经开始使用Qwen-7B,千万注意不要使用错误代码和模型。"
998
+ )
999
+
1000
+ autoset_precision = config.bf16 + config.fp16 + config.fp32 == 0
1001
+
1002
+ if autoset_precision:
1003
+ if SUPPORT_BF16:
1004
+ logger.warn(
1005
+ "The model is automatically converting to bf16 for faster inference. "
1006
+ "If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"."
1007
+ )
1008
+ config.bf16 = True
1009
+ elif SUPPORT_FP16:
1010
+ logger.warn(
1011
+ "The model is automatically converting to fp16 for faster inference. "
1012
+ "If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"."
1013
+ )
1014
+ config.fp16 = True
1015
+ else:
1016
+ config.fp32 = True
1017
+
1018
+ if config.bf16 and SUPPORT_CUDA and not SUPPORT_BF16:
1019
+ logger.warn("Your device does NOT seem to support bf16, you can switch to fp16 or fp32 by by passing fp16/fp32=True in \"AutoModelForCausalLM.from_pretrained\".")
1020
+ if config.fp16 and SUPPORT_CUDA and not SUPPORT_FP16:
1021
+ logger.warn("Your device does NOT support faster inference with fp16, please switch to fp32 which is likely to be faster")
1022
+ if config.fp32:
1023
+ if SUPPORT_BF16:
1024
+ logger.warn("Your device support faster inference by passing bf16=True in \"AutoModelForCausalLM.from_pretrained\".")
1025
+ elif SUPPORT_FP16:
1026
+ logger.warn("Your device support faster inference by passing fp16=True in \"AutoModelForCausalLM.from_pretrained\".")
1027
+
1028
+ if config.use_flash_attn == "auto":
1029
+ if config.bf16 or config.fp16:
1030
+ logger.warn("Try importing flash-attention for faster inference...")
1031
+ config.use_flash_attn = True
1032
+ else:
1033
+ config.use_flash_attn = False
1034
+ if config.use_flash_attn and config.fp32:
1035
+ logger.warn("Flash attention will be disabled because it does NOT support fp32.")
1036
+
1037
+ if config.use_flash_attn:
1038
+ _import_flash_attn()
1039
+
1040
+
1041
+ if hasattr(config, 'use_cache_quantization') and config.use_cache_quantization:
1042
+ config.use_flash_attn = False
1043
+ if hasattr(config, 'use_cache_kernel') and config.use_cache_kernel:
1044
+ try:
1045
+ from kernels.cpp_kernels import cache_autogptq_cuda_256
1046
+ except ImportError:
1047
+ cache_autogptq_cuda_256 = None
1048
+
1049
+ self.transformer = QWenModel(config)
1050
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
1051
+
1052
+ if config.bf16:
1053
+ self.transformer.bfloat16()
1054
+ self.lm_head.bfloat16()
1055
+ if config.fp16:
1056
+ self.transformer.half()
1057
+ self.lm_head.half()
1058
+ self.post_init()
1059
+
1060
+
1061
+ def get_output_embeddings(self):
1062
+ return self.lm_head
1063
+
1064
+ def set_output_embeddings(self, new_embeddings):
1065
+ self.lm_head = new_embeddings
1066
+
1067
+ def prepare_inputs_for_generation(
1068
+ self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
1069
+ ):
1070
+ token_type_ids = kwargs.get("token_type_ids", None)
1071
+ if past_key_values:
1072
+ input_ids = input_ids[:, -1].unsqueeze(-1)
1073
+ if token_type_ids is not None:
1074
+ token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
1075
+
1076
+ attention_mask = kwargs.get("attention_mask", None)
1077
+ position_ids = kwargs.get("position_ids", None)
1078
+
1079
+ if attention_mask is not None and position_ids is None:
1080
+ position_ids = attention_mask.long().cumsum(-1) - 1
1081
+ position_ids.masked_fill_(attention_mask == 0, 1)
1082
+ if past_key_values:
1083
+ position_ids = position_ids[:, -1].unsqueeze(-1)
1084
+ else:
1085
+ position_ids = None
1086
+
1087
+ if inputs_embeds is not None and past_key_values is None:
1088
+ model_inputs = {"inputs_embeds": inputs_embeds}
1089
+ else:
1090
+ model_inputs = {"input_ids": input_ids}
1091
+
1092
+ model_inputs.update(
1093
+ {
1094
+ "past_key_values": past_key_values,
1095
+ "use_cache": kwargs.get("use_cache"),
1096
+ "position_ids": position_ids,
1097
+ "attention_mask": attention_mask,
1098
+ "token_type_ids": token_type_ids,
1099
+ }
1100
+ )
1101
+ return model_inputs
1102
+
1103
+ def forward(
1104
+ self,
1105
+ input_ids: Optional[torch.LongTensor] = None,
1106
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1107
+ attention_mask: Optional[torch.FloatTensor] = None,
1108
+ token_type_ids: Optional[torch.LongTensor] = None,
1109
+ position_ids: Optional[torch.LongTensor] = None,
1110
+ head_mask: Optional[torch.FloatTensor] = None,
1111
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1112
+ encoder_hidden_states: Optional[torch.Tensor] = None,
1113
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
1114
+ labels: Optional[torch.LongTensor] = None,
1115
+ use_cache: Optional[bool] = None,
1116
+ output_attentions: Optional[bool] = None,
1117
+ output_hidden_states: Optional[bool] = None,
1118
+ return_dict: Optional[bool] = None,
1119
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
1120
+
1121
+ return_dict = (
1122
+ return_dict if return_dict is not None else self.config.use_return_dict
1123
+ )
1124
+
1125
+ transformer_outputs = self.transformer(
1126
+ input_ids,
1127
+ past_key_values=past_key_values,
1128
+ attention_mask=attention_mask,
1129
+ token_type_ids=token_type_ids,
1130
+ position_ids=position_ids,
1131
+ head_mask=head_mask,
1132
+ inputs_embeds=inputs_embeds,
1133
+ encoder_hidden_states=encoder_hidden_states,
1134
+ encoder_attention_mask=encoder_attention_mask,
1135
+ use_cache=use_cache,
1136
+ output_attentions=output_attentions,
1137
+ output_hidden_states=output_hidden_states,
1138
+ return_dict=return_dict,
1139
+ )
1140
+ hidden_states = transformer_outputs[0]
1141
+
1142
+ lm_logits = self.lm_head(hidden_states)
1143
+
1144
+ loss = None
1145
+ if labels is not None:
1146
+ labels = labels.to(lm_logits.device)
1147
+ shift_logits = lm_logits[..., :-1, :].contiguous()
1148
+ shift_labels = labels[..., 1:].contiguous()
1149
+ loss_fct = CrossEntropyLoss()
1150
+ loss = loss_fct(
1151
+ shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
1152
+ )
1153
+
1154
+
1155
+ if not return_dict:
1156
+ output = (lm_logits,) + transformer_outputs[1:]
1157
+ return ((loss,) + output) if loss is not None else output
1158
+
1159
+ if self.training:
1160
+ lm_logits=None
1161
+ return CausalLMOutputWithPast(
1162
+ loss=loss,
1163
+ logits=lm_logits,
1164
+ past_key_values=transformer_outputs.past_key_values,
1165
+ hidden_states=transformer_outputs.hidden_states,
1166
+ attentions=transformer_outputs.attentions,
1167
+ )
1168
+
1169
+ @staticmethod
1170
+ def _reorder_cache(
1171
+ past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
1172
+ ) -> Tuple[Tuple[torch.Tensor]]:
1173
+
1174
+ return tuple(
1175
+ tuple(
1176
+ past_state.index_select(0, beam_idx.to(past_state.device))
1177
+ for past_state in layer_past
1178
+ )
1179
+ for layer_past in past_key_values
1180
+ )
1181
+
1182
+ def chat(
1183
+ self,
1184
+ tokenizer: PreTrainedTokenizer,
1185
+ query: str,
1186
+ history: Optional[HistoryType],
1187
+ system: str = "You are a helpful assistant.",
1188
+ append_history: bool = True,
1189
+ stream: Optional[bool] = _SENTINEL,
1190
+ stop_words_ids: Optional[List[List[int]]] = None,
1191
+ generation_config: Optional[GenerationConfig] = None,
1192
+ **kwargs,
1193
+ ) -> Tuple[str, HistoryType]:
1194
+ generation_config = generation_config if generation_config is not None else self.generation_config
1195
+
1196
+ assert stream is _SENTINEL, _ERROR_STREAM_IN_CHAT
1197
+ assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
1198
+ if history is None:
1199
+ history = []
1200
+ if stop_words_ids is None:
1201
+ stop_words_ids = []
1202
+
1203
+ max_window_size = kwargs.get('max_window_size', None)
1204
+ if max_window_size is None:
1205
+ max_window_size = generation_config.max_window_size
1206
+ raw_text, context_tokens = make_context(
1207
+ tokenizer,
1208
+ query,
1209
+ history=history,
1210
+ system=system,
1211
+ max_window_size=max_window_size,
1212
+ chat_format=generation_config.chat_format,
1213
+ )
1214
+
1215
+ stop_words_ids.extend(get_stop_words_ids(
1216
+ generation_config.chat_format, tokenizer
1217
+ ))
1218
+ input_ids = torch.tensor([context_tokens]).to(self.device)
1219
+ outputs = self.generate(
1220
+ input_ids,
1221
+ stop_words_ids=stop_words_ids,
1222
+ return_dict_in_generate=False,
1223
+ generation_config=generation_config,
1224
+ **kwargs,
1225
+ )
1226
+
1227
+ response = decode_tokens(
1228
+ outputs[0],
1229
+ tokenizer,
1230
+ raw_text_len=len(raw_text),
1231
+ context_length=len(context_tokens),
1232
+ chat_format=generation_config.chat_format,
1233
+ verbose=False,
1234
+ errors='replace'
1235
+ )
1236
+
1237
+ if append_history:
1238
+ history.append((query, response))
1239
+
1240
+ return response, history
1241
+
1242
+ def chat_stream(
1243
+ self,
1244
+ tokenizer: PreTrainedTokenizer,
1245
+ query: str,
1246
+ history: Optional[HistoryType],
1247
+ system: str = "You are a helpful assistant.",
1248
+ stop_words_ids: Optional[List[List[int]]] = None,
1249
+ logits_processor: Optional[LogitsProcessorList] = None,
1250
+ generation_config: Optional[GenerationConfig] = None,
1251
+ **kwargs,
1252
+ ) -> Generator[str, Any, None]:
1253
+ generation_config = generation_config if generation_config is not None else self.generation_config
1254
+ assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
1255
+ if history is None:
1256
+ history = []
1257
+ if stop_words_ids is None:
1258
+ stop_words_ids = []
1259
+
1260
+ max_window_size = kwargs.get('max_window_size', None)
1261
+ if max_window_size is None:
1262
+ max_window_size = generation_config.max_window_size
1263
+ raw_text, context_tokens = make_context(
1264
+ tokenizer,
1265
+ query,
1266
+ history=history,
1267
+ system=system,
1268
+ max_window_size=max_window_size,
1269
+ chat_format=generation_config.chat_format,
1270
+ )
1271
+
1272
+ stop_words_ids.extend(get_stop_words_ids(
1273
+ generation_config.chat_format, tokenizer
1274
+ ))
1275
+ if stop_words_ids is not None:
1276
+ stop_words_logits_processor = StopWordsLogitsProcessor(
1277
+ stop_words_ids=stop_words_ids,
1278
+ eos_token_id=generation_config.eos_token_id,
1279
+ )
1280
+ if logits_processor is None:
1281
+ logits_processor = LogitsProcessorList([stop_words_logits_processor])
1282
+ else:
1283
+ logits_processor.append(stop_words_logits_processor)
1284
+ input_ids = torch.tensor([context_tokens]).to(self.device)
1285
+
1286
+ from transformers_stream_generator.main import NewGenerationMixin, StreamGenerationConfig
1287
+ self.__class__.generate_stream = NewGenerationMixin.generate
1288
+ self.__class__.sample_stream = NewGenerationMixin.sample_stream
1289
+ stream_config = StreamGenerationConfig(**generation_config.to_dict(), do_stream=True)
1290
+
1291
+ def stream_generator():
1292
+ outputs = []
1293
+ for token in self.generate_stream(
1294
+ input_ids,
1295
+ return_dict_in_generate=False,
1296
+ generation_config=stream_config,
1297
+ logits_processor=logits_processor,
1298
+ seed=-1,
1299
+ **kwargs):
1300
+ outputs.append(token.item())
1301
+ yield tokenizer.decode(outputs, skip_special_tokens=True, errors='ignore')
1302
+
1303
+ return stream_generator()
1304
+
1305
+ def generate(
1306
+ self,
1307
+ inputs: Optional[torch.Tensor] = None,
1308
+ generation_config: Optional[GenerationConfig] = None,
1309
+ logits_processor: Optional[LogitsProcessorList] = None,
1310
+ stopping_criteria: Optional[StoppingCriteriaList] = None,
1311
+ prefix_allowed_tokens_fn: Optional[
1312
+ Callable[[int, torch.Tensor], List[int]]
1313
+ ] = None,
1314
+ synced_gpus: Optional[bool] = None,
1315
+ assistant_model: Optional["PreTrainedModel"] = None,
1316
+ streamer: Optional["BaseStreamer"] = None,
1317
+ **kwargs,
1318
+ ) -> Union[GenerateOutput, torch.LongTensor]:
1319
+ generation_config = generation_config if generation_config is not None else self.generation_config
1320
+
1321
+ # Process stop_words_ids.
1322
+ stop_words_ids = kwargs.pop("stop_words_ids", None)
1323
+ if stop_words_ids is None and generation_config is not None:
1324
+ stop_words_ids = getattr(generation_config, "stop_words_ids", None)
1325
+ if stop_words_ids is None:
1326
+ stop_words_ids = getattr(generation_config, "stop_words_ids", None)
1327
+
1328
+ if stop_words_ids is not None:
1329
+ stop_words_logits_processor = StopWordsLogitsProcessor(
1330
+ stop_words_ids=stop_words_ids,
1331
+ eos_token_id=generation_config.eos_token_id,
1332
+ )
1333
+ if logits_processor is None:
1334
+ logits_processor = LogitsProcessorList([stop_words_logits_processor])
1335
+ else:
1336
+ logits_processor.append(stop_words_logits_processor)
1337
+
1338
+ return super().generate(
1339
+ inputs,
1340
+ generation_config=generation_config,
1341
+ logits_processor=logits_processor,
1342
+ stopping_criteria=stopping_criteria,
1343
+ prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
1344
+ synced_gpus=synced_gpus,
1345
+ assistant_model=assistant_model,
1346
+ streamer=streamer,
1347
+ **kwargs,
1348
+ )
1349
+
1350
+
1351
+ class RotaryEmbedding(torch.nn.Module):
1352
+ def __init__(self, dim, base=10000):
1353
+ super().__init__()
1354
+ self.dim = dim
1355
+ self.base = base
1356
+ inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
1357
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
1358
+ if importlib.util.find_spec("einops") is None:
1359
+ raise RuntimeError("einops is required for Rotary Embedding")
1360
+
1361
+ self._rotary_pos_emb_cache = None
1362
+ self._seq_len_cached = 0
1363
+ self._ntk_alpha_cached = 1.0
1364
+ self._ntk_alpha_cached_list = [1.0]
1365
+
1366
+ def update_rotary_pos_emb_cache(self, max_seq_len, offset=0, ntk_alpha=1.0):
1367
+ seqlen = max_seq_len + offset
1368
+ if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
1369
+ #计算新的base。ntk_alpha=1时,base不变;ntk_alpha>1时,base变大
1370
+ base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
1371
+ #波长=2*pi*base^(2d/D),即频率的倒数。此处算出了每个维度对应的频率。由于每两个维度对应一个频率,所以频率的个数为dim/2
1372
+ self.inv_freq = 1.0 / (
1373
+ base
1374
+ ** (
1375
+ torch.arange(0, self.dim, 2, device=self.inv_freq.device).float()
1376
+ / self.dim
1377
+ )
1378
+ )
1379
+ self._seq_len_cached = max(seqlen+512, 16)
1380
+ self._ntk_alpha_cached = ntk_alpha
1381
+ seq = torch.arange(self._seq_len_cached, device=self.inv_freq.device)
1382
+ freqs = torch.outer(seq.type_as(self.inv_freq), self.inv_freq) #做外积,获得seq_len*(dim/2)的矩阵
1383
+
1384
+ emb = torch.cat((freqs, freqs), dim=-1) #获得(seq_len*2)*dim的矩阵
1385
+ from einops import rearrange
1386
+
1387
+ emb = rearrange(emb, "n d -> 1 n 1 d") #获得1*(seq_len*2)*1*dim的矩阵
1388
+
1389
+ cos, sin = emb.cos(), emb.sin()
1390
+ self._rotary_pos_emb_cache = [cos, sin]
1391
+
1392
+ def forward(self, max_seq_len, offset=0, ntk_alpha=1.0):
1393
+ self.update_rotary_pos_emb_cache(max_seq_len, offset, ntk_alpha)
1394
+ cos, sin = self._rotary_pos_emb_cache
1395
+ return [cos[:, offset : offset + max_seq_len], sin[:, offset : offset + max_seq_len]]
1396
+
1397
+
1398
+ class YaRNRotaryEmbedding(RotaryEmbedding):
1399
+ def __init__(self, dim, base=10000,interleaved=False,extrapolation_factor=1,
1400
+ attn_factor=1, beta_fast=32, beta_slow=1,original_max_position_embeddings=2048):
1401
+ """
1402
+ interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
1403
+ of 1st half and 2nd half (GPT-NeoX style).
1404
+ pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
1405
+ otherwise they might be in lower precision.
1406
+ This option was added because previously (before 2023-07-02), when we construct
1407
+ the position indices, we use the dtype of self.inv_freq. In most cases this would
1408
+ be fp32, but if the model is trained in pure bf16 (not mixed precision), then
1409
+ self.inv_freq would be bf16, and the position indices are also in bf16.
1410
+ Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
1411
+ embeddings for some positions will coincide.
1412
+ To maintain compatibility with models previously trained in pure bf16,
1413
+ we add this option.
1414
+ scaling_factor: RotaryEmbedding extended with YaRN scaling.
1415
+ """
1416
+ super().__init__(dim, base)
1417
+ self.interleaved = interleaved
1418
+ self.extrapolation_factor = extrapolation_factor
1419
+ self.attn_factor = attn_factor
1420
+ self.beta_fast = beta_fast
1421
+ self.beta_slow = beta_slow
1422
+ self.original_max_position_embeddings = original_max_position_embeddings
1423
+
1424
+ def update_rotary_pos_emb_cache(self, max_seq_len, offset=0, ntk_alpha=1.0):
1425
+ seqlen = max_seq_len + offset
1426
+ if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
1427
+ #波长=2*pi*base^(2d/D),即频率的倒数。此处算出了每个维度对应的频率。由于每两个维度对应一个频率,所以频率的个数为dim/2
1428
+ self._compute_inv_freq(ntk_alpha, device=self.inv_freq.device)
1429
+ self.mscale = float(_yarn_get_mscale(ntk_alpha) * self.attn_factor)
1430
+
1431
+ self._seq_len_cached = max(seqlen+512, 16)
1432
+ self._ntk_alpha_cached = ntk_alpha
1433
+ seq = torch.arange(self._seq_len_cached, device=self.inv_freq.device)
1434
+ freqs = torch.outer(seq.type_as(self.inv_freq), self.inv_freq) #seq和逆频率做外积,获得seq_len*(dim/2)的矩阵
1435
+
1436
+ emb = torch.cat((freqs, freqs), dim=-1) #获得(seq_len*2)*dim的矩阵
1437
+ from einops import rearrange
1438
+
1439
+ emb = rearrange(emb, "n d -> 1 n 1 d") #获得1*(seq_len*2)*1*dim的矩阵
1440
+
1441
+ cos, sin = emb.cos() * self.mscale, emb.sin() * self.mscale
1442
+ self._rotary_pos_emb_cache = [cos, sin]
1443
+
1444
+ def _compute_inv_freq(self, scaling_factor, device=None):
1445
+ pos_freqs = self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim)
1446
+ inv_freq_extrapolation = 1.0 / pos_freqs
1447
+ inv_freq_interpolation = 1.0 / (scaling_factor * pos_freqs)
1448
+
1449
+ low, high = _yarn_find_correction_range(self.beta_fast, self.beta_slow, self.dim, self.base, self.original_max_position_embeddings)
1450
+ inv_freq_mask = (1 - _yarn_linear_ramp_mask(low, high, self.dim // 2).float().to(device)) * self.extrapolation_factor # Get n-d rotational scaling corrected for extrapolation
1451
+ inv_freq = inv_freq_interpolation * (1 - inv_freq_mask) + inv_freq_extrapolation * inv_freq_mask
1452
+ inv_freq=inv_freq.float()
1453
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
1454
+
1455
+ # Inverse dim formula to find dim based on number of rotations
1456
+ def _yarn_find_correction_dim(num_rotations, dim, base=10000, max_position_embeddings=2048):
1457
+ return (dim * math.log(max_position_embeddings/(num_rotations * 2 * math.pi)))/(2 * math.log(base))
1458
+
1459
+ # Find dim range bounds based on rotations
1460
+ def _yarn_find_correction_range(low_rot, high_rot, dim, base=10000, max_position_embeddings=2048):
1461
+ low = math.floor(_yarn_find_correction_dim(
1462
+ low_rot, dim, base, max_position_embeddings))
1463
+ high = math.ceil(_yarn_find_correction_dim(
1464
+ high_rot, dim, base, max_position_embeddings))
1465
+ return max(low, 0), min(high, dim-1) # Clamp values just in case
1466
+
1467
+ def _yarn_linear_ramp_mask(min, max, dim):
1468
+ if min == max:
1469
+ max += 0.001 # Prevent singularity
1470
+
1471
+ linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
1472
+ ramp_func = torch.clamp(linear_func, 0, 1)
1473
+ return ramp_func
1474
+
1475
+ def _yarn_get_mscale(scale=1):
1476
+ if scale <= 1:
1477
+ return 1.0
1478
+ return 0.1 * math.log(scale) + 1.0
1479
+
1480
+
1481
+
1482
+ def _rotate_half(x):
1483
+ from einops import rearrange
1484
+
1485
+ x = rearrange(x, "... (j d) -> ... j d", j=2)
1486
+ x1, x2 = x.unbind(dim=-2)
1487
+ return torch.cat((-x2, x1), dim=-1)
1488
+
1489
+
1490
+ def apply_rotary_pos_emb(t, freqs):
1491
+ cos, sin = freqs
1492
+ if apply_rotary_emb_func is not None and t.is_cuda:
1493
+ t_ = t.float()
1494
+ cos = cos.squeeze(0).squeeze(1)[:, : cos.shape[-1] // 2]
1495
+ sin = sin.squeeze(0).squeeze(1)[:, : sin.shape[-1] // 2]
1496
+ output = apply_rotary_emb_func(t_, cos, sin).type_as(t)
1497
+ return output
1498
+ else:
1499
+ rot_dim = freqs[0].shape[-1]
1500
+ cos, sin = freqs
1501
+ t_, t_pass_ = t[..., :rot_dim], t[..., rot_dim:]
1502
+ t_ = t_.float()
1503
+ t_pass_ = t_pass_.float()
1504
+ t_ = (t_ * cos) + (_rotate_half(t_) * sin)
1505
+ return torch.cat((t_, t_pass_), dim=-1).type_as(t)
1506
+
1507
+
1508
+ class RMSNorm(torch.nn.Module):
1509
+ def __init__(self, dim: int, eps: float = 1e-6):
1510
+ super().__init__()
1511
+ self.eps = eps
1512
+ self.weight = nn.Parameter(torch.ones(dim))
1513
+
1514
+ def _norm(self, x):
1515
+ return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
1516
+
1517
+ def forward(self, x):
1518
+ if rms_norm is not None and x.is_cuda:
1519
+ return rms_norm(x, self.weight, self.eps)
1520
+ else:
1521
+ output = self._norm(x.float()).type_as(x)
1522
+ return output * self.weight
qwen.tiktoken ADDED
The diff for this file is too large to render. See raw diff
 
qwen_generation_utils.py ADDED
@@ -0,0 +1,416 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ """Generation support."""
7
+
8
+ from typing import Tuple, List, Union, Iterable
9
+
10
+ import numpy as np
11
+ import torch
12
+ import torch.nn.functional as F
13
+ from transformers import PreTrainedTokenizer
14
+ from transformers import logging
15
+ from transformers.generation import LogitsProcessor
16
+
17
+ logger = logging.get_logger(__name__)
18
+
19
+ # Types.
20
+ HistoryType = List[Tuple[str, str]]
21
+ TokensType = List[int]
22
+ BatchTokensType = List[List[int]]
23
+
24
+
25
+ def pad_batch(batch: BatchTokensType, pad_id: int, seq_length: int) -> BatchTokensType:
26
+ for tokens in batch:
27
+ context_length = len(tokens)
28
+ if context_length < seq_length:
29
+ tokens.extend([pad_id] * (seq_length - context_length))
30
+ return batch
31
+
32
+
33
+ def get_ltor_masks_and_position_ids(
34
+ data,
35
+ eod_token,
36
+ reset_position_ids,
37
+ reset_attention_mask,
38
+ eod_mask_loss,
39
+ ):
40
+ """Build masks and position id for left to right model."""
41
+
42
+ # Extract batch size and sequence length.
43
+ micro_batch_size, seq_length = data.size()
44
+
45
+ # Attention mask (lower triangular).
46
+ if reset_attention_mask:
47
+ att_mask_batch = micro_batch_size
48
+ else:
49
+ att_mask_batch = 1
50
+ attention_mask = torch.tril(
51
+ torch.ones((att_mask_batch, seq_length, seq_length), device=data.device)
52
+ ).view(att_mask_batch, 1, seq_length, seq_length)
53
+
54
+ # Loss mask.
55
+ loss_mask = torch.ones(data.size(), dtype=torch.float, device=data.device)
56
+ if eod_mask_loss:
57
+ loss_mask[data == eod_token] = 0.0
58
+
59
+ # Position ids.
60
+ position_ids = torch.arange(seq_length, dtype=torch.long, device=data.device)
61
+ position_ids = position_ids.unsqueeze(0).expand_as(data)
62
+ # We need to clone as the ids will be modifed based on batch index.
63
+ if reset_position_ids:
64
+ position_ids = position_ids.clone()
65
+
66
+ if reset_position_ids or reset_attention_mask:
67
+ # Loop through the batches:
68
+ for b in range(micro_batch_size):
69
+
70
+ # Find indecies where EOD token is.
71
+ eod_index = position_ids[b, data[b] == eod_token]
72
+ # Detach indecies from positions if going to modify positions.
73
+ if reset_position_ids:
74
+ eod_index = eod_index.clone()
75
+
76
+ # Loop through EOD indecies:
77
+ prev_index = 0
78
+ for j in range(eod_index.size()[0]):
79
+ i = eod_index[j]
80
+ # Mask attention loss.
81
+ if reset_attention_mask:
82
+ attention_mask[b, 0, (i + 1) :, : (i + 1)] = 0
83
+ # Reset positions.
84
+ if reset_position_ids:
85
+ position_ids[b, (i + 1) :] -= i + 1 - prev_index
86
+ prev_index = i + 1
87
+
88
+ # Convert attention mask to binary:
89
+ attention_mask = attention_mask < 0.5
90
+
91
+ return attention_mask, loss_mask, position_ids
92
+
93
+
94
+ def get_batch(context_tokens: torch.LongTensor, eod_id: int):
95
+ """Generate batch from context tokens."""
96
+ # Move to GPU.
97
+ tokens = context_tokens.contiguous().to(context_tokens.device)
98
+ # Get the attention mask and postition ids.
99
+ attention_mask, _, position_ids = get_ltor_masks_and_position_ids(
100
+ tokens,
101
+ eod_id,
102
+ reset_position_ids=False,
103
+ reset_attention_mask=False,
104
+ eod_mask_loss=False,
105
+ )
106
+ return tokens, attention_mask, position_ids
107
+
108
+
109
+ def get_stop_words_ids(chat_format, tokenizer):
110
+ if chat_format == "raw":
111
+ stop_words_ids = [tokenizer.encode("Human:"), [tokenizer.eod_id]]
112
+ elif chat_format == "chatml":
113
+ stop_words_ids = [[tokenizer.im_end_id], [tokenizer.im_start_id]]
114
+ else:
115
+ raise NotImplementedError(f"Unknown chat format {chat_format!r}")
116
+ return stop_words_ids
117
+
118
+
119
+ def make_context(
120
+ tokenizer: PreTrainedTokenizer,
121
+ query: str,
122
+ history: List[Tuple[str, str]] = None,
123
+ system: str = "",
124
+ max_window_size: int = 6144,
125
+ chat_format: str = "chatml",
126
+ ):
127
+ if history is None:
128
+ history = []
129
+
130
+ if chat_format == "chatml":
131
+ im_start, im_end = "<|im_start|>", "<|im_end|>"
132
+ im_start_tokens = [tokenizer.im_start_id]
133
+ im_end_tokens = [tokenizer.im_end_id]
134
+ nl_tokens = tokenizer.encode("\n")
135
+
136
+ def _tokenize_str(role, content):
137
+ return f"{role}\n{content}", tokenizer.encode(
138
+ role, allowed_special=set()
139
+ ) + nl_tokens + tokenizer.encode(content, allowed_special=set())
140
+
141
+ system_text, system_tokens_part = _tokenize_str("system", system)
142
+ system_tokens = im_start_tokens + system_tokens_part + im_end_tokens
143
+
144
+ raw_text = ""
145
+ context_tokens = []
146
+
147
+ for turn_query, turn_response in reversed(history):
148
+ query_text, query_tokens_part = _tokenize_str("user", turn_query)
149
+ query_tokens = im_start_tokens + query_tokens_part + im_end_tokens
150
+ response_text, response_tokens_part = _tokenize_str(
151
+ "assistant", turn_response
152
+ )
153
+ response_tokens = im_start_tokens + response_tokens_part + im_end_tokens
154
+
155
+ next_context_tokens = nl_tokens + query_tokens + nl_tokens + response_tokens
156
+ prev_chat = (
157
+ f"\n{im_start}{query_text}{im_end}\n{im_start}{response_text}{im_end}"
158
+ )
159
+
160
+ current_context_size = (
161
+ len(system_tokens) + len(next_context_tokens) + len(context_tokens)
162
+ )
163
+ if current_context_size < max_window_size:
164
+ context_tokens = next_context_tokens + context_tokens
165
+ raw_text = prev_chat + raw_text
166
+ else:
167
+ break
168
+
169
+ context_tokens = system_tokens + context_tokens
170
+ raw_text = f"{im_start}{system_text}{im_end}" + raw_text
171
+ context_tokens += (
172
+ nl_tokens
173
+ + im_start_tokens
174
+ + _tokenize_str("user", query)[1]
175
+ + im_end_tokens
176
+ + nl_tokens
177
+ + im_start_tokens
178
+ + tokenizer.encode("assistant")
179
+ + nl_tokens
180
+ )
181
+ raw_text += f"\n{im_start}user\n{query}{im_end}\n{im_start}assistant\n"
182
+
183
+ elif chat_format == "raw":
184
+ raw_text = query
185
+ context_tokens = tokenizer.encode(raw_text)
186
+ else:
187
+ raise NotImplementedError(f"Unknown chat format {chat_format!r}")
188
+
189
+ return raw_text, context_tokens
190
+
191
+
192
+ def _decode_default(
193
+ tokens: List[int],
194
+ *,
195
+ stop_words: List[str],
196
+ eod_words: List[str],
197
+ tokenizer: PreTrainedTokenizer,
198
+ raw_text_len: int,
199
+ verbose: bool = False,
200
+ return_end_reason: bool = False,
201
+ errors: str='replace',
202
+ ):
203
+ trim_decode_tokens = tokenizer.decode(tokens, errors=errors)[raw_text_len:]
204
+ if verbose:
205
+ print("\nRaw Generate: ", trim_decode_tokens)
206
+
207
+ end_reason = f"Gen length {len(tokens)}"
208
+ for stop_word in stop_words:
209
+ trim_decode_tokens = trim_decode_tokens.replace(stop_word, "").strip()
210
+ for eod_word in eod_words:
211
+ if eod_word in trim_decode_tokens:
212
+ end_reason = f"Gen {eod_word!r}"
213
+ trim_decode_tokens = trim_decode_tokens.split(eod_word)[0]
214
+ trim_decode_tokens = trim_decode_tokens.strip()
215
+ if verbose:
216
+ print("\nEnd Reason:", end_reason)
217
+ print("\nGenerate: ", trim_decode_tokens)
218
+
219
+ if return_end_reason:
220
+ return trim_decode_tokens, end_reason
221
+ else:
222
+ return trim_decode_tokens
223
+
224
+
225
+ def _decode_chatml(
226
+ tokens: List[int],
227
+ *,
228
+ stop_words: List[str],
229
+ eod_token_ids: List[int],
230
+ tokenizer: PreTrainedTokenizer,
231
+ raw_text_len: int,
232
+ context_length: int,
233
+ verbose: bool = False,
234
+ return_end_reason: bool = False,
235
+ errors: str='replace'
236
+ ):
237
+ end_reason = f"Gen length {len(tokens)}"
238
+ eod_token_idx = context_length
239
+ for eod_token_idx in range(context_length, len(tokens)):
240
+ if tokens[eod_token_idx] in eod_token_ids:
241
+ end_reason = f"Gen {tokenizer.decode([tokens[eod_token_idx]])!r}"
242
+ break
243
+
244
+ trim_decode_tokens = tokenizer.decode(tokens[:eod_token_idx], errors=errors)[raw_text_len:]
245
+ if verbose:
246
+ print("\nRaw Generate w/o EOD:", tokenizer.decode(tokens, errors=errors)[raw_text_len:])
247
+ print("\nRaw Generate:", trim_decode_tokens)
248
+ print("\nEnd Reason:", end_reason)
249
+ for stop_word in stop_words:
250
+ trim_decode_tokens = trim_decode_tokens.replace(stop_word, "").strip()
251
+ trim_decode_tokens = trim_decode_tokens.strip()
252
+ if verbose:
253
+ print("\nGenerate:", trim_decode_tokens)
254
+
255
+ if return_end_reason:
256
+ return trim_decode_tokens, end_reason
257
+ else:
258
+ return trim_decode_tokens
259
+
260
+
261
+ def decode_tokens(
262
+ tokens: Union[torch.LongTensor, TokensType],
263
+ tokenizer: PreTrainedTokenizer,
264
+ raw_text_len: int,
265
+ context_length: int,
266
+ chat_format: str,
267
+ verbose: bool = False,
268
+ return_end_reason: bool = False,
269
+ errors: str="replace",
270
+ ) -> str:
271
+ if torch.is_tensor(tokens):
272
+ tokens = tokens.cpu().numpy().tolist()
273
+
274
+ if chat_format == "chatml":
275
+ return _decode_chatml(
276
+ tokens,
277
+ stop_words=[],
278
+ eod_token_ids=[tokenizer.im_start_id, tokenizer.im_end_id],
279
+ tokenizer=tokenizer,
280
+ raw_text_len=raw_text_len,
281
+ context_length=context_length,
282
+ verbose=verbose,
283
+ return_end_reason=return_end_reason,
284
+ errors=errors,
285
+ )
286
+ elif chat_format == "raw":
287
+ return _decode_default(
288
+ tokens,
289
+ stop_words=["<|endoftext|>"],
290
+ eod_words=["<|endoftext|>"],
291
+ tokenizer=tokenizer,
292
+ raw_text_len=raw_text_len,
293
+ verbose=verbose,
294
+ return_end_reason=return_end_reason,
295
+ errors=errors,
296
+ )
297
+ else:
298
+ raise NotImplementedError(f"Unknown chat format {chat_format!r}")
299
+
300
+
301
+ class StopWordsLogitsProcessor(LogitsProcessor):
302
+ """
303
+ :class:`transformers.LogitsProcessor` that enforces that when specified sequences appear, stop geration.
304
+
305
+ Args:
306
+ stop_words_ids (:obj:`List[List[int]]`):
307
+ List of list of token ids of stop ids. In order to get the tokens of the words
308
+ that should not appear in the generated text, use :obj:`tokenizer(bad_word,
309
+ add_prefix_space=True).input_ids`.
310
+ eos_token_id (:obj:`int`):
311
+ The id of the `end-of-sequence` token.
312
+ """
313
+
314
+ def __init__(self, stop_words_ids: Iterable[Iterable[int]], eos_token_id: int):
315
+
316
+ if not isinstance(stop_words_ids, List) or len(stop_words_ids) == 0:
317
+ raise ValueError(
318
+ f"`stop_words_ids` has to be a non-emtpy list, but is {stop_words_ids}."
319
+ )
320
+ if any(not isinstance(bad_word_ids, list) for bad_word_ids in stop_words_ids):
321
+ raise ValueError(
322
+ f"`stop_words_ids` has to be a list of lists, but is {stop_words_ids}."
323
+ )
324
+ if any(
325
+ any(
326
+ (not isinstance(token_id, (int, np.integer)) or token_id < 0)
327
+ for token_id in stop_word_ids
328
+ )
329
+ for stop_word_ids in stop_words_ids
330
+ ):
331
+ raise ValueError(
332
+ f"Each list in `stop_words_ids` has to be a list of positive integers, but is {stop_words_ids}."
333
+ )
334
+
335
+ self.stop_words_ids = list(
336
+ filter(
337
+ lambda bad_token_seq: bad_token_seq != [eos_token_id], stop_words_ids
338
+ )
339
+ )
340
+ self.eos_token_id = eos_token_id
341
+ for stop_token_seq in self.stop_words_ids:
342
+ assert (
343
+ len(stop_token_seq) > 0
344
+ ), "Stop words token sequences {} cannot have an empty list".format(
345
+ stop_words_ids
346
+ )
347
+
348
+ def __call__(
349
+ self, input_ids: torch.LongTensor, scores: torch.FloatTensor
350
+ ) -> torch.FloatTensor:
351
+ stopped_samples = self._calc_stopped_samples(input_ids)
352
+ for i, should_stop in enumerate(stopped_samples):
353
+ if should_stop:
354
+ scores[i, self.eos_token_id] = float(2**15)
355
+ return scores
356
+
357
+ def _tokens_match(self, prev_tokens: torch.LongTensor, tokens: List[int]) -> bool:
358
+ if len(tokens) == 0:
359
+ # if bad word tokens is just one token always ban it
360
+ return True
361
+ elif len(tokens) > len(prev_tokens):
362
+ # if bad word tokens are longer then prev input_ids they can't be equal
363
+ return False
364
+ elif prev_tokens[-len(tokens) :].tolist() == tokens:
365
+ # if tokens match
366
+ return True
367
+ else:
368
+ return False
369
+
370
+ def _calc_stopped_samples(self, prev_input_ids: Iterable[int]) -> Iterable[int]:
371
+ stopped_samples = []
372
+ for prev_input_ids_slice in prev_input_ids:
373
+ match = False
374
+ for stop_token_seq in self.stop_words_ids:
375
+ if self._tokens_match(prev_input_ids_slice, stop_token_seq):
376
+ # if tokens do not match continue
377
+ match = True
378
+ break
379
+ stopped_samples.append(match)
380
+
381
+ return stopped_samples
382
+
383
+
384
+ def top_k_logits(logits, top_k=0, top_p=0.0, filter_value=-float("Inf")):
385
+ """This function has been mostly taken from huggingface conversational
386
+ ai code at
387
+ https://medium.com/huggingface/how-to-build-a-state-of-the-art-
388
+ conversational-ai-with-transfer-learning-2d818ac26313"""
389
+
390
+ if top_k > 0:
391
+ # Remove all tokens with a probability less than the
392
+ # last token of the top-k
393
+ indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
394
+ logits[indices_to_remove] = filter_value
395
+
396
+ if top_p > 0.0:
397
+ # Cconvert to 1D
398
+ sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
399
+ cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
400
+
401
+ # Remove tokens with cumulative probability above the threshold
402
+ sorted_indices_to_remove = cumulative_probs > top_p
403
+ # Shift the indices to the right to keep also the first token
404
+ # above the threshold
405
+ sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
406
+ sorted_indices_to_remove[..., 0] = 0
407
+ for i in range(sorted_indices.size(0)):
408
+ indices_to_remove = sorted_indices[i][sorted_indices_to_remove[i]]
409
+ logits[i][indices_to_remove] = filter_value
410
+
411
+ return logits
412
+
413
+
414
+ def switch(val1, val2, boolean):
415
+ boolean = boolean.type_as(val1)
416
+ return (1 - boolean) * val1 + boolean * val2
tokenization_qwen.py ADDED
@@ -0,0 +1,276 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ """Tokenization classes for QWen."""
7
+
8
+ import base64
9
+ import logging
10
+ import os
11
+ import unicodedata
12
+ from typing import Collection, Dict, List, Set, Tuple, Union
13
+
14
+ import tiktoken
15
+ from transformers import PreTrainedTokenizer, AddedToken
16
+
17
+ logger = logging.getLogger(__name__)
18
+
19
+
20
+ VOCAB_FILES_NAMES = {"vocab_file": "qwen.tiktoken"}
21
+
22
+ PAT_STR = r"""(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"""
23
+ ENDOFTEXT = "<|endoftext|>"
24
+ IMSTART = "<|im_start|>"
25
+ IMEND = "<|im_end|>"
26
+ # as the default behavior is changed to allow special tokens in
27
+ # regular texts, the surface forms of special tokens need to be
28
+ # as different as possible to minimize the impact
29
+ EXTRAS = tuple((f"<|extra_{i}|>" for i in range(205)))
30
+ # changed to use actual index to avoid misconfiguration with vocabulary expansion
31
+ SPECIAL_START_ID = 151643
32
+ SPECIAL_TOKENS = tuple(
33
+ enumerate(
34
+ (
35
+ (
36
+ ENDOFTEXT,
37
+ IMSTART,
38
+ IMEND,
39
+ )
40
+ + EXTRAS
41
+ ),
42
+ start=SPECIAL_START_ID,
43
+ )
44
+ )
45
+ SPECIAL_TOKENS_SET = set(t for i, t in SPECIAL_TOKENS)
46
+
47
+
48
+ def _load_tiktoken_bpe(tiktoken_bpe_file: str) -> Dict[bytes, int]:
49
+ with open(tiktoken_bpe_file, "rb") as f:
50
+ contents = f.read()
51
+ return {
52
+ base64.b64decode(token): int(rank)
53
+ for token, rank in (line.split() for line in contents.splitlines() if line)
54
+ }
55
+
56
+
57
+ class QWenTokenizer(PreTrainedTokenizer):
58
+ """QWen tokenizer."""
59
+
60
+ vocab_files_names = VOCAB_FILES_NAMES
61
+
62
+ def __init__(
63
+ self,
64
+ vocab_file,
65
+ errors="replace",
66
+ extra_vocab_file=None,
67
+ **kwargs,
68
+ ):
69
+ super().__init__(**kwargs)
70
+
71
+ # how to handle errors in decoding UTF-8 byte sequences
72
+ # use ignore if you are in streaming inference
73
+ self.errors = errors
74
+
75
+ self.mergeable_ranks = _load_tiktoken_bpe(vocab_file) # type: Dict[bytes, int]
76
+ self.special_tokens = {
77
+ token: index
78
+ for index, token in SPECIAL_TOKENS
79
+ }
80
+
81
+ # try load extra vocab from file
82
+ if extra_vocab_file is not None:
83
+ used_ids = set(self.mergeable_ranks.values()) | set(self.special_tokens.values())
84
+ extra_mergeable_ranks = _load_tiktoken_bpe(extra_vocab_file)
85
+ for token, index in extra_mergeable_ranks.items():
86
+ if token in self.mergeable_ranks:
87
+ logger.info(f"extra token {token} exists, skipping")
88
+ continue
89
+ if index in used_ids:
90
+ logger.info(f'the index {index} for extra token {token} exists, skipping')
91
+ continue
92
+ self.mergeable_ranks[token] = index
93
+ # the index may be sparse after this, but don't worry tiktoken.Encoding will handle this
94
+
95
+ enc = tiktoken.Encoding(
96
+ "Qwen",
97
+ pat_str=PAT_STR,
98
+ mergeable_ranks=self.mergeable_ranks,
99
+ special_tokens=self.special_tokens,
100
+ )
101
+ assert (
102
+ len(self.mergeable_ranks) + len(self.special_tokens) == enc.n_vocab
103
+ ), f"{len(self.mergeable_ranks) + len(self.special_tokens)} != {enc.n_vocab} in encoding"
104
+
105
+ self.decoder = {
106
+ v: k for k, v in self.mergeable_ranks.items()
107
+ } # type: dict[int, bytes|str]
108
+ self.decoder.update({v: k for k, v in self.special_tokens.items()})
109
+
110
+ self.tokenizer = enc # type: tiktoken.Encoding
111
+
112
+ self.eod_id = self.tokenizer.eot_token
113
+ self.im_start_id = self.special_tokens[IMSTART]
114
+ self.im_end_id = self.special_tokens[IMEND]
115
+
116
+ def __getstate__(self):
117
+ # for pickle lovers
118
+ state = self.__dict__.copy()
119
+ del state["tokenizer"]
120
+ return state
121
+
122
+ def __setstate__(self, state):
123
+ # tokenizer is not python native; don't pass it; rebuild it
124
+ self.__dict__.update(state)
125
+ enc = tiktoken.Encoding(
126
+ "Qwen",
127
+ pat_str=PAT_STR,
128
+ mergeable_ranks=self.mergeable_ranks,
129
+ special_tokens=self.special_tokens,
130
+ )
131
+ self.tokenizer = enc
132
+
133
+ def __len__(self) -> int:
134
+ return self.tokenizer.n_vocab
135
+
136
+ def get_vocab(self) -> Dict[bytes, int]:
137
+ return self.mergeable_ranks
138
+
139
+ def convert_tokens_to_ids(
140
+ self, tokens: Union[bytes, str, List[Union[bytes, str]]]
141
+ ) -> List[int]:
142
+ ids = []
143
+ if isinstance(tokens, (str, bytes)):
144
+ if tokens in self.special_tokens:
145
+ return self.special_tokens[tokens]
146
+ else:
147
+ return self.mergeable_ranks.get(tokens)
148
+ for token in tokens:
149
+ if token in self.special_tokens:
150
+ ids.append(self.special_tokens[token])
151
+ else:
152
+ ids.append(self.mergeable_ranks.get(token))
153
+ return ids
154
+
155
+ def _add_tokens(
156
+ self,
157
+ new_tokens: Union[List[str], List[AddedToken]],
158
+ special_tokens: bool = False,
159
+ ) -> int:
160
+ if not special_tokens and new_tokens:
161
+ raise ValueError("Adding regular tokens is not supported")
162
+ for token in new_tokens:
163
+ surface_form = token.content if isinstance(token, AddedToken) else token
164
+ if surface_form not in SPECIAL_TOKENS_SET:
165
+ raise ValueError("Adding unknown special tokens is not supported")
166
+ return 0
167
+
168
+ def save_vocabulary(self, save_directory: str, **kwargs) -> Tuple[str]:
169
+ """
170
+ Save only the vocabulary of the tokenizer (vocabulary).
171
+
172
+ Returns:
173
+ `Tuple(str)`: Paths to the files saved.
174
+ """
175
+ file_path = os.path.join(save_directory, "qwen.tiktoken")
176
+ with open(file_path, "w", encoding="utf8") as w:
177
+ for k, v in self.mergeable_ranks.items():
178
+ line = base64.b64encode(k).decode("utf8") + " " + str(v) + "\n"
179
+ w.write(line)
180
+ return (file_path,)
181
+
182
+ def tokenize(
183
+ self,
184
+ text: str,
185
+ allowed_special: Union[Set, str] = "all",
186
+ disallowed_special: Union[Collection, str] = (),
187
+ **kwargs,
188
+ ) -> List[Union[bytes, str]]:
189
+ """
190
+ Converts a string in a sequence of tokens.
191
+
192
+ Args:
193
+ text (`str`):
194
+ The sequence to be encoded.
195
+ allowed_special (`Literal["all"]` or `set`):
196
+ The surface forms of the tokens to be encoded as special tokens in regular texts.
197
+ Default to "all".
198
+ disallowed_special (`Literal["all"]` or `Collection`):
199
+ The surface forms of the tokens that should not be in regular texts and trigger errors.
200
+ Default to an empty tuple.
201
+
202
+ kwargs (additional keyword arguments, *optional*):
203
+ Will be passed to the underlying model specific encode method.
204
+
205
+ Returns:
206
+ `List[bytes|str]`: The list of tokens.
207
+ """
208
+ tokens = []
209
+ text = unicodedata.normalize("NFC", text)
210
+
211
+ # this implementation takes a detour: text -> token id -> token surface forms
212
+ for t in self.tokenizer.encode(
213
+ text, allowed_special=allowed_special, disallowed_special=disallowed_special
214
+ ):
215
+ tokens.append(self.decoder[t])
216
+ return tokens
217
+
218
+ def convert_tokens_to_string(self, tokens: List[Union[bytes, str]]) -> str:
219
+ """
220
+ Converts a sequence of tokens in a single string.
221
+ """
222
+ text = ""
223
+ temp = b""
224
+ for t in tokens:
225
+ if isinstance(t, str):
226
+ if temp:
227
+ text += temp.decode("utf-8", errors=self.errors)
228
+ temp = b""
229
+ text += t
230
+ elif isinstance(t, bytes):
231
+ temp += t
232
+ else:
233
+ raise TypeError("token should only be of type types or str")
234
+ if temp:
235
+ text += temp.decode("utf-8", errors=self.errors)
236
+ return text
237
+
238
+ @property
239
+ def vocab_size(self):
240
+ return self.tokenizer.n_vocab
241
+
242
+ def _convert_id_to_token(self, index: int) -> Union[bytes, str]:
243
+ """Converts an id to a token, special tokens included"""
244
+ if index in self.decoder:
245
+ return self.decoder[index]
246
+ raise ValueError("unknown ids")
247
+
248
+ def _convert_token_to_id(self, token: Union[bytes, str]) -> int:
249
+ """Converts a token to an id using the vocab, special tokens included"""
250
+ if token in self.special_tokens:
251
+ return self.special_tokens[token]
252
+ if token in self.mergeable_ranks:
253
+ return self.mergeable_ranks[token]
254
+ raise ValueError("unknown token")
255
+
256
+ def _tokenize(self, text: str, **kwargs):
257
+ """
258
+ Converts a string in a sequence of tokens (string), using the tokenizer. Split in words for word-based
259
+ vocabulary or sub-words for sub-word-based vocabularies (BPE/SentencePieces/WordPieces).
260
+
261
+ Do NOT take care of added tokens.
262
+ """
263
+ raise NotImplementedError
264
+
265
+ def _decode(
266
+ self,
267
+ token_ids: Union[int, List[int]],
268
+ skip_special_tokens: bool = False,
269
+ errors: str = None,
270
+ **kwargs,
271
+ ) -> str:
272
+ if isinstance(token_ids, int):
273
+ token_ids = [token_ids]
274
+ if skip_special_tokens:
275
+ token_ids = [i for i in token_ids if i < self.eod_id]
276
+ return self.tokenizer.decode(token_ids, errors=errors or self.errors)
tokenizer_config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_max_length": 8192,
3
+ "tokenizer_class": "QWenTokenizer",
4
+ "auto_map": {
5
+ "AutoTokenizer": [
6
+ "tokenization_qwen.QWenTokenizer",
7
+ null
8
+ ]
9
+ }
10
+ }