prince-canuma commited on
Commit
8fdf423
1 Parent(s): 35de202

Upload folder using huggingface_hub

Browse files
README.md ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ tags:
4
+ - mlx
5
+ license_name: apple-sample-code-license
6
+ license_link: LICENSE
7
+ ---
8
+
9
+ # mlx-community/OpenELM-1_1B-instruct-8bit
10
+ This model was converted to MLX format from [`apple/OpenELM-1_1B-instruct`]() using mlx-lm version **0.10.0**.
11
+ Refer to the [original model card](https://huggingface.co/apple/OpenELM-1_1B-instruct) for more details on the model.
12
+ ## Use with mlx
13
+
14
+ ```bash
15
+ pip install mlx-lm
16
+ ```
17
+
18
+ ```python
19
+ from mlx_lm import load, generate
20
+
21
+ model, tokenizer = load("mlx-community/OpenELM-1_1B-instruct-8bit")
22
+ response = generate(model, tokenizer, prompt="hello", verbose=True)
23
+ ```
config.json ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "activation_fn_name": "swish",
3
+ "architectures": [
4
+ "OpenELMForCausalLM"
5
+ ],
6
+ "auto_map": {
7
+ "AutoConfig": "configuration_openelm.OpenELMConfig",
8
+ "AutoModelForCausalLM": "modeling_openelm.OpenELMForCausalLM"
9
+ },
10
+ "bos_token_id": 1,
11
+ "eos_token_id": 2,
12
+ "ffn_dim_divisor": 256,
13
+ "ffn_multipliers": [
14
+ 0.5,
15
+ 0.63,
16
+ 0.76,
17
+ 0.89,
18
+ 1.02,
19
+ 1.15,
20
+ 1.28,
21
+ 1.41,
22
+ 1.54,
23
+ 1.67,
24
+ 1.8,
25
+ 1.93,
26
+ 2.06,
27
+ 2.19,
28
+ 2.31,
29
+ 2.44,
30
+ 2.57,
31
+ 2.7,
32
+ 2.83,
33
+ 2.96,
34
+ 3.09,
35
+ 3.22,
36
+ 3.35,
37
+ 3.48,
38
+ 3.61,
39
+ 3.74,
40
+ 3.87,
41
+ 4.0
42
+ ],
43
+ "ffn_with_glu": true,
44
+ "head_dim": 64,
45
+ "initializer_range": 0.02,
46
+ "max_context_length": 2048,
47
+ "model_dim": 2048,
48
+ "model_type": "openelm",
49
+ "normalization_layer_name": "rms_norm",
50
+ "normalize_qk_projections": true,
51
+ "num_gqa_groups": 4,
52
+ "num_kv_heads": [
53
+ 4,
54
+ 4,
55
+ 4,
56
+ 5,
57
+ 5,
58
+ 5,
59
+ 5,
60
+ 5,
61
+ 5,
62
+ 5,
63
+ 6,
64
+ 6,
65
+ 6,
66
+ 6,
67
+ 6,
68
+ 6,
69
+ 6,
70
+ 6,
71
+ 7,
72
+ 7,
73
+ 7,
74
+ 7,
75
+ 7,
76
+ 7,
77
+ 8,
78
+ 8,
79
+ 8,
80
+ 8
81
+ ],
82
+ "num_query_heads": [
83
+ 16,
84
+ 16,
85
+ 16,
86
+ 20,
87
+ 20,
88
+ 20,
89
+ 20,
90
+ 20,
91
+ 20,
92
+ 20,
93
+ 24,
94
+ 24,
95
+ 24,
96
+ 24,
97
+ 24,
98
+ 24,
99
+ 24,
100
+ 24,
101
+ 28,
102
+ 28,
103
+ 28,
104
+ 28,
105
+ 28,
106
+ 28,
107
+ 32,
108
+ 32,
109
+ 32,
110
+ 32
111
+ ],
112
+ "num_transformer_layers": 28,
113
+ "qkv_multipliers": [
114
+ 0.5,
115
+ 1.0
116
+ ],
117
+ "quantization": {
118
+ "group_size": 64,
119
+ "bits": 8
120
+ },
121
+ "rope_freq_constant": 10000,
122
+ "rope_max_length": 4096,
123
+ "share_input_output_layers": true,
124
+ "torch_dtype": "bfloat16",
125
+ "transformers_version": "4.39.3",
126
+ "use_cache": true,
127
+ "vocab_size": 32000
128
+ }
configuration_openelm.py ADDED
@@ -0,0 +1,318 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #
2
+ # For licensing see accompanying LICENSE file.
3
+ # Copyright (C) 2024 Apple Inc. All Rights Reserved.
4
+ #
5
+
6
+ """Implements HF OpenELMConfig based on PretrainedConfig"""
7
+ from numbers import Number
8
+ from typing import List, Optional, Union
9
+
10
+ import numpy as np
11
+ from transformers import PretrainedConfig
12
+
13
+
14
+ def make_divisible(
15
+ v: Union[float, int],
16
+ divisor: Optional[int] = 8,
17
+ min_value: Optional[Union[float, int]] = None,
18
+ ) -> Union[float, int]:
19
+ """
20
+ This function is taken from the original tf repo.
21
+ It ensures that all layers have a channel number that is divisible by the divisor
22
+ It can be seen at:
23
+ https://github.com/tensorflow/models/blob/2cfc99eff5e5eb729c6793d2f3d03aa1c9be2b15/research/slim/nets/mobilenet/mobilenet.py#L62
24
+
25
+ Args:
26
+ v: input value
27
+ divisor: default to 8
28
+ min_value: minimum divisor value
29
+ Returns:
30
+ new_v: new divisible value
31
+ """
32
+ if min_value is None:
33
+ min_value = divisor
34
+ new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
35
+ # Make sure that round down does not go down by more than 10%.
36
+ if new_v < 0.9 * v:
37
+ new_v += divisor
38
+ return new_v
39
+
40
+
41
+ def compute_heads(model_dim: int, head_dim: int) -> int:
42
+ """Compute the number of heads.
43
+
44
+ Args:
45
+ model_dim: Model dimension.
46
+ head_dim: Head dimension.
47
+
48
+ Returns:
49
+ An integer denoting number of heads in multi-head attention is returned.
50
+
51
+ Raises:
52
+ ValueError: if model dimension is not divisible by head dimension.
53
+ """
54
+ if model_dim % head_dim == 0:
55
+ return model_dim // head_dim
56
+ else:
57
+ raise ValueError(
58
+ f"Model dimension should be divisible by head dimension. Got: {model_dim} and {head_dim}."
59
+ )
60
+
61
+
62
+ OpenELM_CONFIGS = {
63
+ "OpenELM-270M": dict(
64
+ num_transformer_layers=16,
65
+ model_dim=1280,
66
+ head_dim=64,
67
+ num_gqa_groups=4,
68
+ normalize_qk_projections=True,
69
+ share_input_output_layers=True,
70
+ # Vary the FFN and QKV multipliers to create variable FFN and attention layers respectively.
71
+ ffn_multipliers=(0.5, 4.0),
72
+ qkv_multipliers=(0.5, 1.0),
73
+ ),
74
+ "OpenELM-450M": dict(
75
+ num_transformer_layers=20,
76
+ model_dim=1536,
77
+ head_dim=64,
78
+ num_gqa_groups=4,
79
+ normalize_qk_projections=True,
80
+ share_input_output_layers=True,
81
+ # Vary the FFN and QKV multipliers to create variable FFN and attention layers respectively.
82
+ ffn_multipliers=(0.5, 4.0),
83
+ qkv_multipliers=(0.5, 1.0),
84
+ ),
85
+ "OpenELM-1_1B": dict(
86
+ num_transformer_layers=28,
87
+ model_dim=2048,
88
+ head_dim=64,
89
+ num_gqa_groups=4,
90
+ normalize_qk_projections=True,
91
+ share_input_output_layers=True,
92
+ # Vary the FFN and QKV multipliers to create variable FFN and attention layers respectively.
93
+ ffn_multipliers=(0.5, 4.0),
94
+ qkv_multipliers=(0.5, 1.0),
95
+ ),
96
+ "OpenELM-3B": dict(
97
+ num_transformer_layers=36,
98
+ model_dim=3072,
99
+ head_dim=128,
100
+ num_gqa_groups=4,
101
+ normalize_qk_projections=True,
102
+ share_input_output_layers=True,
103
+ # Vary the FFN and QKV multipliers to create variable FFN and attention layers respectively.
104
+ ffn_multipliers=(0.5, 4.0),
105
+ qkv_multipliers=(0.5, 1.0),
106
+ ),
107
+ }
108
+
109
+
110
+ class OpenELMConfig(PretrainedConfig):
111
+ r"""
112
+ This is the configuration class to store the configuration of a [`OpenELMModel`]. It is used to instantiate an OpenELM model according to the specified arguments, defining the model architecture.
113
+
114
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
115
+ documentation from [`PretrainedConfig`] for more information.
116
+
117
+ Args:
118
+ vocab_size (`int`, *optional*, defaults to 32000):
119
+ Vocabulary size of the OpenELM model.
120
+ max_context_length (`int`, *optional*, defaults to 2048):
121
+ Maximum number of input tokens.
122
+ num_transformer_layers (`int`, *optional*, defaults to 12):
123
+ Number of hidden layers in the Transformer decoder.
124
+ model_dim (`int`, *optional*, defaults to 2048):
125
+ Dimension of the hidden representations.
126
+ head_dim (`int`, *optional*, defaults to 128):
127
+ The attention head dimension.
128
+ qkv_multipliers (`Union[Number, List[Number]]`, *optional*, defaults to 1.0):
129
+ If the qkv_multipliers is a Number, then all attention layers have the same latent dimensions,
130
+ resulting in uniform allocation of parameters.
131
+ If the qkv_multipliers is a List of Number, then each attention layer have different latent dimensions
132
+ assuming qkv_multipliers[0] != qkv_multipliers[1]. This results in variable allocation of parameters in attention layer.
133
+ This scaling is known as layer-wise or block-wise scaling: https://arxiv.org/abs/2008.00623
134
+ num_query_heads (`Union[int, None]`, *optional*, defaults to None):
135
+ The number of query heads, computed from `compute_heads(model_dim=model_dim, head_dim=head_dim)`.
136
+ num_gqa_groups (`int`, *optional*, defaults to 1):
137
+ This variable allows to switch between multi-head attention, group query attention, and multi-query attention.
138
+ When num_gqa_groups == 1, then it is multi-head attention.
139
+ When 1 < num_gqa_groups < num_heads and num_heads is divisible by num_gqa_groups, then it is group query attention
140
+ When num_gqa_groups == num_heads, then it is multi-query attention
141
+ ffn_multipliers (`Union[Number, List[Number]]`, *optional*, defaults to 4.0):
142
+ Feed-forward network (FFN) multipliers.
143
+ If the ffn_multipliers is a Number, then all FFN layers have the same latent dimensions,
144
+ resulting in uniform allocation of parameters.
145
+ If the ffn_multipliers is a List of Number, then each FFN layer have different latent dimensions
146
+ assuming ffn_multipliers[0] != ffn_multipliers[1]. This results in variable allocation of parameters in FFN layer.
147
+ This scaling is known as layer-wise or block-wise scaling: https://arxiv.org/abs/2008.00623
148
+ ffn_with_glu (`bool`, *optional*, defaults to True):
149
+ Whether to use FFN with Gated Linear Unit (GLU)
150
+ ffn_dim_divisor (`int`, *optional*, defaults to 256):
151
+ The ffn layer dimension divisor.
152
+ activation_fn_name (`str` or `function`, *optional*, defaults to `"swish"`):
153
+ The non-linear activation function (function or string) in the decoder.
154
+ normalization_layer_name (`str` or `function`, *optional*, defaults to `"rms_norm"`):
155
+ Type of normalization layer.
156
+ normalize_qk_projections (`bool`, *optional*, defaults to False):
157
+ Whether to normalize queries and keys after projections
158
+ share_input_output_layers (`bool`, *optional*, defaults to False):
159
+ Whether to share the embedding between input and output linear layer
160
+ rope_freq_constant (`int`, *optional*, defaults to 10000):
161
+ The base period of the RoPE embeddings.
162
+ rope_max_length (`int`, *optional*, defaults to 4096):
163
+ That rope_max_length is set to twice of max_context_length.
164
+ This allows flexibility in token lengths during training or fine-tuning.
165
+ initializer_range (`float`, *optional*, defaults to 0.02):
166
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
167
+ use_cache (`bool`, *optional*, defaults to `True`):
168
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
169
+ relevant if `config.is_decoder=True`.
170
+ bos_token_id (`int`, *optional*, defaults to 2):
171
+ Beginning of stream token id.
172
+ eos_token_id (`int`, *optional*, defaults to 1):
173
+ End of stream token id.
174
+ """
175
+
176
+ model_type = "openelm"
177
+
178
+ def __init__(
179
+ self,
180
+ vocab_size: int = 32000,
181
+ max_context_length: int = 2048,
182
+ num_transformer_layers: int = 12,
183
+ model_dim: int = 2048,
184
+ head_dim: int = 128,
185
+ qkv_multipliers: Union[Number, List[Number]] = 1.0,
186
+ num_query_heads: Union[int, None] = None,
187
+ num_gqa_groups: int = 1,
188
+ ffn_multipliers: Union[Number, List[Number]] = 4.0,
189
+ ffn_with_glu: bool = True,
190
+ ffn_dim_divisor: int = 256,
191
+ activation_fn_name: str = "swish",
192
+ normalization_layer_name: str = "rms_norm",
193
+ normalize_qk_projections: bool = False,
194
+ share_input_output_layers: bool = False,
195
+ rope_freq_constant: int = 10000,
196
+ rope_max_length: int = 4096,
197
+ initializer_range: float = 0.02,
198
+ use_cache: bool = True,
199
+ bos_token_id: int = 1,
200
+ eos_token_id: int = 2,
201
+ **kwargs,
202
+ ) -> None:
203
+ self.vocab_size = vocab_size
204
+ self.max_context_length = max_context_length
205
+ self.num_transformer_layers = num_transformer_layers
206
+ self.model_dim = model_dim
207
+ self.head_dim = head_dim
208
+ self.qkv_multipliers = qkv_multipliers
209
+ self.num_query_heads = num_query_heads
210
+ self.num_gqa_groups = num_gqa_groups
211
+ self.ffn_multipliers = ffn_multipliers
212
+ self.ffn_with_glu = ffn_with_glu
213
+ self.ffn_dim_divisor = ffn_dim_divisor
214
+ self.activation_fn_name = activation_fn_name
215
+ self.normalization_layer_name = normalization_layer_name
216
+ self.normalize_qk_projections = normalize_qk_projections
217
+ self.share_input_output_layers = share_input_output_layers
218
+ self.rope_freq_constant = rope_freq_constant
219
+ self.rope_max_length = rope_max_length
220
+ self.num_query_heads = (
221
+ compute_heads(model_dim=model_dim, head_dim=head_dim)
222
+ if num_query_heads is None
223
+ else num_query_heads
224
+ )
225
+ self.initializer_range = initializer_range
226
+
227
+ self.__post_init__()
228
+ super().__init__(
229
+ use_cache=use_cache,
230
+ bos_token_id=bos_token_id,
231
+ eos_token_id=eos_token_id,
232
+ **kwargs,
233
+ )
234
+
235
+ def __post_init__(self) -> None:
236
+ if self.num_gqa_groups is not None:
237
+ head_multiple_of = self.num_gqa_groups
238
+ else:
239
+ head_multiple_of = 2
240
+
241
+ if isinstance(self.qkv_multipliers, Number):
242
+ # All attention layers have the same latent dimensions, resulting in uniform allocation of parameters.
243
+ qkv_dim = make_divisible(
244
+ self.model_dim * self.qkv_multipliers,
245
+ divisor=self.head_dim * head_multiple_of,
246
+ )
247
+ query_dims = [int(qkv_dim)] * self.num_transformer_layers
248
+
249
+ elif (
250
+ isinstance(self.qkv_multipliers, (tuple, list))
251
+ and len(self.qkv_multipliers) == 2
252
+ ):
253
+ # Each attention layer have different latent dimensions assuming qkv_multipliers[0] != qkv_multipliers[1].
254
+ # This results in variable allocation of parameters in attention layer.
255
+ # This scaling is known as layer-wise or block-wise scaling: https://arxiv.org/abs/2008.00623
256
+ qkv_multipliers = [
257
+ round(v, 2)
258
+ for v in np.linspace(
259
+ self.qkv_multipliers[0],
260
+ self.qkv_multipliers[1],
261
+ num=self.num_transformer_layers,
262
+ dtype=float,
263
+ )
264
+ ]
265
+ # Make sure that scaled model dimension is divisible by scaled head dimension.
266
+ query_dims = [
267
+ int(
268
+ make_divisible(
269
+ self.model_dim * m, divisor=self.head_dim * head_multiple_of
270
+ )
271
+ )
272
+ for m in qkv_multipliers
273
+ ]
274
+ else:
275
+ raise NotImplementedError(
276
+ f"QKV multipliers should be a single number or a list containing exactly two numbers. Got: {qkv_multipliers}."
277
+ )
278
+
279
+ # compute the number of query, key, and value heads
280
+ # For multi-head and multi-query attention, the number of heads for query, key, and value are the same.
281
+ # For group query attention, the number of key and value heads are the same.
282
+ self.num_query_heads = [
283
+ int(compute_heads(q_dim, self.head_dim)) for q_dim in query_dims
284
+ ]
285
+ self.num_kv_heads = [
286
+ q_heads // self.num_gqa_groups for q_heads in self.num_query_heads
287
+ ]
288
+
289
+ # Feed-forward network (FFN) multipliers
290
+ if isinstance(self.ffn_multipliers, Number):
291
+ # All FFN layers have the same latent dimensions, resulting in uniform allocation of parameters.
292
+ self.ffn_multipliers = [self.ffn_multipliers] * self.num_transformer_layers
293
+ elif isinstance(self.ffn_multipliers, (tuple, list)):
294
+ # Each FFN layer have different latent dimensions assuming ffn_multipliers[0] != ffn_multipliers[1].
295
+ # This results in variable allocation of parameters in FFN layer.
296
+ # This scaling is known as layer-wise or block-wise scaling: https://arxiv.org/abs/2008.00623
297
+ if len(self.ffn_multipliers) == 2:
298
+ self.ffn_multipliers = [
299
+ round(v, 2)
300
+ for v in np.linspace(
301
+ self.ffn_multipliers[0],
302
+ self.ffn_multipliers[1],
303
+ num=self.num_transformer_layers,
304
+ dtype=float,
305
+ )
306
+ ]
307
+ else:
308
+ assert (
309
+ len(self.ffn_multipliers) == self.num_transformer_layers
310
+ ), f"{len(self.ffn_multipliers)=}!={self.num_transformer_layers=}"
311
+ else:
312
+ raise NotImplementedError(
313
+ f"FFN multipliers should be a single number or a list containing exactly two numbers. Got: {qkv_multipliers}."
314
+ )
315
+
316
+ # check num_query_heads divisible by num_kv_heads for every layer
317
+ for layer_idx in range(len(query_dims)):
318
+ assert self.num_query_heads[layer_idx] % self.num_kv_heads[layer_idx] == 0
generate_openelm.py ADDED
@@ -0,0 +1,240 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #
2
+ # For licensing see accompanying LICENSE file.
3
+ # Copyright (C) 2024 Apple Inc. All Rights Reserved.
4
+ #
5
+
6
+ """Module to generate OpenELM output given a model and an input prompt."""
7
+ import os
8
+ import logging
9
+ import time
10
+ import argparse
11
+ from typing import Optional, Union
12
+ import torch
13
+
14
+ from transformers import AutoTokenizer, AutoModelForCausalLM
15
+
16
+
17
+ def generate(
18
+ prompt: str,
19
+ model: Union[str, AutoModelForCausalLM],
20
+ hf_access_token: str = None,
21
+ tokenizer: Union[str, AutoTokenizer] = 'meta-llama/Llama-2-7b-hf',
22
+ device: Optional[str] = None,
23
+ max_length: int = 1024,
24
+ assistant_model: Optional[Union[str, AutoModelForCausalLM]] = None,
25
+ generate_kwargs: Optional[dict] = None,
26
+ ) -> str:
27
+ """ Generates output given a prompt.
28
+
29
+ Args:
30
+ prompt: The string prompt.
31
+ model: The LLM Model. If a string is passed, it should be the path to
32
+ the hf converted checkpoint.
33
+ hf_access_token: Hugging face access token.
34
+ tokenizer: Tokenizer instance. If model is set as a string path,
35
+ the tokenizer will be loaded from the checkpoint.
36
+ device: String representation of device to run the model on. If None
37
+ and cuda available it would be set to cuda:0 else cpu.
38
+ max_length: Maximum length of tokens, input prompt + generated tokens.
39
+ assistant_model: If set, this model will be used for
40
+ speculative generation. If a string is passed, it should be the
41
+ path to the hf converted checkpoint.
42
+ generate_kwargs: Extra kwargs passed to the hf generate function.
43
+
44
+ Returns:
45
+ output_text: output generated as a string.
46
+ generation_time: generation time in seconds.
47
+
48
+ Raises:
49
+ ValueError: If device is set to CUDA but no CUDA device is detected.
50
+ ValueError: If tokenizer is not set.
51
+ ValueError: If hf_access_token is not specified.
52
+ """
53
+ if not device:
54
+ if torch.cuda.is_available() and torch.cuda.device_count():
55
+ device = "cuda:0"
56
+ logging.warning(
57
+ 'inference device is not set, using cuda:0, %s',
58
+ torch.cuda.get_device_name(0)
59
+ )
60
+ else:
61
+ device = 'cpu'
62
+ logging.warning(
63
+ (
64
+ 'No CUDA device detected, using cpu, '
65
+ 'expect slower speeds.'
66
+ )
67
+ )
68
+
69
+ if 'cuda' in device and not torch.cuda.is_available():
70
+ raise ValueError('CUDA device requested but no CUDA device detected.')
71
+
72
+ if not tokenizer:
73
+ raise ValueError('Tokenizer is not set in the generate function.')
74
+
75
+ if not hf_access_token:
76
+ raise ValueError((
77
+ 'Hugging face access token needs to be specified. '
78
+ 'Please refer to https://huggingface.co/docs/hub/security-tokens'
79
+ ' to obtain one.'
80
+ )
81
+ )
82
+
83
+ if isinstance(model, str):
84
+ checkpoint_path = model
85
+ model = AutoModelForCausalLM.from_pretrained(
86
+ checkpoint_path,
87
+ trust_remote_code=True
88
+ )
89
+ model.to(device).eval()
90
+ if isinstance(tokenizer, str):
91
+ tokenizer = AutoTokenizer.from_pretrained(
92
+ tokenizer,
93
+ token=hf_access_token,
94
+ )
95
+
96
+ # Speculative mode
97
+ draft_model = None
98
+ if assistant_model:
99
+ draft_model = assistant_model
100
+ if isinstance(assistant_model, str):
101
+ draft_model = AutoModelForCausalLM.from_pretrained(
102
+ assistant_model,
103
+ trust_remote_code=True
104
+ )
105
+ draft_model.to(device).eval()
106
+
107
+ # Prepare the prompt
108
+ tokenized_prompt = tokenizer(prompt)
109
+ tokenized_prompt = torch.tensor(
110
+ tokenized_prompt['input_ids'],
111
+ device=device
112
+ )
113
+
114
+ tokenized_prompt = tokenized_prompt.unsqueeze(0)
115
+
116
+ # Generate
117
+ stime = time.time()
118
+ output_ids = model.generate(
119
+ tokenized_prompt,
120
+ max_length=max_length,
121
+ pad_token_id=0,
122
+ assistant_model=draft_model,
123
+ **(generate_kwargs if generate_kwargs else {}),
124
+ )
125
+ generation_time = time.time() - stime
126
+
127
+ output_text = tokenizer.decode(
128
+ output_ids[0].tolist(),
129
+ skip_special_tokens=True
130
+ )
131
+
132
+ return output_text, generation_time
133
+
134
+
135
+ def openelm_generate_parser():
136
+ """Argument Parser"""
137
+
138
+ class KwargsParser(argparse.Action):
139
+ """Parser action class to parse kwargs of form key=value"""
140
+ def __call__(self, parser, namespace, values, option_string=None):
141
+ setattr(namespace, self.dest, dict())
142
+ for val in values:
143
+ if '=' not in val:
144
+ raise ValueError(
145
+ (
146
+ 'Argument parsing error, kwargs are expected in'
147
+ ' the form of key=value.'
148
+ )
149
+ )
150
+ kwarg_k, kwarg_v = val.split('=')
151
+ try:
152
+ converted_v = int(kwarg_v)
153
+ except ValueError:
154
+ try:
155
+ converted_v = float(kwarg_v)
156
+ except ValueError:
157
+ converted_v = kwarg_v
158
+ getattr(namespace, self.dest)[kwarg_k] = converted_v
159
+
160
+ parser = argparse.ArgumentParser('OpenELM Generate Module')
161
+ parser.add_argument(
162
+ '--model',
163
+ dest='model',
164
+ help='Path to the hf converted model.',
165
+ required=True,
166
+ type=str,
167
+ )
168
+ parser.add_argument(
169
+ '--hf_access_token',
170
+ dest='hf_access_token',
171
+ help='Hugging face access token, starting with "hf_".',
172
+ type=str,
173
+ )
174
+ parser.add_argument(
175
+ '--prompt',
176
+ dest='prompt',
177
+ help='Prompt for LLM call.',
178
+ default='',
179
+ type=str,
180
+ )
181
+ parser.add_argument(
182
+ '--device',
183
+ dest='device',
184
+ help='Device used for inference.',
185
+ type=str,
186
+ )
187
+ parser.add_argument(
188
+ '--max_length',
189
+ dest='max_length',
190
+ help='Maximum length of tokens.',
191
+ default=256,
192
+ type=int,
193
+ )
194
+ parser.add_argument(
195
+ '--assistant_model',
196
+ dest='assistant_model',
197
+ help=(
198
+ (
199
+ 'If set, this is used as a draft model '
200
+ 'for assisted speculative generation.'
201
+ )
202
+ ),
203
+ type=str,
204
+ )
205
+ parser.add_argument(
206
+ '--generate_kwargs',
207
+ dest='generate_kwargs',
208
+ help='Additional kwargs passed to the HF generate function.',
209
+ type=str,
210
+ nargs='*',
211
+ action=KwargsParser,
212
+ )
213
+ return parser.parse_args()
214
+
215
+
216
+ if __name__ == '__main__':
217
+ args = openelm_generate_parser()
218
+ prompt = args.prompt
219
+
220
+ output_text, genertaion_time = generate(
221
+ prompt=prompt,
222
+ model=args.model,
223
+ device=args.device,
224
+ max_length=args.max_length,
225
+ assistant_model=args.assistant_model,
226
+ generate_kwargs=args.generate_kwargs,
227
+ hf_access_token=args.hf_access_token,
228
+ )
229
+
230
+ print_txt = (
231
+ f'\r\n{"=" * os.get_terminal_size().columns}\r\n'
232
+ '\033[1m Prompt + Generated Output\033[0m\r\n'
233
+ f'{"-" * os.get_terminal_size().columns}\r\n'
234
+ f'{output_text}\r\n'
235
+ f'{"-" * os.get_terminal_size().columns}\r\n'
236
+ '\r\nGeneration took'
237
+ f'\033[1m\033[92m {round(genertaion_time, 2)} \033[0m'
238
+ 'seconds.\r\n'
239
+ )
240
+ print(print_txt)
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fdbc9b1fff3c1d0fffeb5063d8296449a57a0d1bc7b3c25b59d89f898b986ada
3
+ size 1147548674
model.safetensors.index.json ADDED
@@ -0,0 +1,459 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 1147497472
4
+ },
5
+ "weight_map": {
6
+ "transformer.layers.0.attn.k_norm.weight": "model.safetensors",
7
+ "transformer.layers.0.attn.out_proj.biases": "model.safetensors",
8
+ "transformer.layers.0.attn.out_proj.scales": "model.safetensors",
9
+ "transformer.layers.0.attn.out_proj.weight": "model.safetensors",
10
+ "transformer.layers.0.attn.q_norm.weight": "model.safetensors",
11
+ "transformer.layers.0.attn.qkv_proj.biases": "model.safetensors",
12
+ "transformer.layers.0.attn.qkv_proj.scales": "model.safetensors",
13
+ "transformer.layers.0.attn.qkv_proj.weight": "model.safetensors",
14
+ "transformer.layers.0.attn_norm.weight": "model.safetensors",
15
+ "transformer.layers.0.ffn.proj_1.biases": "model.safetensors",
16
+ "transformer.layers.0.ffn.proj_1.scales": "model.safetensors",
17
+ "transformer.layers.0.ffn.proj_1.weight": "model.safetensors",
18
+ "transformer.layers.0.ffn.proj_2.biases": "model.safetensors",
19
+ "transformer.layers.0.ffn.proj_2.scales": "model.safetensors",
20
+ "transformer.layers.0.ffn.proj_2.weight": "model.safetensors",
21
+ "transformer.layers.0.ffn_norm.weight": "model.safetensors",
22
+ "transformer.layers.1.attn.k_norm.weight": "model.safetensors",
23
+ "transformer.layers.1.attn.out_proj.biases": "model.safetensors",
24
+ "transformer.layers.1.attn.out_proj.scales": "model.safetensors",
25
+ "transformer.layers.1.attn.out_proj.weight": "model.safetensors",
26
+ "transformer.layers.1.attn.q_norm.weight": "model.safetensors",
27
+ "transformer.layers.1.attn.qkv_proj.biases": "model.safetensors",
28
+ "transformer.layers.1.attn.qkv_proj.scales": "model.safetensors",
29
+ "transformer.layers.1.attn.qkv_proj.weight": "model.safetensors",
30
+ "transformer.layers.1.attn_norm.weight": "model.safetensors",
31
+ "transformer.layers.1.ffn.proj_1.biases": "model.safetensors",
32
+ "transformer.layers.1.ffn.proj_1.scales": "model.safetensors",
33
+ "transformer.layers.1.ffn.proj_1.weight": "model.safetensors",
34
+ "transformer.layers.1.ffn.proj_2.biases": "model.safetensors",
35
+ "transformer.layers.1.ffn.proj_2.scales": "model.safetensors",
36
+ "transformer.layers.1.ffn.proj_2.weight": "model.safetensors",
37
+ "transformer.layers.1.ffn_norm.weight": "model.safetensors",
38
+ "transformer.layers.10.attn.k_norm.weight": "model.safetensors",
39
+ "transformer.layers.10.attn.out_proj.biases": "model.safetensors",
40
+ "transformer.layers.10.attn.out_proj.scales": "model.safetensors",
41
+ "transformer.layers.10.attn.out_proj.weight": "model.safetensors",
42
+ "transformer.layers.10.attn.q_norm.weight": "model.safetensors",
43
+ "transformer.layers.10.attn.qkv_proj.biases": "model.safetensors",
44
+ "transformer.layers.10.attn.qkv_proj.scales": "model.safetensors",
45
+ "transformer.layers.10.attn.qkv_proj.weight": "model.safetensors",
46
+ "transformer.layers.10.attn_norm.weight": "model.safetensors",
47
+ "transformer.layers.10.ffn.proj_1.biases": "model.safetensors",
48
+ "transformer.layers.10.ffn.proj_1.scales": "model.safetensors",
49
+ "transformer.layers.10.ffn.proj_1.weight": "model.safetensors",
50
+ "transformer.layers.10.ffn.proj_2.biases": "model.safetensors",
51
+ "transformer.layers.10.ffn.proj_2.scales": "model.safetensors",
52
+ "transformer.layers.10.ffn.proj_2.weight": "model.safetensors",
53
+ "transformer.layers.10.ffn_norm.weight": "model.safetensors",
54
+ "transformer.layers.11.attn.k_norm.weight": "model.safetensors",
55
+ "transformer.layers.11.attn.out_proj.biases": "model.safetensors",
56
+ "transformer.layers.11.attn.out_proj.scales": "model.safetensors",
57
+ "transformer.layers.11.attn.out_proj.weight": "model.safetensors",
58
+ "transformer.layers.11.attn.q_norm.weight": "model.safetensors",
59
+ "transformer.layers.11.attn.qkv_proj.biases": "model.safetensors",
60
+ "transformer.layers.11.attn.qkv_proj.scales": "model.safetensors",
61
+ "transformer.layers.11.attn.qkv_proj.weight": "model.safetensors",
62
+ "transformer.layers.11.attn_norm.weight": "model.safetensors",
63
+ "transformer.layers.11.ffn.proj_1.biases": "model.safetensors",
64
+ "transformer.layers.11.ffn.proj_1.scales": "model.safetensors",
65
+ "transformer.layers.11.ffn.proj_1.weight": "model.safetensors",
66
+ "transformer.layers.11.ffn.proj_2.biases": "model.safetensors",
67
+ "transformer.layers.11.ffn.proj_2.scales": "model.safetensors",
68
+ "transformer.layers.11.ffn.proj_2.weight": "model.safetensors",
69
+ "transformer.layers.11.ffn_norm.weight": "model.safetensors",
70
+ "transformer.layers.12.attn.k_norm.weight": "model.safetensors",
71
+ "transformer.layers.12.attn.out_proj.biases": "model.safetensors",
72
+ "transformer.layers.12.attn.out_proj.scales": "model.safetensors",
73
+ "transformer.layers.12.attn.out_proj.weight": "model.safetensors",
74
+ "transformer.layers.12.attn.q_norm.weight": "model.safetensors",
75
+ "transformer.layers.12.attn.qkv_proj.biases": "model.safetensors",
76
+ "transformer.layers.12.attn.qkv_proj.scales": "model.safetensors",
77
+ "transformer.layers.12.attn.qkv_proj.weight": "model.safetensors",
78
+ "transformer.layers.12.attn_norm.weight": "model.safetensors",
79
+ "transformer.layers.12.ffn.proj_1.biases": "model.safetensors",
80
+ "transformer.layers.12.ffn.proj_1.scales": "model.safetensors",
81
+ "transformer.layers.12.ffn.proj_1.weight": "model.safetensors",
82
+ "transformer.layers.12.ffn.proj_2.biases": "model.safetensors",
83
+ "transformer.layers.12.ffn.proj_2.scales": "model.safetensors",
84
+ "transformer.layers.12.ffn.proj_2.weight": "model.safetensors",
85
+ "transformer.layers.12.ffn_norm.weight": "model.safetensors",
86
+ "transformer.layers.13.attn.k_norm.weight": "model.safetensors",
87
+ "transformer.layers.13.attn.out_proj.biases": "model.safetensors",
88
+ "transformer.layers.13.attn.out_proj.scales": "model.safetensors",
89
+ "transformer.layers.13.attn.out_proj.weight": "model.safetensors",
90
+ "transformer.layers.13.attn.q_norm.weight": "model.safetensors",
91
+ "transformer.layers.13.attn.qkv_proj.biases": "model.safetensors",
92
+ "transformer.layers.13.attn.qkv_proj.scales": "model.safetensors",
93
+ "transformer.layers.13.attn.qkv_proj.weight": "model.safetensors",
94
+ "transformer.layers.13.attn_norm.weight": "model.safetensors",
95
+ "transformer.layers.13.ffn.proj_1.biases": "model.safetensors",
96
+ "transformer.layers.13.ffn.proj_1.scales": "model.safetensors",
97
+ "transformer.layers.13.ffn.proj_1.weight": "model.safetensors",
98
+ "transformer.layers.13.ffn.proj_2.biases": "model.safetensors",
99
+ "transformer.layers.13.ffn.proj_2.scales": "model.safetensors",
100
+ "transformer.layers.13.ffn.proj_2.weight": "model.safetensors",
101
+ "transformer.layers.13.ffn_norm.weight": "model.safetensors",
102
+ "transformer.layers.14.attn.k_norm.weight": "model.safetensors",
103
+ "transformer.layers.14.attn.out_proj.biases": "model.safetensors",
104
+ "transformer.layers.14.attn.out_proj.scales": "model.safetensors",
105
+ "transformer.layers.14.attn.out_proj.weight": "model.safetensors",
106
+ "transformer.layers.14.attn.q_norm.weight": "model.safetensors",
107
+ "transformer.layers.14.attn.qkv_proj.biases": "model.safetensors",
108
+ "transformer.layers.14.attn.qkv_proj.scales": "model.safetensors",
109
+ "transformer.layers.14.attn.qkv_proj.weight": "model.safetensors",
110
+ "transformer.layers.14.attn_norm.weight": "model.safetensors",
111
+ "transformer.layers.14.ffn.proj_1.biases": "model.safetensors",
112
+ "transformer.layers.14.ffn.proj_1.scales": "model.safetensors",
113
+ "transformer.layers.14.ffn.proj_1.weight": "model.safetensors",
114
+ "transformer.layers.14.ffn.proj_2.biases": "model.safetensors",
115
+ "transformer.layers.14.ffn.proj_2.scales": "model.safetensors",
116
+ "transformer.layers.14.ffn.proj_2.weight": "model.safetensors",
117
+ "transformer.layers.14.ffn_norm.weight": "model.safetensors",
118
+ "transformer.layers.15.attn.k_norm.weight": "model.safetensors",
119
+ "transformer.layers.15.attn.out_proj.biases": "model.safetensors",
120
+ "transformer.layers.15.attn.out_proj.scales": "model.safetensors",
121
+ "transformer.layers.15.attn.out_proj.weight": "model.safetensors",
122
+ "transformer.layers.15.attn.q_norm.weight": "model.safetensors",
123
+ "transformer.layers.15.attn.qkv_proj.biases": "model.safetensors",
124
+ "transformer.layers.15.attn.qkv_proj.scales": "model.safetensors",
125
+ "transformer.layers.15.attn.qkv_proj.weight": "model.safetensors",
126
+ "transformer.layers.15.attn_norm.weight": "model.safetensors",
127
+ "transformer.layers.15.ffn.proj_1.biases": "model.safetensors",
128
+ "transformer.layers.15.ffn.proj_1.scales": "model.safetensors",
129
+ "transformer.layers.15.ffn.proj_1.weight": "model.safetensors",
130
+ "transformer.layers.15.ffn.proj_2.biases": "model.safetensors",
131
+ "transformer.layers.15.ffn.proj_2.scales": "model.safetensors",
132
+ "transformer.layers.15.ffn.proj_2.weight": "model.safetensors",
133
+ "transformer.layers.15.ffn_norm.weight": "model.safetensors",
134
+ "transformer.layers.16.attn.k_norm.weight": "model.safetensors",
135
+ "transformer.layers.16.attn.out_proj.biases": "model.safetensors",
136
+ "transformer.layers.16.attn.out_proj.scales": "model.safetensors",
137
+ "transformer.layers.16.attn.out_proj.weight": "model.safetensors",
138
+ "transformer.layers.16.attn.q_norm.weight": "model.safetensors",
139
+ "transformer.layers.16.attn.qkv_proj.biases": "model.safetensors",
140
+ "transformer.layers.16.attn.qkv_proj.scales": "model.safetensors",
141
+ "transformer.layers.16.attn.qkv_proj.weight": "model.safetensors",
142
+ "transformer.layers.16.attn_norm.weight": "model.safetensors",
143
+ "transformer.layers.16.ffn.proj_1.biases": "model.safetensors",
144
+ "transformer.layers.16.ffn.proj_1.scales": "model.safetensors",
145
+ "transformer.layers.16.ffn.proj_1.weight": "model.safetensors",
146
+ "transformer.layers.16.ffn.proj_2.biases": "model.safetensors",
147
+ "transformer.layers.16.ffn.proj_2.scales": "model.safetensors",
148
+ "transformer.layers.16.ffn.proj_2.weight": "model.safetensors",
149
+ "transformer.layers.16.ffn_norm.weight": "model.safetensors",
150
+ "transformer.layers.17.attn.k_norm.weight": "model.safetensors",
151
+ "transformer.layers.17.attn.out_proj.biases": "model.safetensors",
152
+ "transformer.layers.17.attn.out_proj.scales": "model.safetensors",
153
+ "transformer.layers.17.attn.out_proj.weight": "model.safetensors",
154
+ "transformer.layers.17.attn.q_norm.weight": "model.safetensors",
155
+ "transformer.layers.17.attn.qkv_proj.biases": "model.safetensors",
156
+ "transformer.layers.17.attn.qkv_proj.scales": "model.safetensors",
157
+ "transformer.layers.17.attn.qkv_proj.weight": "model.safetensors",
158
+ "transformer.layers.17.attn_norm.weight": "model.safetensors",
159
+ "transformer.layers.17.ffn.proj_1.biases": "model.safetensors",
160
+ "transformer.layers.17.ffn.proj_1.scales": "model.safetensors",
161
+ "transformer.layers.17.ffn.proj_1.weight": "model.safetensors",
162
+ "transformer.layers.17.ffn.proj_2.biases": "model.safetensors",
163
+ "transformer.layers.17.ffn.proj_2.scales": "model.safetensors",
164
+ "transformer.layers.17.ffn.proj_2.weight": "model.safetensors",
165
+ "transformer.layers.17.ffn_norm.weight": "model.safetensors",
166
+ "transformer.layers.18.attn.k_norm.weight": "model.safetensors",
167
+ "transformer.layers.18.attn.out_proj.biases": "model.safetensors",
168
+ "transformer.layers.18.attn.out_proj.scales": "model.safetensors",
169
+ "transformer.layers.18.attn.out_proj.weight": "model.safetensors",
170
+ "transformer.layers.18.attn.q_norm.weight": "model.safetensors",
171
+ "transformer.layers.18.attn.qkv_proj.biases": "model.safetensors",
172
+ "transformer.layers.18.attn.qkv_proj.scales": "model.safetensors",
173
+ "transformer.layers.18.attn.qkv_proj.weight": "model.safetensors",
174
+ "transformer.layers.18.attn_norm.weight": "model.safetensors",
175
+ "transformer.layers.18.ffn.proj_1.biases": "model.safetensors",
176
+ "transformer.layers.18.ffn.proj_1.scales": "model.safetensors",
177
+ "transformer.layers.18.ffn.proj_1.weight": "model.safetensors",
178
+ "transformer.layers.18.ffn.proj_2.biases": "model.safetensors",
179
+ "transformer.layers.18.ffn.proj_2.scales": "model.safetensors",
180
+ "transformer.layers.18.ffn.proj_2.weight": "model.safetensors",
181
+ "transformer.layers.18.ffn_norm.weight": "model.safetensors",
182
+ "transformer.layers.19.attn.k_norm.weight": "model.safetensors",
183
+ "transformer.layers.19.attn.out_proj.biases": "model.safetensors",
184
+ "transformer.layers.19.attn.out_proj.scales": "model.safetensors",
185
+ "transformer.layers.19.attn.out_proj.weight": "model.safetensors",
186
+ "transformer.layers.19.attn.q_norm.weight": "model.safetensors",
187
+ "transformer.layers.19.attn.qkv_proj.biases": "model.safetensors",
188
+ "transformer.layers.19.attn.qkv_proj.scales": "model.safetensors",
189
+ "transformer.layers.19.attn.qkv_proj.weight": "model.safetensors",
190
+ "transformer.layers.19.attn_norm.weight": "model.safetensors",
191
+ "transformer.layers.19.ffn.proj_1.biases": "model.safetensors",
192
+ "transformer.layers.19.ffn.proj_1.scales": "model.safetensors",
193
+ "transformer.layers.19.ffn.proj_1.weight": "model.safetensors",
194
+ "transformer.layers.19.ffn.proj_2.biases": "model.safetensors",
195
+ "transformer.layers.19.ffn.proj_2.scales": "model.safetensors",
196
+ "transformer.layers.19.ffn.proj_2.weight": "model.safetensors",
197
+ "transformer.layers.19.ffn_norm.weight": "model.safetensors",
198
+ "transformer.layers.2.attn.k_norm.weight": "model.safetensors",
199
+ "transformer.layers.2.attn.out_proj.biases": "model.safetensors",
200
+ "transformer.layers.2.attn.out_proj.scales": "model.safetensors",
201
+ "transformer.layers.2.attn.out_proj.weight": "model.safetensors",
202
+ "transformer.layers.2.attn.q_norm.weight": "model.safetensors",
203
+ "transformer.layers.2.attn.qkv_proj.biases": "model.safetensors",
204
+ "transformer.layers.2.attn.qkv_proj.scales": "model.safetensors",
205
+ "transformer.layers.2.attn.qkv_proj.weight": "model.safetensors",
206
+ "transformer.layers.2.attn_norm.weight": "model.safetensors",
207
+ "transformer.layers.2.ffn.proj_1.biases": "model.safetensors",
208
+ "transformer.layers.2.ffn.proj_1.scales": "model.safetensors",
209
+ "transformer.layers.2.ffn.proj_1.weight": "model.safetensors",
210
+ "transformer.layers.2.ffn.proj_2.biases": "model.safetensors",
211
+ "transformer.layers.2.ffn.proj_2.scales": "model.safetensors",
212
+ "transformer.layers.2.ffn.proj_2.weight": "model.safetensors",
213
+ "transformer.layers.2.ffn_norm.weight": "model.safetensors",
214
+ "transformer.layers.20.attn.k_norm.weight": "model.safetensors",
215
+ "transformer.layers.20.attn.out_proj.biases": "model.safetensors",
216
+ "transformer.layers.20.attn.out_proj.scales": "model.safetensors",
217
+ "transformer.layers.20.attn.out_proj.weight": "model.safetensors",
218
+ "transformer.layers.20.attn.q_norm.weight": "model.safetensors",
219
+ "transformer.layers.20.attn.qkv_proj.biases": "model.safetensors",
220
+ "transformer.layers.20.attn.qkv_proj.scales": "model.safetensors",
221
+ "transformer.layers.20.attn.qkv_proj.weight": "model.safetensors",
222
+ "transformer.layers.20.attn_norm.weight": "model.safetensors",
223
+ "transformer.layers.20.ffn.proj_1.biases": "model.safetensors",
224
+ "transformer.layers.20.ffn.proj_1.scales": "model.safetensors",
225
+ "transformer.layers.20.ffn.proj_1.weight": "model.safetensors",
226
+ "transformer.layers.20.ffn.proj_2.biases": "model.safetensors",
227
+ "transformer.layers.20.ffn.proj_2.scales": "model.safetensors",
228
+ "transformer.layers.20.ffn.proj_2.weight": "model.safetensors",
229
+ "transformer.layers.20.ffn_norm.weight": "model.safetensors",
230
+ "transformer.layers.21.attn.k_norm.weight": "model.safetensors",
231
+ "transformer.layers.21.attn.out_proj.biases": "model.safetensors",
232
+ "transformer.layers.21.attn.out_proj.scales": "model.safetensors",
233
+ "transformer.layers.21.attn.out_proj.weight": "model.safetensors",
234
+ "transformer.layers.21.attn.q_norm.weight": "model.safetensors",
235
+ "transformer.layers.21.attn.qkv_proj.biases": "model.safetensors",
236
+ "transformer.layers.21.attn.qkv_proj.scales": "model.safetensors",
237
+ "transformer.layers.21.attn.qkv_proj.weight": "model.safetensors",
238
+ "transformer.layers.21.attn_norm.weight": "model.safetensors",
239
+ "transformer.layers.21.ffn.proj_1.biases": "model.safetensors",
240
+ "transformer.layers.21.ffn.proj_1.scales": "model.safetensors",
241
+ "transformer.layers.21.ffn.proj_1.weight": "model.safetensors",
242
+ "transformer.layers.21.ffn.proj_2.biases": "model.safetensors",
243
+ "transformer.layers.21.ffn.proj_2.scales": "model.safetensors",
244
+ "transformer.layers.21.ffn.proj_2.weight": "model.safetensors",
245
+ "transformer.layers.21.ffn_norm.weight": "model.safetensors",
246
+ "transformer.layers.22.attn.k_norm.weight": "model.safetensors",
247
+ "transformer.layers.22.attn.out_proj.biases": "model.safetensors",
248
+ "transformer.layers.22.attn.out_proj.scales": "model.safetensors",
249
+ "transformer.layers.22.attn.out_proj.weight": "model.safetensors",
250
+ "transformer.layers.22.attn.q_norm.weight": "model.safetensors",
251
+ "transformer.layers.22.attn.qkv_proj.biases": "model.safetensors",
252
+ "transformer.layers.22.attn.qkv_proj.scales": "model.safetensors",
253
+ "transformer.layers.22.attn.qkv_proj.weight": "model.safetensors",
254
+ "transformer.layers.22.attn_norm.weight": "model.safetensors",
255
+ "transformer.layers.22.ffn.proj_1.biases": "model.safetensors",
256
+ "transformer.layers.22.ffn.proj_1.scales": "model.safetensors",
257
+ "transformer.layers.22.ffn.proj_1.weight": "model.safetensors",
258
+ "transformer.layers.22.ffn.proj_2.biases": "model.safetensors",
259
+ "transformer.layers.22.ffn.proj_2.scales": "model.safetensors",
260
+ "transformer.layers.22.ffn.proj_2.weight": "model.safetensors",
261
+ "transformer.layers.22.ffn_norm.weight": "model.safetensors",
262
+ "transformer.layers.23.attn.k_norm.weight": "model.safetensors",
263
+ "transformer.layers.23.attn.out_proj.biases": "model.safetensors",
264
+ "transformer.layers.23.attn.out_proj.scales": "model.safetensors",
265
+ "transformer.layers.23.attn.out_proj.weight": "model.safetensors",
266
+ "transformer.layers.23.attn.q_norm.weight": "model.safetensors",
267
+ "transformer.layers.23.attn.qkv_proj.biases": "model.safetensors",
268
+ "transformer.layers.23.attn.qkv_proj.scales": "model.safetensors",
269
+ "transformer.layers.23.attn.qkv_proj.weight": "model.safetensors",
270
+ "transformer.layers.23.attn_norm.weight": "model.safetensors",
271
+ "transformer.layers.23.ffn.proj_1.biases": "model.safetensors",
272
+ "transformer.layers.23.ffn.proj_1.scales": "model.safetensors",
273
+ "transformer.layers.23.ffn.proj_1.weight": "model.safetensors",
274
+ "transformer.layers.23.ffn.proj_2.biases": "model.safetensors",
275
+ "transformer.layers.23.ffn.proj_2.scales": "model.safetensors",
276
+ "transformer.layers.23.ffn.proj_2.weight": "model.safetensors",
277
+ "transformer.layers.23.ffn_norm.weight": "model.safetensors",
278
+ "transformer.layers.24.attn.k_norm.weight": "model.safetensors",
279
+ "transformer.layers.24.attn.out_proj.biases": "model.safetensors",
280
+ "transformer.layers.24.attn.out_proj.scales": "model.safetensors",
281
+ "transformer.layers.24.attn.out_proj.weight": "model.safetensors",
282
+ "transformer.layers.24.attn.q_norm.weight": "model.safetensors",
283
+ "transformer.layers.24.attn.qkv_proj.biases": "model.safetensors",
284
+ "transformer.layers.24.attn.qkv_proj.scales": "model.safetensors",
285
+ "transformer.layers.24.attn.qkv_proj.weight": "model.safetensors",
286
+ "transformer.layers.24.attn_norm.weight": "model.safetensors",
287
+ "transformer.layers.24.ffn.proj_1.biases": "model.safetensors",
288
+ "transformer.layers.24.ffn.proj_1.scales": "model.safetensors",
289
+ "transformer.layers.24.ffn.proj_1.weight": "model.safetensors",
290
+ "transformer.layers.24.ffn.proj_2.biases": "model.safetensors",
291
+ "transformer.layers.24.ffn.proj_2.scales": "model.safetensors",
292
+ "transformer.layers.24.ffn.proj_2.weight": "model.safetensors",
293
+ "transformer.layers.24.ffn_norm.weight": "model.safetensors",
294
+ "transformer.layers.25.attn.k_norm.weight": "model.safetensors",
295
+ "transformer.layers.25.attn.out_proj.biases": "model.safetensors",
296
+ "transformer.layers.25.attn.out_proj.scales": "model.safetensors",
297
+ "transformer.layers.25.attn.out_proj.weight": "model.safetensors",
298
+ "transformer.layers.25.attn.q_norm.weight": "model.safetensors",
299
+ "transformer.layers.25.attn.qkv_proj.biases": "model.safetensors",
300
+ "transformer.layers.25.attn.qkv_proj.scales": "model.safetensors",
301
+ "transformer.layers.25.attn.qkv_proj.weight": "model.safetensors",
302
+ "transformer.layers.25.attn_norm.weight": "model.safetensors",
303
+ "transformer.layers.25.ffn.proj_1.biases": "model.safetensors",
304
+ "transformer.layers.25.ffn.proj_1.scales": "model.safetensors",
305
+ "transformer.layers.25.ffn.proj_1.weight": "model.safetensors",
306
+ "transformer.layers.25.ffn.proj_2.biases": "model.safetensors",
307
+ "transformer.layers.25.ffn.proj_2.scales": "model.safetensors",
308
+ "transformer.layers.25.ffn.proj_2.weight": "model.safetensors",
309
+ "transformer.layers.25.ffn_norm.weight": "model.safetensors",
310
+ "transformer.layers.26.attn.k_norm.weight": "model.safetensors",
311
+ "transformer.layers.26.attn.out_proj.biases": "model.safetensors",
312
+ "transformer.layers.26.attn.out_proj.scales": "model.safetensors",
313
+ "transformer.layers.26.attn.out_proj.weight": "model.safetensors",
314
+ "transformer.layers.26.attn.q_norm.weight": "model.safetensors",
315
+ "transformer.layers.26.attn.qkv_proj.biases": "model.safetensors",
316
+ "transformer.layers.26.attn.qkv_proj.scales": "model.safetensors",
317
+ "transformer.layers.26.attn.qkv_proj.weight": "model.safetensors",
318
+ "transformer.layers.26.attn_norm.weight": "model.safetensors",
319
+ "transformer.layers.26.ffn.proj_1.biases": "model.safetensors",
320
+ "transformer.layers.26.ffn.proj_1.scales": "model.safetensors",
321
+ "transformer.layers.26.ffn.proj_1.weight": "model.safetensors",
322
+ "transformer.layers.26.ffn.proj_2.biases": "model.safetensors",
323
+ "transformer.layers.26.ffn.proj_2.scales": "model.safetensors",
324
+ "transformer.layers.26.ffn.proj_2.weight": "model.safetensors",
325
+ "transformer.layers.26.ffn_norm.weight": "model.safetensors",
326
+ "transformer.layers.27.attn.k_norm.weight": "model.safetensors",
327
+ "transformer.layers.27.attn.out_proj.biases": "model.safetensors",
328
+ "transformer.layers.27.attn.out_proj.scales": "model.safetensors",
329
+ "transformer.layers.27.attn.out_proj.weight": "model.safetensors",
330
+ "transformer.layers.27.attn.q_norm.weight": "model.safetensors",
331
+ "transformer.layers.27.attn.qkv_proj.biases": "model.safetensors",
332
+ "transformer.layers.27.attn.qkv_proj.scales": "model.safetensors",
333
+ "transformer.layers.27.attn.qkv_proj.weight": "model.safetensors",
334
+ "transformer.layers.27.attn_norm.weight": "model.safetensors",
335
+ "transformer.layers.27.ffn.proj_1.biases": "model.safetensors",
336
+ "transformer.layers.27.ffn.proj_1.scales": "model.safetensors",
337
+ "transformer.layers.27.ffn.proj_1.weight": "model.safetensors",
338
+ "transformer.layers.27.ffn.proj_2.biases": "model.safetensors",
339
+ "transformer.layers.27.ffn.proj_2.scales": "model.safetensors",
340
+ "transformer.layers.27.ffn.proj_2.weight": "model.safetensors",
341
+ "transformer.layers.27.ffn_norm.weight": "model.safetensors",
342
+ "transformer.layers.3.attn.k_norm.weight": "model.safetensors",
343
+ "transformer.layers.3.attn.out_proj.biases": "model.safetensors",
344
+ "transformer.layers.3.attn.out_proj.scales": "model.safetensors",
345
+ "transformer.layers.3.attn.out_proj.weight": "model.safetensors",
346
+ "transformer.layers.3.attn.q_norm.weight": "model.safetensors",
347
+ "transformer.layers.3.attn.qkv_proj.biases": "model.safetensors",
348
+ "transformer.layers.3.attn.qkv_proj.scales": "model.safetensors",
349
+ "transformer.layers.3.attn.qkv_proj.weight": "model.safetensors",
350
+ "transformer.layers.3.attn_norm.weight": "model.safetensors",
351
+ "transformer.layers.3.ffn.proj_1.biases": "model.safetensors",
352
+ "transformer.layers.3.ffn.proj_1.scales": "model.safetensors",
353
+ "transformer.layers.3.ffn.proj_1.weight": "model.safetensors",
354
+ "transformer.layers.3.ffn.proj_2.biases": "model.safetensors",
355
+ "transformer.layers.3.ffn.proj_2.scales": "model.safetensors",
356
+ "transformer.layers.3.ffn.proj_2.weight": "model.safetensors",
357
+ "transformer.layers.3.ffn_norm.weight": "model.safetensors",
358
+ "transformer.layers.4.attn.k_norm.weight": "model.safetensors",
359
+ "transformer.layers.4.attn.out_proj.biases": "model.safetensors",
360
+ "transformer.layers.4.attn.out_proj.scales": "model.safetensors",
361
+ "transformer.layers.4.attn.out_proj.weight": "model.safetensors",
362
+ "transformer.layers.4.attn.q_norm.weight": "model.safetensors",
363
+ "transformer.layers.4.attn.qkv_proj.biases": "model.safetensors",
364
+ "transformer.layers.4.attn.qkv_proj.scales": "model.safetensors",
365
+ "transformer.layers.4.attn.qkv_proj.weight": "model.safetensors",
366
+ "transformer.layers.4.attn_norm.weight": "model.safetensors",
367
+ "transformer.layers.4.ffn.proj_1.biases": "model.safetensors",
368
+ "transformer.layers.4.ffn.proj_1.scales": "model.safetensors",
369
+ "transformer.layers.4.ffn.proj_1.weight": "model.safetensors",
370
+ "transformer.layers.4.ffn.proj_2.biases": "model.safetensors",
371
+ "transformer.layers.4.ffn.proj_2.scales": "model.safetensors",
372
+ "transformer.layers.4.ffn.proj_2.weight": "model.safetensors",
373
+ "transformer.layers.4.ffn_norm.weight": "model.safetensors",
374
+ "transformer.layers.5.attn.k_norm.weight": "model.safetensors",
375
+ "transformer.layers.5.attn.out_proj.biases": "model.safetensors",
376
+ "transformer.layers.5.attn.out_proj.scales": "model.safetensors",
377
+ "transformer.layers.5.attn.out_proj.weight": "model.safetensors",
378
+ "transformer.layers.5.attn.q_norm.weight": "model.safetensors",
379
+ "transformer.layers.5.attn.qkv_proj.biases": "model.safetensors",
380
+ "transformer.layers.5.attn.qkv_proj.scales": "model.safetensors",
381
+ "transformer.layers.5.attn.qkv_proj.weight": "model.safetensors",
382
+ "transformer.layers.5.attn_norm.weight": "model.safetensors",
383
+ "transformer.layers.5.ffn.proj_1.biases": "model.safetensors",
384
+ "transformer.layers.5.ffn.proj_1.scales": "model.safetensors",
385
+ "transformer.layers.5.ffn.proj_1.weight": "model.safetensors",
386
+ "transformer.layers.5.ffn.proj_2.biases": "model.safetensors",
387
+ "transformer.layers.5.ffn.proj_2.scales": "model.safetensors",
388
+ "transformer.layers.5.ffn.proj_2.weight": "model.safetensors",
389
+ "transformer.layers.5.ffn_norm.weight": "model.safetensors",
390
+ "transformer.layers.6.attn.k_norm.weight": "model.safetensors",
391
+ "transformer.layers.6.attn.out_proj.biases": "model.safetensors",
392
+ "transformer.layers.6.attn.out_proj.scales": "model.safetensors",
393
+ "transformer.layers.6.attn.out_proj.weight": "model.safetensors",
394
+ "transformer.layers.6.attn.q_norm.weight": "model.safetensors",
395
+ "transformer.layers.6.attn.qkv_proj.biases": "model.safetensors",
396
+ "transformer.layers.6.attn.qkv_proj.scales": "model.safetensors",
397
+ "transformer.layers.6.attn.qkv_proj.weight": "model.safetensors",
398
+ "transformer.layers.6.attn_norm.weight": "model.safetensors",
399
+ "transformer.layers.6.ffn.proj_1.biases": "model.safetensors",
400
+ "transformer.layers.6.ffn.proj_1.scales": "model.safetensors",
401
+ "transformer.layers.6.ffn.proj_1.weight": "model.safetensors",
402
+ "transformer.layers.6.ffn.proj_2.biases": "model.safetensors",
403
+ "transformer.layers.6.ffn.proj_2.scales": "model.safetensors",
404
+ "transformer.layers.6.ffn.proj_2.weight": "model.safetensors",
405
+ "transformer.layers.6.ffn_norm.weight": "model.safetensors",
406
+ "transformer.layers.7.attn.k_norm.weight": "model.safetensors",
407
+ "transformer.layers.7.attn.out_proj.biases": "model.safetensors",
408
+ "transformer.layers.7.attn.out_proj.scales": "model.safetensors",
409
+ "transformer.layers.7.attn.out_proj.weight": "model.safetensors",
410
+ "transformer.layers.7.attn.q_norm.weight": "model.safetensors",
411
+ "transformer.layers.7.attn.qkv_proj.biases": "model.safetensors",
412
+ "transformer.layers.7.attn.qkv_proj.scales": "model.safetensors",
413
+ "transformer.layers.7.attn.qkv_proj.weight": "model.safetensors",
414
+ "transformer.layers.7.attn_norm.weight": "model.safetensors",
415
+ "transformer.layers.7.ffn.proj_1.biases": "model.safetensors",
416
+ "transformer.layers.7.ffn.proj_1.scales": "model.safetensors",
417
+ "transformer.layers.7.ffn.proj_1.weight": "model.safetensors",
418
+ "transformer.layers.7.ffn.proj_2.biases": "model.safetensors",
419
+ "transformer.layers.7.ffn.proj_2.scales": "model.safetensors",
420
+ "transformer.layers.7.ffn.proj_2.weight": "model.safetensors",
421
+ "transformer.layers.7.ffn_norm.weight": "model.safetensors",
422
+ "transformer.layers.8.attn.k_norm.weight": "model.safetensors",
423
+ "transformer.layers.8.attn.out_proj.biases": "model.safetensors",
424
+ "transformer.layers.8.attn.out_proj.scales": "model.safetensors",
425
+ "transformer.layers.8.attn.out_proj.weight": "model.safetensors",
426
+ "transformer.layers.8.attn.q_norm.weight": "model.safetensors",
427
+ "transformer.layers.8.attn.qkv_proj.biases": "model.safetensors",
428
+ "transformer.layers.8.attn.qkv_proj.scales": "model.safetensors",
429
+ "transformer.layers.8.attn.qkv_proj.weight": "model.safetensors",
430
+ "transformer.layers.8.attn_norm.weight": "model.safetensors",
431
+ "transformer.layers.8.ffn.proj_1.biases": "model.safetensors",
432
+ "transformer.layers.8.ffn.proj_1.scales": "model.safetensors",
433
+ "transformer.layers.8.ffn.proj_1.weight": "model.safetensors",
434
+ "transformer.layers.8.ffn.proj_2.biases": "model.safetensors",
435
+ "transformer.layers.8.ffn.proj_2.scales": "model.safetensors",
436
+ "transformer.layers.8.ffn.proj_2.weight": "model.safetensors",
437
+ "transformer.layers.8.ffn_norm.weight": "model.safetensors",
438
+ "transformer.layers.9.attn.k_norm.weight": "model.safetensors",
439
+ "transformer.layers.9.attn.out_proj.biases": "model.safetensors",
440
+ "transformer.layers.9.attn.out_proj.scales": "model.safetensors",
441
+ "transformer.layers.9.attn.out_proj.weight": "model.safetensors",
442
+ "transformer.layers.9.attn.q_norm.weight": "model.safetensors",
443
+ "transformer.layers.9.attn.qkv_proj.biases": "model.safetensors",
444
+ "transformer.layers.9.attn.qkv_proj.scales": "model.safetensors",
445
+ "transformer.layers.9.attn.qkv_proj.weight": "model.safetensors",
446
+ "transformer.layers.9.attn_norm.weight": "model.safetensors",
447
+ "transformer.layers.9.ffn.proj_1.biases": "model.safetensors",
448
+ "transformer.layers.9.ffn.proj_1.scales": "model.safetensors",
449
+ "transformer.layers.9.ffn.proj_1.weight": "model.safetensors",
450
+ "transformer.layers.9.ffn.proj_2.biases": "model.safetensors",
451
+ "transformer.layers.9.ffn.proj_2.scales": "model.safetensors",
452
+ "transformer.layers.9.ffn.proj_2.weight": "model.safetensors",
453
+ "transformer.layers.9.ffn_norm.weight": "model.safetensors",
454
+ "transformer.norm.weight": "model.safetensors",
455
+ "transformer.token_embeddings.biases": "model.safetensors",
456
+ "transformer.token_embeddings.scales": "model.safetensors",
457
+ "transformer.token_embeddings.weight": "model.safetensors"
458
+ }
459
+ }
modeling_openelm.py ADDED
@@ -0,0 +1,1008 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #
2
+ # For licensing see accompanying LICENSE file.
3
+ # Copyright (C) 2024 Apple Inc. All Rights Reserved.
4
+ #
5
+
6
+ from typing import List, Optional, Tuple, Union
7
+
8
+ import torch
9
+ import torch.utils.checkpoint
10
+ from torch import Tensor, nn
11
+ from torch.nn import CrossEntropyLoss
12
+ from torch.nn import functional as F
13
+ from transformers import PreTrainedModel
14
+ from transformers.activations import ACT2FN
15
+ from transformers.cache_utils import Cache, DynamicCache, StaticCache
16
+ from transformers.modeling_outputs import (
17
+ BaseModelOutputWithPast,
18
+ CausalLMOutputWithPast,
19
+ )
20
+ from transformers.utils import logging
21
+
22
+ logger = logging.get_logger(__name__)
23
+
24
+ # this import has to be relative, otherwise, when setting trust_remote_code=True
25
+ # huggingface transformers won't be able to load the module correctly
26
+ from .configuration_openelm import OpenELMConfig, make_divisible
27
+
28
+
29
+ class OpenELMRMSNorm(nn.Module):
30
+ def __init__(self, num_features: int, eps: float = 1e-6):
31
+ """
32
+ Initialize the OpenELMRMSNorm normalization layer.
33
+
34
+ Args:
35
+ dim (int): The dimension of the input tensor.
36
+ eps (float, optional): A small value added to the denominator for numerical stability. Default is 1e-6.
37
+
38
+ Attributes:
39
+ eps (float): A small value added to the denominator for numerical stability.
40
+ weight (nn.Parameter): Learnable scaling parameter.
41
+
42
+ """
43
+ super().__init__()
44
+ self.eps = eps
45
+ self.weight = nn.Parameter(torch.ones(num_features))
46
+ self.num_features = num_features
47
+
48
+ def _norm(self, x: Tensor) -> Tensor:
49
+ """
50
+ Apply the OpenELMRMSNorm normalization to the input tensor.
51
+
52
+ Args:
53
+ x (torch.Tensor): The input tensor.
54
+
55
+ Returns:
56
+ torch.Tensor: The normalized tensor.
57
+
58
+ """
59
+ return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
60
+
61
+ def forward(self, x: Tensor) -> Tensor:
62
+ """
63
+ Forward pass through the OpenELMRMSNorm layer.
64
+
65
+ Args:
66
+ x (torch.Tensor): The input tensor.
67
+
68
+ Returns:
69
+ torch.Tensor: The output tensor after applying OpenELMRMSNorm.
70
+
71
+ """
72
+ output = self._norm(x.float()).type_as(x)
73
+ return output * self.weight
74
+
75
+ def extra_repr(self) -> str:
76
+ return (
77
+ super().extra_repr() + f"num_features={self.num_features}, eps={self.eps}"
78
+ )
79
+
80
+
81
+ class OpenELMPreTrainedModel(PreTrainedModel):
82
+ config_class = OpenELMConfig
83
+ base_model_prefix = "transformer"
84
+ supports_gradient_checkpointing = True
85
+ _no_split_modules = ["OpenELMDecoderLayer"]
86
+ _skip_keys_device_placement = "past_key_values"
87
+
88
+ def __init__(self, *inputs, **kwargs) -> None:
89
+ super().__init__(*inputs, **kwargs)
90
+
91
+ def _init_weights(self, module: nn.Module) -> None:
92
+ """Initialize the weights."""
93
+ if isinstance(module, nn.Linear):
94
+ # Slightly different from the TF version which uses truncated_normal for initialization
95
+ # cf https://github.com/pytorch/pytorch/pull/5617
96
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
97
+ if module.bias is not None:
98
+ module.bias.data.zero_()
99
+ elif isinstance(module, nn.Embedding):
100
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
101
+ if module.padding_idx is not None:
102
+ module.weight.data[module.padding_idx].zero_()
103
+ elif isinstance(module, OpenELMRMSNorm):
104
+ module.weight.data.fill_(1.0)
105
+
106
+
107
+ def _rotate_half(x: Tensor) -> Tensor:
108
+ x1, x2 = x.chunk(2, dim=-1)
109
+ return torch.cat((-x2, x1), dim=-1)
110
+
111
+
112
+ def _apply_rotary_pos_emb(x: Tensor, pos_sin: Tensor, pos_cos: Tensor) -> Tensor:
113
+ return (x * pos_cos) + (_rotate_half(x) * pos_sin)
114
+
115
+
116
+ class OpenELMRotaryEmbedding(torch.nn.Module):
117
+ """
118
+ The rotary position embeddings (aka RoPE) from `RoFormer <https://arxiv.org/abs/2104.09864>`_.
119
+
120
+ RoPE encodes the position information of tokens using a rotation matrix, and is able to capture
121
+ explicit relative positional dependencies.
122
+
123
+ Args:
124
+ model_dim: The dimensionality of the model's hidden state.
125
+ max_seq_length: Maximum sequence length.
126
+ freq_constant: A constant used for computing frequencies.
127
+ """
128
+
129
+ def __init__(
130
+ self, model_dim: int, max_seq_length: int, freq_constant: int = 10000
131
+ ) -> None:
132
+ inv_freq = 1.0 / (
133
+ freq_constant
134
+ ** (torch.arange(0, model_dim, 2, dtype=torch.float32) / model_dim)
135
+ )
136
+ super().__init__()
137
+
138
+ self.model_dim = model_dim
139
+ self.freq_constant = freq_constant
140
+ self.max_seq_length = max_seq_length
141
+
142
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
143
+ self._cached_cos = None
144
+ self._cached_sin = None
145
+ self._cached_seq_length = max_seq_length
146
+ self._compute_sin_cos_embeddings(max_seq_length)
147
+
148
+ def extra_repr(self) -> str:
149
+ return f"\tmodel_dim={self.model_dim}, max_seq_length={self.max_seq_length}, freq_constant={self.freq_constant}"
150
+
151
+ def _compute_sin_cos_embeddings(
152
+ self,
153
+ key_len: int,
154
+ key_device: torch.device = torch.device("cpu"),
155
+ key_dtype: torch.dtype = torch.float32,
156
+ ) -> None:
157
+ """
158
+ Compute sine and cos embeddings.
159
+
160
+ Args:
161
+ key_len: Number of tokens in the key embeddings in the transformer model.
162
+ device: Device where the key embeddings are stored.
163
+ key_dtype: Data type of the key embeddings.
164
+
165
+ Returns:
166
+ None
167
+
168
+ ...note:
169
+ We recalculate the sine and cosine embeddings if any of the following conditions are met:
170
+ 1. The number of tokens in key embeddings are greater than the cached sequence length.
171
+ 2. Sine and cosine caches are empty.
172
+ 3. The device and data type of sine and cosine embeddings does not match with the key embeddings.
173
+ """
174
+ if (
175
+ key_len > self._cached_seq_length
176
+ or self._cached_cos is None
177
+ or (self._cached_cos is not None and self._cached_cos.device != key_device)
178
+ or (self._cached_cos is not None and self._cached_cos.dtype != key_dtype)
179
+ or self._cached_sin is None
180
+ or (self._cached_sin is not None and self._cached_sin.device != key_device)
181
+ or (self._cached_sin is not None and self._cached_sin.dtype != key_dtype)
182
+ ):
183
+ self._cached_seq_length = max(key_len, self._cached_seq_length)
184
+
185
+ # The shape of 'pos_index' is [number of key tokens]
186
+ pos_index = torch.arange(
187
+ self._cached_seq_length,
188
+ dtype=torch.float32,
189
+ device=self.inv_freq.device,
190
+ )
191
+ # The shape of 'pos_index_theta' is [number of key tokens, model dimension]
192
+ pos_index_theta = torch.einsum("i,j->ij", pos_index, self.inv_freq)
193
+ # The shape of 'emb' is [number of key tokens, model dimension]
194
+ emb = torch.cat((pos_index_theta, pos_index_theta), dim=-1)
195
+
196
+ # the shape of cos and sin embeddings is [number of key tokens, model_dim]
197
+ cos_emb = emb.cos().to(dtype=key_dtype, device=key_device)
198
+ sin_emb = emb.sin().to(dtype=key_dtype, device=key_device)
199
+
200
+ # the shape of cached cos and sin embeddings is [1, 1, number of key tokens, model_dim]
201
+ self._cached_cos = cos_emb[None, None, :, :]
202
+ self._cached_sin = sin_emb[None, None, :, :]
203
+
204
+ def forward(
205
+ self,
206
+ query: torch.Tensor,
207
+ key: torch.Tensor,
208
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
209
+ """
210
+ The forward function of RoPE embeddings.
211
+
212
+ Args:
213
+ query: Query embeddings in the transformer model. The shape of query embeddings is
214
+ [Batch, number of query heads, number of query tokens, model dimension].
215
+ key: Key embeddings in the transformer model. The shape of key embeddings is
216
+ [Batch, number of key heads, number of key tokens, model dimension].
217
+
218
+ Returns:
219
+ A tuple containing the query and key embeddings with positional information. The shape of the returned query
220
+ and key embeddings is the same as the input query and key embeddings respectively.
221
+
222
+ ...note:
223
+ The RoPE embedding computation is done in full-precision. After the computation, input query and key tensors
224
+ are casted to original input datatype.
225
+ """
226
+ dim = key.shape[-1]
227
+ key_len = key.shape[2]
228
+ query_len = query.shape[2]
229
+
230
+ assert dim == self.model_dim
231
+ assert key.device == query.device
232
+ assert key.dtype == query.dtype
233
+
234
+ # In the context of self-attention, the lengths of keys and queries are equal.
235
+ # However, in generation tasks, such as predicting the next token in a sequence, the lengths of keys and queries
236
+ # can differ. For instance, when employing key-value (KV) caching for sequence prediction, the keys
237
+ # represent embeddings of previous tokens and the current token, while the query corresponds
238
+ # to the embedding of the current token only.
239
+ assert (
240
+ key_len >= query_len
241
+ ), "Number of keys has to be greater than or equal to number of queries."
242
+
243
+ query_float = query.float()
244
+ key_float = key.float()
245
+
246
+ self._compute_sin_cos_embeddings(
247
+ key_len, key_device=key_float.device, key_dtype=key_float.dtype
248
+ )
249
+ query_float = _apply_rotary_pos_emb(
250
+ x=query_float,
251
+ pos_sin=self._cached_sin[..., key_len - query_len : key_len, :],
252
+ pos_cos=self._cached_cos[..., key_len - query_len : key_len, :],
253
+ )
254
+ key_float = _apply_rotary_pos_emb(
255
+ x=key_float,
256
+ pos_sin=self._cached_sin[..., :key_len, :],
257
+ pos_cos=self._cached_cos[..., :key_len, :],
258
+ )
259
+
260
+ return query_float.type_as(query), key_float.type_as(key)
261
+
262
+
263
+ class OpenELMMultiHeadCausalAttention(nn.Module):
264
+ def __init__(self, config: OpenELMConfig, layer_idx: int) -> None:
265
+ super().__init__()
266
+ self.layer_idx = layer_idx
267
+ head_dim = config.head_dim
268
+ q_heads = config.num_query_heads[layer_idx]
269
+ k_heads = config.num_kv_heads[layer_idx]
270
+ v_heads = config.num_kv_heads[layer_idx]
271
+
272
+ self.qkv_proj = nn.Linear(
273
+ in_features=config.model_dim,
274
+ out_features=(q_heads + k_heads + v_heads) * head_dim,
275
+ bias=False,
276
+ )
277
+
278
+ self.pos_embedding = OpenELMRotaryEmbedding(
279
+ model_dim=config.head_dim,
280
+ max_seq_length=config.rope_max_length,
281
+ freq_constant=config.rope_freq_constant,
282
+ )
283
+
284
+ if config.normalize_qk_projections:
285
+ self.q_norm = OpenELMRMSNorm(
286
+ num_features=config.head_dim,
287
+ )
288
+ self.k_norm = OpenELMRMSNorm(
289
+ num_features=config.head_dim,
290
+ )
291
+ else:
292
+ self.q_norm = None
293
+ self.k_norm = None
294
+
295
+ self.out_proj = nn.Linear(
296
+ in_features=q_heads * head_dim,
297
+ out_features=config.model_dim,
298
+ bias=False,
299
+ )
300
+
301
+ self.head_dim = config.head_dim
302
+ self.num_q_heads = q_heads
303
+ self.num_k_heads = k_heads
304
+ self.num_v_heads = v_heads
305
+ self.transformer_dim = config.model_dim
306
+ self.num_groups = self.num_q_heads // self.num_k_heads
307
+
308
+ def extra_repr(self) -> str:
309
+ return (
310
+ super().extra_repr()
311
+ + f"query_heads={self.num_q_heads}, key_heads={self.num_k_heads}, value_heads={self.num_v_heads}"
312
+ )
313
+
314
+ def forward(
315
+ self,
316
+ hidden_states: torch.Tensor,
317
+ attention_mask: Optional[torch.Tensor] = None,
318
+ past_key_value: Optional[Cache] = None,
319
+ output_attentions: bool = False,
320
+ use_cache: bool = False,
321
+ cache_position: Optional[torch.LongTensor] = None,
322
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
323
+ """
324
+ Forward pass of multi-head self-attention.
325
+
326
+ Args:
327
+ hidden_states: Input tensor of the shape [batch size, sequence length, model dimension].
328
+ past_key_value: Tensor storing the cached keys and values.
329
+ output_attentions: output attention weights.
330
+ use_cache: Specifies whether to use kv-cache for generation.
331
+ cache_position: used for updating the kv-cache.
332
+
333
+ Returns:
334
+ The output of the same shape as the input, optionally with a tensor containing cached keys and values.
335
+ """
336
+
337
+ # scaled_dot_product_attention does not return attention weights, set output_attentions to False
338
+ output_attentions = False
339
+ batch_size, seq_length, d_model = hidden_states.size()
340
+
341
+ # [B, S, d] --> [B, S, (q_h + k_h + v_h) * h]
342
+ qkv = self.qkv_proj(hidden_states)
343
+ # [B, S, (q_h + k_h + v_h) * h] --> [B, S, (q_h + k_h + v_h), h]
344
+ qkv = qkv.reshape(
345
+ batch_size,
346
+ seq_length,
347
+ self.num_q_heads + self.num_k_heads + self.num_v_heads,
348
+ self.head_dim,
349
+ )
350
+ # [B, S, (q_h + k_h + v_h), h] --> [B, (q_h + k_h + v_h), S, h]
351
+ qkv = qkv.transpose(1, 2)
352
+ # [B, (q_h + k_h + v_h), S, h] --> [B, q_h, S h], [B, k_h, S, h], [B, v_h, S, h]
353
+ queries, keys, values = qkv.split(
354
+ [self.num_q_heads, self.num_k_heads, self.num_v_heads], dim=1
355
+ )
356
+
357
+ if self.q_norm is not None:
358
+ queries = self.q_norm(queries)
359
+
360
+ if self.k_norm is not None:
361
+ keys = self.k_norm(keys)
362
+
363
+ past_key_value = getattr(self, "past_key_value", past_key_value)
364
+
365
+ if past_key_value is not None:
366
+ # sin and cos are specific to RoPE models; position_ids needed for the static cache
367
+ # cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
368
+ cache_kwargs = {"cache_position": cache_position}
369
+ keys, values = past_key_value.update(
370
+ keys, values, self.layer_idx, cache_kwargs
371
+ )
372
+
373
+ # Add positional embedding
374
+ queries, keys = self.pos_embedding(queries, keys)
375
+
376
+ if self.num_groups != 1:
377
+ # GQA
378
+ # [B, k_h, S, h] --> [B, q_h, S, h]
379
+ keys = keys.repeat_interleave(self.num_groups, dim=1)
380
+ # [B, v_h, S, h] --> [B, q_h, S, h]
381
+ values = values.repeat_interleave(self.num_groups, dim=1)
382
+
383
+ causal_mask = attention_mask
384
+ if attention_mask is not None and cache_position is not None:
385
+ causal_mask = causal_mask[:, :, cache_position, : keys.shape[-2]]
386
+
387
+ attn_output = F.scaled_dot_product_attention(
388
+ queries,
389
+ keys,
390
+ values,
391
+ attn_mask=causal_mask,
392
+ dropout_p=0,
393
+ )
394
+
395
+ attn_output = attn_output.transpose(1, 2).contiguous()
396
+ attn_output = attn_output.reshape(
397
+ batch_size, seq_length, self.num_q_heads * self.head_dim
398
+ )
399
+ attn_output = self.out_proj(attn_output)
400
+ if not output_attentions:
401
+ attn_weights = None
402
+ return attn_output, attn_weights, past_key_value
403
+
404
+
405
+ class OpenELMFeedForwardNetwork(nn.Module):
406
+ def __init__(self, config: OpenELMConfig, layer_idx: int) -> None:
407
+ super().__init__()
408
+ ffn_multiplier = config.ffn_multipliers[layer_idx]
409
+ intermediate_dim = int(
410
+ make_divisible(
411
+ ffn_multiplier * config.model_dim,
412
+ divisor=config.ffn_dim_divisor,
413
+ )
414
+ )
415
+ if config.ffn_with_glu:
416
+ # FFN with Gated linear unit, as described in https://arxiv.org/abs/2002.05202v1.
417
+ self.proj_1 = nn.Linear(
418
+ in_features=config.model_dim,
419
+ out_features=2 * intermediate_dim,
420
+ bias=False,
421
+ )
422
+ self.proj_2 = nn.Linear(
423
+ in_features=intermediate_dim,
424
+ out_features=config.model_dim,
425
+ bias=False,
426
+ )
427
+ self.ffn_with_glu = True
428
+ else:
429
+ # Standard FFN, as described in https://arxiv.org/abs/1706.03762
430
+ self.proj_1 = nn.Linear(
431
+ in_features=config.model_dim,
432
+ out_features=intermediate_dim,
433
+ bias=False,
434
+ )
435
+ self.proj_2 = nn.Linear(
436
+ in_features=intermediate_dim,
437
+ out_features=config.model_dim,
438
+ bias=False,
439
+ )
440
+ self.ffn_with_glu = False
441
+
442
+ self.act = ACT2FN[config.activation_fn_name]
443
+
444
+ def extra_repr(self) -> str:
445
+ return super().extra_repr() + f"(ffn_with_glu) : {self.ffn_with_glu}"
446
+
447
+ def forward(self, x: Tensor) -> Tensor:
448
+ """Forward function of FFN layer.
449
+
450
+ Args:
451
+ x: Input tensor of the shape [batch size, sequence length, model dimension].
452
+
453
+ Returns:
454
+ A tensor of the same shape as the input.
455
+ """
456
+ if self.ffn_with_glu:
457
+ y_12 = self.proj_1(x)
458
+ y_1, y_2 = y_12.chunk(2, dim=-1)
459
+ y = self.act(y_1) * y_2
460
+ return self.proj_2(y)
461
+ else:
462
+ return self.proj_2(self.act(self.proj_1(x)))
463
+
464
+
465
+ class OpenELMDecoderLayer(nn.Module):
466
+ def __init__(self, config: OpenELMConfig, layer_idx: int) -> None:
467
+ super().__init__()
468
+ self.attn = OpenELMMultiHeadCausalAttention(config=config, layer_idx=layer_idx)
469
+ self.ffn = OpenELMFeedForwardNetwork(config=config, layer_idx=layer_idx)
470
+ self.ffn_norm = OpenELMRMSNorm(
471
+ num_features=config.model_dim,
472
+ )
473
+ self.attn_norm = OpenELMRMSNorm(
474
+ num_features=config.model_dim,
475
+ )
476
+
477
+ def forward(
478
+ self,
479
+ hidden_states: torch.Tensor,
480
+ attention_mask: Optional[torch.Tensor] = None,
481
+ position_ids: Optional[torch.LongTensor] = None,
482
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
483
+ output_attentions: Optional[bool] = False,
484
+ use_cache: Optional[bool] = False,
485
+ cache_position: Optional[torch.LongTensor] = None,
486
+ **kwargs,
487
+ ) -> Tuple[
488
+ torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
489
+ ]:
490
+ """
491
+ Args:
492
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
493
+ attention_mask (`torch.FloatTensor`, *optional*):
494
+ attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
495
+ query_sequence_length, key_sequence_length)` if default attention is used.
496
+ output_attentions (`bool`, *optional*):
497
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
498
+ returned tensors for more detail.
499
+ use_cache (`bool`, *optional*):
500
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
501
+ (see `past_key_values`).
502
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
503
+ """
504
+ residual = hidden_states
505
+ hidden_states = self.attn_norm(hidden_states)
506
+
507
+ # Self Attention
508
+ hidden_states, self_attn_weights, present_key_value = self.attn(
509
+ hidden_states=hidden_states,
510
+ attention_mask=attention_mask,
511
+ past_key_value=past_key_value,
512
+ output_attentions=output_attentions,
513
+ use_cache=use_cache,
514
+ cache_position=cache_position,
515
+ **kwargs,
516
+ )
517
+ hidden_states = residual + hidden_states
518
+
519
+ # Fully Connected
520
+ residual = hidden_states
521
+ hidden_states = self.ffn_norm(hidden_states)
522
+ hidden_states = self.ffn(hidden_states)
523
+ hidden_states = residual + hidden_states
524
+
525
+ outputs = (hidden_states,)
526
+
527
+ if output_attentions:
528
+ outputs += (self_attn_weights,)
529
+
530
+ if use_cache:
531
+ outputs += (present_key_value,)
532
+
533
+ return outputs
534
+
535
+
536
+ class OpenELMModel(OpenELMPreTrainedModel):
537
+ config_class = OpenELMConfig
538
+
539
+ def __init__(self, config: OpenELMConfig):
540
+ super().__init__(config)
541
+ self.config = config
542
+
543
+ self.token_embeddings = nn.Embedding(
544
+ embedding_dim=config.model_dim,
545
+ num_embeddings=config.vocab_size,
546
+ )
547
+
548
+ self.layers = nn.ModuleList(
549
+ OpenELMDecoderLayer(config=config, layer_idx=layer_idx)
550
+ for layer_idx in range(config.num_transformer_layers)
551
+ )
552
+ self.norm = OpenELMRMSNorm(num_features=config.model_dim)
553
+ if config.share_input_output_layers:
554
+ self.classifier = None
555
+ else:
556
+ self.classifier = nn.Linear(
557
+ in_features=config.model_dim,
558
+ out_features=config.vocab_size,
559
+ bias=False,
560
+ )
561
+ self.num_transformer_layers = config.num_transformer_layers
562
+ self.gradient_checkpointing = False
563
+
564
+ # Register a causal mask to separate causal and padding mask creation. Merging happens in the attention class.
565
+ # NOTE: This is not friendly with TorchScript, ONNX, ExportedProgram serialization for very large `max_context_length`.
566
+ causal_mask = torch.full(
567
+ (config.max_context_length, config.max_context_length),
568
+ fill_value=True,
569
+ dtype=torch.bool,
570
+ )
571
+ self.register_buffer(
572
+ "causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False
573
+ )
574
+
575
+ # Initialize weights and apply final processing
576
+ self.post_init()
577
+ self.reset_parameters(config=config)
578
+
579
+ def get_input_embeddings(self):
580
+ return self.token_embeddings
581
+
582
+ def set_input_embeddings(self, new_embeddings: torch.Tensor):
583
+ self.token_embeddings = new_embeddings
584
+
585
+ def reset_parameters(self, config: OpenELMConfig) -> None:
586
+ """Initialize the layers in Language Model
587
+
588
+ The initialization scheme is followed, following `OPT <https://arxiv.org/pdf/2205.01068.pdf>`_.
589
+
590
+ Args:
591
+ use_megatron_std: Use standard deviation as described in Megatron-LM.
592
+
593
+ Returns:
594
+ None
595
+ """
596
+ for module in self.modules():
597
+ if isinstance(module, nn.Linear):
598
+ std = module.in_features**-0.5
599
+ torch.nn.init.normal_(module.weight, mean=0.0, std=std)
600
+ if module.bias is not None:
601
+ torch.nn.init.zeros_(module.bias)
602
+ elif isinstance(module, nn.Embedding):
603
+ std = module.embedding_dim**-0.5
604
+ torch.nn.init.normal_(module.weight, mean=0.0, std=std)
605
+ elif isinstance(module, OpenELMRMSNorm):
606
+ if module.weight is not None:
607
+ torch.nn.init.ones_(module.weight)
608
+ if hasattr(module, "bias") and module.bias is not None:
609
+ torch.nn.init.zeros_(module.bias)
610
+
611
+ model_dim = config.model_dim
612
+ n_layers = config.num_transformer_layers
613
+ std = (model_dim**-0.5) * ((2 * n_layers) ** -0.5)
614
+ for param_name, param in self.named_parameters():
615
+ if param_name.endswith("out_proj.weight") or param_name.endswith(
616
+ "ffn.proj_2.weight"
617
+ ):
618
+ torch.nn.init.normal_(param, mean=0.0, std=std)
619
+
620
+ def forward(
621
+ self,
622
+ input_ids: torch.LongTensor = None,
623
+ attention_mask: Optional[torch.Tensor] = None,
624
+ position_ids: Optional[torch.LongTensor] = None,
625
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
626
+ inputs_embeds: Optional[torch.FloatTensor] = None,
627
+ use_cache: Optional[bool] = None,
628
+ output_attentions: Optional[bool] = None,
629
+ output_hidden_states: Optional[bool] = None,
630
+ return_dict: Optional[bool] = None,
631
+ cache_position: Optional[torch.LongTensor] = None,
632
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
633
+ output_attentions = (
634
+ output_attentions
635
+ if output_attentions is not None
636
+ else self.config.output_attentions
637
+ )
638
+ output_hidden_states = (
639
+ output_hidden_states
640
+ if output_hidden_states is not None
641
+ else self.config.output_hidden_states
642
+ )
643
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
644
+ return_dict = (
645
+ return_dict if return_dict is not None else self.config.use_return_dict
646
+ )
647
+
648
+ if (input_ids is None) ^ (inputs_embeds is not None):
649
+ raise ValueError(
650
+ "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
651
+ )
652
+
653
+ if self.gradient_checkpointing and self.training and use_cache:
654
+ logger.warning_once(
655
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
656
+ )
657
+ use_cache = False
658
+
659
+ if inputs_embeds is None:
660
+ inputs_embeds = self.token_embeddings(input_ids)
661
+
662
+ past_seen_tokens = 0
663
+ if use_cache: # kept for BC (cache positions)
664
+ if not isinstance(past_key_values, StaticCache):
665
+ past_key_values = DynamicCache.from_legacy_cache(past_key_values)
666
+ past_seen_tokens = past_key_values.get_seq_length()
667
+
668
+ if cache_position is None:
669
+ cache_position = torch.arange(
670
+ past_seen_tokens,
671
+ past_seen_tokens + inputs_embeds.shape[1],
672
+ device=inputs_embeds.device,
673
+ )
674
+
675
+ if position_ids is None:
676
+ position_ids = cache_position.unsqueeze(0)
677
+
678
+ causal_mask = self._update_causal_mask(attention_mask, inputs_embeds)
679
+
680
+ # embed positions
681
+ hidden_states = inputs_embeds
682
+
683
+ # decoder layers
684
+ all_hidden_states = () if output_hidden_states else None
685
+ all_self_attns = () if output_attentions else None
686
+ next_decoder_cache = None
687
+
688
+ for decoder_layer in self.layers:
689
+ if output_hidden_states:
690
+ all_hidden_states += (hidden_states,)
691
+
692
+ if self.gradient_checkpointing and self.training:
693
+ layer_outputs = self._gradient_checkpointing_func(
694
+ decoder_layer.__call__,
695
+ hidden_states,
696
+ causal_mask,
697
+ position_ids,
698
+ past_key_values,
699
+ output_attentions,
700
+ use_cache,
701
+ cache_position,
702
+ )
703
+ else:
704
+ layer_outputs = decoder_layer(
705
+ hidden_states,
706
+ attention_mask=causal_mask,
707
+ position_ids=position_ids,
708
+ past_key_value=past_key_values,
709
+ output_attentions=output_attentions,
710
+ use_cache=use_cache,
711
+ cache_position=cache_position,
712
+ )
713
+
714
+ hidden_states = layer_outputs[0]
715
+
716
+ if use_cache:
717
+ next_decoder_cache = layer_outputs[2 if output_attentions else 1]
718
+
719
+ if output_attentions:
720
+ all_self_attns += (layer_outputs[1],)
721
+
722
+ hidden_states = self.norm(hidden_states)
723
+
724
+ # add hidden states from the last decoder layer
725
+ if output_hidden_states:
726
+ all_hidden_states += (hidden_states,)
727
+
728
+ next_cache = None
729
+ if use_cache:
730
+ next_cache = (
731
+ next_decoder_cache.to_legacy_cache()
732
+ if isinstance(next_decoder_cache, Cache)
733
+ else next_decoder_cache
734
+ )
735
+ if not return_dict:
736
+ return tuple(
737
+ v
738
+ for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
739
+ if v is not None
740
+ )
741
+ return BaseModelOutputWithPast(
742
+ last_hidden_state=hidden_states,
743
+ past_key_values=next_cache,
744
+ hidden_states=all_hidden_states,
745
+ attentions=all_self_attns,
746
+ )
747
+
748
+ def _update_causal_mask(self, attention_mask, input_tensor):
749
+ if self.config._attn_implementation == "flash_attention_2":
750
+ if attention_mask is not None and 0.0 in attention_mask:
751
+ return attention_mask
752
+ return None
753
+
754
+ batch_size, seq_length = input_tensor.shape[:2]
755
+ dtype = input_tensor.dtype
756
+ device = input_tensor.device
757
+
758
+ # support going beyond cached `max_position_embedding`
759
+ if seq_length > self.causal_mask.shape[-1]:
760
+ causal_mask = torch.full(
761
+ (2 * self.causal_mask.shape[-1], 2 * self.causal_mask.shape[-1]),
762
+ fill_value=1,
763
+ )
764
+ self.register_buffer(
765
+ "causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False
766
+ )
767
+
768
+ # We use the current dtype to avoid any overflows
769
+ min_dtype = torch.finfo(dtype).min
770
+ causal_mask = (
771
+ self.causal_mask[None, None, :, :].repeat(batch_size, 1, 1, 1).to(dtype)
772
+ * min_dtype
773
+ )
774
+
775
+ causal_mask = causal_mask.to(dtype=dtype, device=device)
776
+ if attention_mask is not None and attention_mask.dim() == 2:
777
+ mask_length = attention_mask.shape[-1]
778
+ padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[
779
+ :, None, None, :
780
+ ].eq(0.0)
781
+ causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(
782
+ padding_mask, min_dtype
783
+ )
784
+
785
+ if self.config._attn_implementation == "sdpa" and attention_mask is not None:
786
+ # For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
787
+ is_tracing = (
788
+ torch.jit.is_tracing()
789
+ or isinstance(input_tensor, torch.fx.Proxy)
790
+ or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
791
+ )
792
+ if not is_tracing and torch.any(attention_mask != 1):
793
+ # Attend to all tokens in masked rows from the causal_mask, for example the relevant first rows when
794
+ # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
795
+ # Details: https://github.com/pytorch/pytorch/issues/110213
796
+ causal_mask = causal_mask.mul(
797
+ ~torch.all(causal_mask == min_dtype, dim=-1, keepdim=True)
798
+ ).to(dtype)
799
+
800
+ return causal_mask
801
+
802
+
803
+ class OpenELMForCausalLM(OpenELMPreTrainedModel):
804
+ _tied_weights_keys = ["lm_head.weight"]
805
+
806
+ def __init__(self, config: OpenELMConfig):
807
+ super().__init__(config)
808
+ self.transformer = OpenELMModel(config)
809
+ self.vocab_size = config.vocab_size
810
+ if config.share_input_output_layers:
811
+ self.lm_head = None
812
+ else:
813
+ self.lm_head = nn.Linear(config.model_dim, config.vocab_size, bias=False)
814
+
815
+ # Initialize weights and apply final processing
816
+ self.post_init()
817
+
818
+ def get_input_embeddings(self):
819
+ return self.transformer.token_embeddings
820
+
821
+ def set_input_embeddings(self, value):
822
+ self.transformer.token_embeddings = value
823
+
824
+ def get_output_embeddings(self):
825
+ return self.lm_head
826
+
827
+ def set_output_embeddings(self, new_embeddings):
828
+ self.lm_head = new_embeddings
829
+
830
+ def set_decoder(self, decoder):
831
+ self.transformer = decoder
832
+
833
+ def get_decoder(self):
834
+ return self.transformer
835
+
836
+ def forward(
837
+ self,
838
+ input_ids: torch.LongTensor = None,
839
+ attention_mask: Optional[torch.Tensor] = None,
840
+ position_ids: Optional[torch.LongTensor] = None,
841
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
842
+ inputs_embeds: Optional[torch.FloatTensor] = None,
843
+ labels: Optional[torch.LongTensor] = None,
844
+ use_cache: Optional[bool] = None,
845
+ output_attentions: Optional[bool] = None,
846
+ output_hidden_states: Optional[bool] = None,
847
+ return_dict: Optional[bool] = None,
848
+ cache_position: Optional[torch.LongTensor] = None,
849
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
850
+ output_attentions = (
851
+ output_attentions
852
+ if output_attentions is not None
853
+ else self.config.output_attentions
854
+ )
855
+ output_hidden_states = (
856
+ output_hidden_states
857
+ if output_hidden_states is not None
858
+ else self.config.output_hidden_states
859
+ )
860
+ return_dict = (
861
+ return_dict if return_dict is not None else self.config.use_return_dict
862
+ )
863
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
864
+ outputs = self.transformer(
865
+ input_ids=input_ids,
866
+ attention_mask=attention_mask,
867
+ position_ids=position_ids,
868
+ past_key_values=past_key_values,
869
+ inputs_embeds=inputs_embeds,
870
+ use_cache=use_cache,
871
+ output_attentions=output_attentions,
872
+ output_hidden_states=output_hidden_states,
873
+ return_dict=return_dict,
874
+ cache_position=cache_position,
875
+ )
876
+
877
+ hidden_states = outputs[0]
878
+ if self.lm_head is None:
879
+ # shared
880
+ logits = F.linear(
881
+ hidden_states, weight=self.transformer.token_embeddings.weight
882
+ )
883
+ else:
884
+ logits = self.lm_head(hidden_states)
885
+ logits = logits[:, : self.config.vocab_size]
886
+ loss = None
887
+ if labels is not None:
888
+ # Shift so that tokens < n predict n
889
+ shift_logits = logits[..., :-1, :].contiguous()
890
+ shift_labels = labels[..., 1:].contiguous()
891
+ # Flatten the tokens
892
+ loss_fct = CrossEntropyLoss()
893
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
894
+ shift_labels = shift_labels.view(-1)
895
+ # Enable model parallelism
896
+ shift_labels = shift_labels.to(shift_logits.device)
897
+ loss = loss_fct(shift_logits, shift_labels)
898
+
899
+ if not return_dict:
900
+ output = (logits,) + outputs[1:]
901
+ return (loss,) + output if loss is not None else output
902
+
903
+ return CausalLMOutputWithPast(
904
+ loss=loss,
905
+ logits=logits,
906
+ past_key_values=outputs.past_key_values,
907
+ hidden_states=outputs.hidden_states,
908
+ attentions=outputs.attentions,
909
+ )
910
+
911
+ def prepare_inputs_for_generation(
912
+ self,
913
+ input_ids,
914
+ past_key_values=None,
915
+ attention_mask=None,
916
+ inputs_embeds=None,
917
+ **kwargs,
918
+ ):
919
+ past_length = 0
920
+ if past_key_values is not None:
921
+ if isinstance(past_key_values, Cache):
922
+ cache_length = past_key_values.get_seq_length()
923
+ past_length = past_key_values.seen_tokens
924
+ max_cache_length = past_key_values.get_max_length()
925
+ else:
926
+ cache_length = past_length = past_key_values[0][0].shape[2]
927
+ max_cache_length = None
928
+
929
+ # Keep only the unprocessed tokens:
930
+ # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
931
+ # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
932
+ # input)
933
+ if (
934
+ attention_mask is not None
935
+ and attention_mask.shape[1] > input_ids.shape[1]
936
+ ):
937
+ input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
938
+ # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
939
+ # input_ids based on the past_length.
940
+ elif past_length < input_ids.shape[1]:
941
+ input_ids = input_ids[:, past_length:]
942
+ # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
943
+
944
+ # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
945
+ if (
946
+ max_cache_length is not None
947
+ and attention_mask is not None
948
+ and cache_length + input_ids.shape[1] > max_cache_length
949
+ ):
950
+ attention_mask = attention_mask[:, -max_cache_length:]
951
+
952
+ position_ids = kwargs.get("position_ids", None)
953
+ if attention_mask is not None and position_ids is None:
954
+ # create position_ids on the fly for batch generation
955
+ position_ids = attention_mask.long().cumsum(-1) - 1
956
+ position_ids.masked_fill_(attention_mask == 0, 1)
957
+ if past_key_values:
958
+ position_ids = position_ids[:, -input_ids.shape[1] :]
959
+
960
+ if self.generation_config.cache_implementation == "static":
961
+ # generation with static cache
962
+ cache_position = kwargs.get("cache_position", None)
963
+ if cache_position is None:
964
+ past_length = 0
965
+ else:
966
+ past_length = cache_position[-1] + 1
967
+ input_ids = input_ids[:, past_length:]
968
+ position_ids = position_ids[:, past_length:]
969
+
970
+ # we should only keep a `cache_position` in generate, and do +=1.
971
+ # same goes for position ids. Could also help with continued generation.
972
+ cache_position = torch.arange(
973
+ past_length,
974
+ past_length + position_ids.shape[-1],
975
+ device=position_ids.device,
976
+ )
977
+
978
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
979
+ if inputs_embeds is not None and past_key_values is None:
980
+ model_inputs = {"inputs_embeds": inputs_embeds}
981
+ else:
982
+ # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
983
+ # recompiles graphs as the stride of the inputs is a guard. Ref: https://github.com/huggingface/transformers/pull/29114
984
+ # We could use `next_tokens` directly instead.
985
+ model_inputs = {"input_ids": input_ids.contiguous()}
986
+
987
+ model_inputs.update(
988
+ {
989
+ "position_ids": position_ids.contiguous(),
990
+ "cache_position": cache_position,
991
+ "past_key_values": past_key_values,
992
+ "use_cache": kwargs.get("use_cache"),
993
+ "attention_mask": attention_mask,
994
+ }
995
+ )
996
+ return model_inputs
997
+
998
+ @staticmethod
999
+ def _reorder_cache(past_key_values, beam_idx):
1000
+ reordered_past = ()
1001
+ for layer_past in past_key_values:
1002
+ reordered_past += (
1003
+ tuple(
1004
+ past_state.index_select(0, beam_idx.to(past_state.device))
1005
+ for past_state in layer_past
1006
+ ),
1007
+ )
1008
+ return reordered_past
special_tokens_map.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "</s>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "unk_token": {
17
+ "content": "<unk>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ }
23
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
3
+ size 499723
tokenizer_config.json ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": true,
3
+ "add_eos_token": false,
4
+ "added_tokens_decoder": {
5
+ "0": {
6
+ "content": "<unk>",
7
+ "lstrip": false,
8
+ "normalized": false,
9
+ "rstrip": false,
10
+ "single_word": false,
11
+ "special": true
12
+ },
13
+ "1": {
14
+ "content": "<s>",
15
+ "lstrip": false,
16
+ "normalized": false,
17
+ "rstrip": false,
18
+ "single_word": false,
19
+ "special": true
20
+ },
21
+ "2": {
22
+ "content": "</s>",
23
+ "lstrip": false,
24
+ "normalized": false,
25
+ "rstrip": false,
26
+ "single_word": false,
27
+ "special": true
28
+ }
29
+ },
30
+ "bos_token": "<s>",
31
+ "clean_up_tokenization_spaces": false,
32
+ "eos_token": "</s>",
33
+ "legacy": false,
34
+ "model_max_length": 1000000000000000019884624838656,
35
+ "pad_token": null,
36
+ "padding_side": "right",
37
+ "sp_model_kwargs": {},
38
+ "tokenizer_class": "LlamaTokenizer",
39
+ "unk_token": "<unk>",
40
+ "use_default_system_prompt": false
41
+ }