Jason0214 commited on
Commit
b02c73e
1 Parent(s): 2f33f25

Upload folder using huggingface_hub

Browse files
config.json ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "MBZUAI-LLM/SlimPajama-DC-6",
3
+ "architectures": [
4
+ "BTLMLMHeadModel"
5
+ ],
6
+ "auto_map": {
7
+ "AutoConfig": "MBZUAI-LLM/SlimPajama-DC-6--configuration_btlm.BTLMConfig",
8
+ "AutoModel": "MBZUAI-LLM/SlimPajama-DC-6--modeling_btlm.BTLMModel",
9
+ "AutoModelForCausalLM": "MBZUAI-LLM/SlimPajama-DC-6--modeling_btlm.BTLMLMHeadModel"
10
+ },
11
+ "model_type": "btlm",
12
+ "eos_token_id": 0,
13
+ "pad_token_id": 0,
14
+ "attn_pdrop": 0.0,
15
+ "scale_attn_weights": true,
16
+ "resid_pdrop": 0.0,
17
+ "embeddings_scale": 14.6,
18
+ "n_inner": 5461,
19
+ "n_embd": 2048,
20
+ "layer_norm_epsilon": 1e-05,
21
+ "n_positions": 2048,
22
+ "activation_function": "swiglu",
23
+ "n_head": 16,
24
+ "n_layer": 24,
25
+ "width_scale": 0.2775,
26
+ "position_embedding_type": "alibi",
27
+ "scale_qk_dot_by_d": true,
28
+ "tie_word_embeddings": true,
29
+ "vocab_size": 50277,
30
+ "embd_pdrop": 0.0
31
+ }
configuration_btlm.py ADDED
@@ -0,0 +1,160 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2023 The OpenAI Team Authors and HuggingFace Inc. team.
3
+ # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
4
+ # Copyright 2023 Cerebras Systems.
5
+ #
6
+ # Licensed under the Apache License, Version 2.0 (the "License");
7
+ # you may not use this file except in compliance with the License.
8
+ # You may obtain a copy of the License at
9
+ #
10
+ # http://www.apache.org/licenses/LICENSE-2.0
11
+ #
12
+ # Unless required by applicable law or agreed to in writing, software
13
+ # distributed under the License is distributed on an "AS IS" BASIS,
14
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15
+ # See the License for the specific language governing permissions and
16
+ # limitations under the License.
17
+ """ BTLM configuration"""
18
+
19
+ from transformers.configuration_utils import PretrainedConfig
20
+ from transformers.utils import logging
21
+
22
+
23
+ logger = logging.get_logger(__name__)
24
+
25
+ BTLM_PRETRAINED_CONFIG_ARCHIVE_MAP = {
26
+ "cerebras/BTLM-2.7B": "https://huggingface.co/cerebras/BTLM-2.7B/resolve/main/config.json",
27
+ }
28
+
29
+ class BTLMConfig(PretrainedConfig):
30
+ """
31
+ This is the configuration class to store the configuration of a [`BTLMModel`]. It is used to
32
+ instantiate a BTLM model according to the specified arguments, defining the model architecture.
33
+
34
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
35
+ documentation from [`PretrainedConfig`] for more information.
36
+
37
+
38
+ Args:
39
+ vocab_size (`int`, *optional*, defaults to 50257):
40
+ Vocabulary size of the BTLM model. Defines the number of different tokens that can be represented by the
41
+ `inputs_ids` passed when calling [`BTLMModel`].
42
+ n_positions (`int`, *optional*, defaults to 1024):
43
+ The maximum sequence length that this model might ever be used with. Typically set this to something large
44
+ just in case (e.g., 512 or 1024 or 2048).
45
+ n_embd (`int`, *optional*, defaults to 768):
46
+ Dimensionality of the embeddings and hidden states.
47
+ n_layer (`int`, *optional*, defaults to 12):
48
+ Number of hidden layers in the Transformer encoder.
49
+ n_head (`int`, *optional*, defaults to 12):
50
+ Number of attention heads for each attention layer in the Transformer encoder.
51
+ n_inner (`int`, *optional*, defaults to None):
52
+ Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
53
+ activation_function (`str`, *optional*, defaults to `"gelu"`):
54
+ Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new", "swiglu"]`.
55
+ resid_pdrop (`float`, *optional*, defaults to 0.1):
56
+ The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
57
+ embd_pdrop (`float`, *optional*, defaults to 0.1):
58
+ The dropout ratio for the embeddings.
59
+ attn_pdrop (`float`, *optional*, defaults to 0.1):
60
+ The dropout ratio for the attention.
61
+ layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
62
+ The epsilon to use in the layer normalization layers.
63
+ initializer_range (`float`, *optional*, defaults to 0.02):
64
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
65
+ scale_attn_weights (`bool`, *optional*, defaults to `True`):
66
+ Scale attention weights by dividing by sqrt(hidden_size)..
67
+ use_cache (`bool`, *optional*, defaults to `True`):
68
+ Whether or not the model should return the last key/values attentions (not used by all models).
69
+ scale_attn_by_inverse_layer_idx (`bool`, *optional*, defaults to `False`):
70
+ Whether to additionally scale attention weights by `1 / layer_idx + 1`.
71
+ reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`):
72
+ Whether to scale keys (K) prior to computing attention (dot-product) and upcast attention
73
+ dot-product/softmax to float() when training with mixed precision.
74
+ position_embedding_type (`str`, *optional*, defaults to `"learned"`):
75
+ Positional embedding can be either `"alibi"` or `"learned"`.
76
+ width_scale (`float`, *optional*, defaults to 1.0):
77
+ muP parameter to scale output logits and initializers. Calculated as (`d_model,0 / d_model`),
78
+ where `d_model` is the model's width and `d_model,0` is the proxy model's width.
79
+ embeddings_scale (`float`, *optional*, defaults to 1.0):
80
+ muP parameter to scale token and position embeddings.
81
+ scale_qk_dot_by_d (`bool`, *optional*, defaults to `False`):
82
+ Scale attention weights by dividing by hidden_size instead of sqrt(hidden_size).
83
+ Need to set scale_attn_weights to `True` as well.
84
+
85
+ Example:
86
+
87
+ ```python
88
+ >>> from transformers import BTLMConfig, BTLMModel
89
+
90
+ >>> # Initializing a BTLM configuration
91
+ >>> configuration = BTLMConfig()
92
+
93
+ >>> # Initializing a model (with random weights) from the configuration
94
+ >>> model = BTLMModel(configuration)
95
+
96
+ >>> # Accessing the model configuration
97
+ >>> configuration = model.config
98
+ ```"""
99
+
100
+ model_type = "btlm"
101
+ keys_to_ignore_at_inference = ["past_key_values"]
102
+ attribute_map = {
103
+ "hidden_size": "n_embd",
104
+ "max_position_embeddings": "n_positions",
105
+ "num_attention_heads": "n_head",
106
+ "num_hidden_layers": "n_layer",
107
+ }
108
+
109
+ def __init__(
110
+ self,
111
+ vocab_size=50257,
112
+ n_positions=1024,
113
+ n_embd=768,
114
+ n_layer=12,
115
+ n_head=12,
116
+ n_inner=None,
117
+ activation_function="gelu_new",
118
+ resid_pdrop=0.1,
119
+ embd_pdrop=0.1,
120
+ attn_pdrop=0.1,
121
+ layer_norm_epsilon=1e-5,
122
+ initializer_range=0.02,
123
+ scale_attn_weights=True,
124
+ use_cache=True,
125
+ bos_token_id=50256,
126
+ eos_token_id=50256,
127
+ scale_attn_by_inverse_layer_idx=False,
128
+ reorder_and_upcast_attn=False,
129
+ position_embedding_type="learned",
130
+ width_scale=1.0,
131
+ embeddings_scale=1.0,
132
+ scale_qk_dot_by_d=False,
133
+ **kwargs,
134
+ ):
135
+ self.vocab_size = vocab_size
136
+ self.n_positions = n_positions
137
+ self.n_embd = n_embd
138
+ self.n_layer = n_layer
139
+ self.n_head = n_head
140
+ self.n_inner = n_inner
141
+ self.activation_function = activation_function
142
+ self.resid_pdrop = resid_pdrop
143
+ self.embd_pdrop = embd_pdrop
144
+ self.attn_pdrop = attn_pdrop
145
+ self.layer_norm_epsilon = layer_norm_epsilon
146
+ self.initializer_range = initializer_range
147
+ self.scale_attn_weights = scale_attn_weights
148
+ self.use_cache = use_cache
149
+ self.scale_attn_by_inverse_layer_idx = scale_attn_by_inverse_layer_idx
150
+ self.reorder_and_upcast_attn = reorder_and_upcast_attn
151
+
152
+ self.bos_token_id = bos_token_id
153
+ self.eos_token_id = eos_token_id
154
+
155
+ self.position_embedding_type = position_embedding_type
156
+ self.width_scale = width_scale
157
+ self.embeddings_scale = embeddings_scale
158
+ self.scale_qk_dot_by_d = scale_qk_dot_by_d
159
+
160
+ super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
modeling_btlm.py ADDED
@@ -0,0 +1,1528 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2023 The OpenAI Team Authors and HuggingFace Inc. team.
3
+ # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
4
+ # Copyright 2023 Cerebras Systems.
5
+ #
6
+ # Licensed under the Apache License, Version 2.0 (the "License");
7
+ # you may not use this file except in compliance with the License.
8
+ # You may obtain a copy of the License at
9
+ #
10
+ # http://www.apache.org/licenses/LICENSE-2.0
11
+ #
12
+ # Unless required by applicable law or agreed to in writing, software
13
+ # distributed under the License is distributed on an "AS IS" BASIS,
14
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15
+ # See the License for the specific language governing permissions and
16
+ # limitations under the License.
17
+ """ PyTorch BTLM model."""
18
+
19
+ import math
20
+ import os
21
+ import warnings
22
+ from typing import Optional, Tuple, Union
23
+
24
+ import torch
25
+ from torch import Tensor, nn
26
+ from torch.cuda.amp import autocast
27
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
28
+
29
+ from transformers.activations import ACT2FN
30
+ from transformers.modeling_outputs import (
31
+ BaseModelOutputWithPastAndCrossAttentions,
32
+ CausalLMOutputWithCrossAttentions,
33
+ QuestionAnsweringModelOutput,
34
+ SequenceClassifierOutputWithPast,
35
+ TokenClassifierOutput,
36
+ )
37
+ from transformers.modeling_utils import PreTrainedModel
38
+ from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
39
+ from transformers.utils import (
40
+ add_code_sample_docstrings,
41
+ add_start_docstrings,
42
+ add_start_docstrings_to_model_forward,
43
+ logging,
44
+ )
45
+ from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
46
+ # from .configuration_btlm import BTLMConfig
47
+ from configuration_btlm import BTLMConfig
48
+
49
+
50
+ logger = logging.get_logger(__name__)
51
+
52
+ _CHECKPOINT_FOR_DOC = "cerebras/BTLM-2.7B"
53
+ _CONFIG_FOR_DOC = "BTLMConfig"
54
+
55
+ BTLM_PRETRAINED_MODEL_ARCHIVE_LIST = [
56
+ "cerebras/BTLM-2.7B",
57
+ # See all BTLM models at https://huggingface.co/models?filter=btlm
58
+ ]
59
+
60
+
61
+ class SwiGLUActivation(nn.Module):
62
+ def forward(self, x1: Tensor, x2: Tensor) -> Tensor:
63
+ return x1 * nn.functional.silu(x2)
64
+
65
+
66
+ class AlibiPositionEmbeddingLayer(nn.Module):
67
+ def __init__(self, num_heads):
68
+ super(AlibiPositionEmbeddingLayer, self).__init__()
69
+
70
+ self.num_heads = num_heads
71
+ slopes = torch.tensor(
72
+ AlibiPositionEmbeddingLayer._get_alibi_slopes(num_heads)
73
+ ).unsqueeze(-1)
74
+ self.slopes = nn.parameter.Parameter(slopes, requires_grad=False)
75
+
76
+ def forward(self, seq_length, key_length, cached_qk_len):
77
+ context_position = torch.arange(
78
+ cached_qk_len, cached_qk_len + seq_length, device=self.slopes.device
79
+ )[:, None]
80
+ memory_position = torch.arange(
81
+ key_length + cached_qk_len, device=self.slopes.device
82
+ )[None, :]
83
+ relative_position = memory_position - context_position
84
+ relative_position = torch.abs(relative_position).unsqueeze(0).expand(self.num_heads, -1, -1)
85
+ alibi = (self.slopes * -1.0).unsqueeze(1) * relative_position
86
+ return alibi
87
+
88
+ @staticmethod
89
+ def _get_alibi_slopes(n):
90
+ def get_slopes_power_of_2(n):
91
+ start = 2 ** (-(2 ** -(math.log2(n) - 3)))
92
+ ratio = start
93
+ return [start * ratio ** i for i in range(n)]
94
+
95
+ if math.log2(n).is_integer():
96
+ return get_slopes_power_of_2(
97
+ n
98
+ ) # In the paper, we only train models that have 2^a heads for some a. This function has
99
+ else: # some good properties that only occur when the input is a power of 2. To maintain that even
100
+ closest_power_of_2 = 2 ** math.floor(
101
+ math.log2(n)
102
+ ) # when the number of heads is not a power of 2, we use this workaround.
103
+ return (
104
+ get_slopes_power_of_2(closest_power_of_2)
105
+ + AlibiPositionEmbeddingLayer._get_alibi_slopes(
106
+ 2 * closest_power_of_2
107
+ )[0::2][: n - closest_power_of_2]
108
+ )
109
+
110
+
111
+ def load_tf_weights_in_btlm(model, config, btlm_checkpoint_path):
112
+ """Load tf checkpoints in a pytorch model"""
113
+ try:
114
+ import re
115
+
116
+ import tensorflow as tf
117
+ except ImportError:
118
+ logger.error(
119
+ "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
120
+ "https://www.tensorflow.org/install/ for installation instructions."
121
+ )
122
+ raise
123
+ tf_path = os.path.abspath(btlm_checkpoint_path)
124
+ logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
125
+ # Load weights from TF model
126
+ init_vars = tf.train.list_variables(tf_path)
127
+ names = []
128
+ arrays = []
129
+ for name, shape in init_vars:
130
+ logger.info(f"Loading TF weight {name} with shape {shape}")
131
+ array = tf.train.load_variable(tf_path, name)
132
+ names.append(name)
133
+ arrays.append(array.squeeze())
134
+
135
+ for name, array in zip(names, arrays):
136
+ name = name[6:] # skip "model/"
137
+ name = name.split("/")
138
+ pointer = model
139
+ for m_name in name:
140
+ if re.fullmatch(r"[A-Za-z]+\d+", m_name):
141
+ scope_names = re.split(r"(\d+)", m_name)
142
+ else:
143
+ scope_names = [m_name]
144
+ if scope_names[0] == "w" or scope_names[0] == "g":
145
+ pointer = getattr(pointer, "weight")
146
+ elif scope_names[0] == "b":
147
+ pointer = getattr(pointer, "bias")
148
+ elif scope_names[0] == "wpe" or scope_names[0] == "wte":
149
+ pointer = getattr(pointer, scope_names[0])
150
+ pointer = getattr(pointer, "weight")
151
+ else:
152
+ pointer = getattr(pointer, scope_names[0])
153
+ if len(scope_names) >= 2:
154
+ num = int(scope_names[1])
155
+ pointer = pointer[num]
156
+ try:
157
+ assert (
158
+ pointer.shape == array.shape
159
+ ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
160
+ except AssertionError as e:
161
+ e.args += (pointer.shape, array.shape)
162
+ raise
163
+ logger.info(f"Initialize PyTorch weight {name}")
164
+ pointer.data = torch.from_numpy(array)
165
+ return model
166
+
167
+
168
+ class BTLMAttention(nn.Module):
169
+ def __init__(self, config, is_cross_attention=False, layer_idx=None):
170
+ super().__init__()
171
+
172
+ max_positions = config.max_position_embeddings
173
+ self.register_buffer(
174
+ "bias",
175
+ torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(
176
+ 1, 1, max_positions, max_positions
177
+ ),
178
+ persistent=False,
179
+ )
180
+ self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)
181
+
182
+ self.embed_dim = config.hidden_size
183
+ self.num_heads = config.num_attention_heads
184
+ self.head_dim = self.embed_dim // self.num_heads
185
+ self.split_size = self.embed_dim
186
+ if self.head_dim * self.num_heads != self.embed_dim:
187
+ raise ValueError(
188
+ f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
189
+ f" {self.num_heads})."
190
+ )
191
+
192
+ self.scale_attn_weights = config.scale_attn_weights
193
+ self.is_cross_attention = is_cross_attention
194
+
195
+ # Layer-wise attention scaling, reordering, and upcasting
196
+ self.scale_attn_by_inverse_layer_idx = config.scale_attn_by_inverse_layer_idx
197
+ self.layer_idx = layer_idx
198
+ self.reorder_and_upcast_attn = config.reorder_and_upcast_attn
199
+
200
+ if self.is_cross_attention:
201
+ self.c_attn = Conv1D(2 * self.embed_dim, self.embed_dim)
202
+ self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
203
+ else:
204
+ self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
205
+ self.c_proj = Conv1D(self.embed_dim, self.embed_dim)
206
+
207
+ self.attn_dropout = nn.Dropout(config.attn_pdrop)
208
+ self.resid_dropout = nn.Dropout(config.resid_pdrop)
209
+
210
+ self.pruned_heads = set()
211
+
212
+ self.attn_scale_power = 1.0 if config.scale_qk_dot_by_d else 0.5
213
+
214
+ def prune_heads(self, heads):
215
+ if len(heads) == 0:
216
+ return
217
+ heads, index = find_pruneable_heads_and_indices(heads, self.num_heads, self.head_dim, self.pruned_heads)
218
+ index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])
219
+
220
+ # Prune conv1d layers
221
+ self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
222
+ self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
223
+
224
+ # Update hyper params
225
+ self.split_size = (self.split_size // self.num_heads) * (self.num_heads - len(heads))
226
+ self.num_heads = self.num_heads - len(heads)
227
+ self.pruned_heads = self.pruned_heads.union(heads)
228
+
229
+ def _attn(self, query, key, value, attention_mask=None, head_mask=None, position_bias=None):
230
+ attn_weights = torch.matmul(query, key.transpose(-1, -2))
231
+
232
+ if self.scale_attn_weights:
233
+ attn_weights = attn_weights / torch.full(
234
+ [], value.size(-1) ** self.attn_scale_power, dtype=attn_weights.dtype, device=attn_weights.device
235
+ )
236
+
237
+ # Layer-wise attention scaling
238
+ if self.scale_attn_by_inverse_layer_idx:
239
+ attn_weights = attn_weights / float(self.layer_idx + 1)
240
+
241
+ if not self.is_cross_attention:
242
+ # if only "normal" attention layer implements causal mask
243
+ query_length, key_length = query.size(-2), key.size(-2)
244
+ causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
245
+ mask_value = torch.finfo(attn_weights.dtype).min
246
+ # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
247
+ # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
248
+ mask_value = torch.full([], mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
249
+ attn_weights = torch.where(causal_mask, attn_weights.to(attn_weights.dtype), mask_value)
250
+
251
+ if attention_mask is not None:
252
+ # Apply the attention mask
253
+ attn_weights = attn_weights + attention_mask
254
+
255
+ if position_bias is not None:
256
+ attn_weights += position_bias.type_as(attn_weights).unsqueeze(0)
257
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1)
258
+
259
+ # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
260
+ attn_weights = attn_weights.type(value.dtype)
261
+ attn_weights = self.attn_dropout(attn_weights)
262
+
263
+ # Mask heads if we want to
264
+ if head_mask is not None:
265
+ attn_weights = attn_weights * head_mask
266
+
267
+ attn_output = torch.matmul(attn_weights, value)
268
+
269
+ return attn_output, attn_weights
270
+
271
+ def _upcast_and_reordered_attn(self, query, key, value, attention_mask=None, head_mask=None, position_bias=None):
272
+ # Use `torch.baddbmm` (a bit more efficient w/ alpha param for scaling -- from Megatron-LM)
273
+ bsz, num_heads, q_seq_len, dk = query.size()
274
+ _, _, k_seq_len, _ = key.size()
275
+
276
+ # Preallocate attn_weights for `baddbmm`
277
+ attn_weights = torch.empty(bsz * num_heads, q_seq_len, k_seq_len, dtype=torch.float32, device=query.device)
278
+
279
+ # Compute Scale Factor
280
+ scale_factor = 1.0
281
+ if self.scale_attn_weights:
282
+ scale_factor /= float(value.size(-1)) ** self.attn_scale_power
283
+
284
+ if self.scale_attn_by_inverse_layer_idx:
285
+ scale_factor /= float(self.layer_idx + 1)
286
+
287
+ # Upcast (turn off autocast) and reorder (Scale K by 1 / root(dk))
288
+ with autocast(enabled=False):
289
+ q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(-1, dk, k_seq_len)
290
+ attn_weights = torch.baddbmm(attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor)
291
+ attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
292
+
293
+ if not self.is_cross_attention:
294
+ # if only "normal" attention layer implements causal mask
295
+ query_length, key_length = query.size(-2), key.size(-2)
296
+ causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
297
+ mask_value = torch.finfo(attn_weights.dtype).min
298
+ # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
299
+ # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
300
+ mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
301
+ attn_weights = torch.where(causal_mask, attn_weights, mask_value)
302
+
303
+ if attention_mask is not None:
304
+ # Apply the attention mask
305
+ attn_weights = attn_weights + attention_mask
306
+
307
+ if position_bias is not None:
308
+ attn_weights += position_bias.type_as(attn_weights).unsqueeze(0)
309
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1)
310
+
311
+ # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op if otherwise
312
+ if attn_weights.dtype != torch.float32:
313
+ raise RuntimeError("Error with upcasting, attn_weights does not have dtype torch.float32")
314
+ attn_weights = attn_weights.type(value.dtype)
315
+ attn_weights = self.attn_dropout(attn_weights)
316
+
317
+ # Mask heads if we want to
318
+ if head_mask is not None:
319
+ attn_weights = attn_weights * head_mask
320
+
321
+ attn_output = torch.matmul(attn_weights, value)
322
+
323
+ return attn_output, attn_weights
324
+
325
+ def _split_heads(self, tensor, num_heads, attn_head_size):
326
+ """
327
+ Splits hidden_size dim into attn_head_size and num_heads
328
+ """
329
+ new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
330
+ tensor = tensor.view(new_shape)
331
+ return tensor.permute(0, 2, 1, 3) # (batch, head, seq_length, head_features)
332
+
333
+ def _merge_heads(self, tensor, num_heads, attn_head_size):
334
+ """
335
+ Merges attn_head_size dim and num_attn_heads dim into hidden_size
336
+ """
337
+ tensor = tensor.permute(0, 2, 1, 3).contiguous()
338
+ new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
339
+ return tensor.view(new_shape)
340
+
341
+ def forward(
342
+ self,
343
+ hidden_states: Optional[Tuple[torch.FloatTensor]],
344
+ layer_past: Optional[Tuple[torch.Tensor]] = None,
345
+ attention_mask: Optional[torch.FloatTensor] = None,
346
+ head_mask: Optional[torch.FloatTensor] = None,
347
+ encoder_hidden_states: Optional[torch.Tensor] = None,
348
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
349
+ use_cache: Optional[bool] = False,
350
+ output_attentions: Optional[bool] = False,
351
+ position_bias: Optional[torch.FloatTensor] = None,
352
+ ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
353
+ if encoder_hidden_states is not None:
354
+ if not hasattr(self, "q_attn"):
355
+ raise ValueError(
356
+ "If class is used as cross attention, the weights `q_attn` have to be defined. "
357
+ "Please make sure to instantiate class with `BTLMAttention(..., is_cross_attention=True)`."
358
+ )
359
+
360
+ query = self.q_attn(hidden_states)
361
+ key, value = self.c_attn(encoder_hidden_states).split(self.split_size, dim=2)
362
+ attention_mask = encoder_attention_mask
363
+ else:
364
+ query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
365
+
366
+ query = self._split_heads(query, self.num_heads, self.head_dim)
367
+ key = self._split_heads(key, self.num_heads, self.head_dim)
368
+ value = self._split_heads(value, self.num_heads, self.head_dim)
369
+
370
+ if layer_past is not None:
371
+ past_key, past_value = layer_past
372
+ key = torch.cat((past_key, key), dim=-2)
373
+ value = torch.cat((past_value, value), dim=-2)
374
+
375
+ if use_cache is True:
376
+ present = (key, value)
377
+ else:
378
+ present = None
379
+
380
+ if self.reorder_and_upcast_attn:
381
+ attn_output, attn_weights = self._upcast_and_reordered_attn(query, key, value, attention_mask, head_mask, position_bias)
382
+ else:
383
+ attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask, position_bias)
384
+
385
+ attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
386
+ attn_output = self.c_proj(attn_output)
387
+ attn_output = self.resid_dropout(attn_output)
388
+
389
+ outputs = (attn_output, present)
390
+ if output_attentions:
391
+ outputs += (attn_weights,)
392
+
393
+ return outputs # a, present, (attentions)
394
+
395
+
396
+ class BTLMMLP(nn.Module):
397
+ def __init__(self, intermediate_size, config):
398
+ super().__init__()
399
+ embed_dim = config.hidden_size
400
+ self.swiglu = config.activation_function == "swiglu"
401
+ self.c_fc = Conv1D(intermediate_size, embed_dim)
402
+ self.c_fc2 = Conv1D(intermediate_size, embed_dim) if self.swiglu else None
403
+ self.c_proj = Conv1D(embed_dim, intermediate_size)
404
+ self.act = SwiGLUActivation() if self.swiglu else ACT2FN[config.activation_function]
405
+ self.dropout = nn.Dropout(config.resid_pdrop)
406
+
407
+ def forward(self, hidden_states: Optional[Tuple[torch.FloatTensor]]) -> torch.FloatTensor:
408
+ if self.swiglu:
409
+ hidden_states2 = self.c_fc2(hidden_states)
410
+ hidden_states = self.c_fc(hidden_states)
411
+ hidden_states = self.act(hidden_states, hidden_states2) if self.swiglu else self.act(hidden_states)
412
+ hidden_states = self.c_proj(hidden_states)
413
+ hidden_states = self.dropout(hidden_states)
414
+ return hidden_states
415
+
416
+
417
+ class BTLMBlock(nn.Module):
418
+ def __init__(self, config, layer_idx=None):
419
+ super().__init__()
420
+ hidden_size = config.hidden_size
421
+ inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
422
+
423
+ self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
424
+ self.attn = BTLMAttention(config, layer_idx=layer_idx)
425
+ self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
426
+
427
+ if config.add_cross_attention:
428
+ self.crossattention = BTLMAttention(config, is_cross_attention=True, layer_idx=layer_idx)
429
+ self.ln_cross_attn = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
430
+
431
+ self.mlp = BTLMMLP(inner_dim, config)
432
+
433
+ def forward(
434
+ self,
435
+ hidden_states: Optional[Tuple[torch.FloatTensor]],
436
+ layer_past: Optional[Tuple[torch.Tensor]] = None,
437
+ attention_mask: Optional[torch.FloatTensor] = None,
438
+ head_mask: Optional[torch.FloatTensor] = None,
439
+ encoder_hidden_states: Optional[torch.Tensor] = None,
440
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
441
+ use_cache: Optional[bool] = False,
442
+ output_attentions: Optional[bool] = False,
443
+ position_bias: Optional[torch.FloatTensor] = None,
444
+ ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
445
+ residual = hidden_states
446
+ hidden_states = self.ln_1(hidden_states)
447
+ attn_outputs = self.attn(
448
+ hidden_states,
449
+ layer_past=layer_past,
450
+ attention_mask=attention_mask,
451
+ head_mask=head_mask,
452
+ use_cache=use_cache,
453
+ output_attentions=output_attentions,
454
+ position_bias=position_bias,
455
+ )
456
+ attn_output = attn_outputs[0] # output_attn: a, present, (attentions)
457
+ outputs = attn_outputs[1:]
458
+ # residual connection
459
+ hidden_states = attn_output + residual
460
+
461
+ if encoder_hidden_states is not None:
462
+ # add one self-attention block for cross-attention
463
+ if not hasattr(self, "crossattention"):
464
+ raise ValueError(
465
+ f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
466
+ "cross-attention layers by setting `config.add_cross_attention=True`"
467
+ )
468
+ residual = hidden_states
469
+ hidden_states = self.ln_cross_attn(hidden_states)
470
+ cross_attn_outputs = self.crossattention(
471
+ hidden_states,
472
+ attention_mask=attention_mask,
473
+ head_mask=head_mask,
474
+ encoder_hidden_states=encoder_hidden_states,
475
+ encoder_attention_mask=encoder_attention_mask,
476
+ output_attentions=output_attentions,
477
+ position_bias=position_bias,
478
+ )
479
+ attn_output = cross_attn_outputs[0]
480
+ # residual connection
481
+ hidden_states = residual + attn_output
482
+ outputs = outputs + cross_attn_outputs[2:] # add cross attentions if we output attention weights
483
+
484
+ residual = hidden_states
485
+ hidden_states = self.ln_2(hidden_states)
486
+ feed_forward_hidden_states = self.mlp(hidden_states)
487
+ # residual connection
488
+ hidden_states = residual + feed_forward_hidden_states
489
+
490
+ if use_cache:
491
+ outputs = (hidden_states,) + outputs
492
+ else:
493
+ outputs = (hidden_states,) + outputs[1:]
494
+
495
+ return outputs # hidden_states, present, (attentions, cross_attentions)
496
+
497
+
498
+ class BTLMPreTrainedModel(PreTrainedModel):
499
+ """
500
+ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
501
+ models.
502
+ """
503
+
504
+ config_class = BTLMConfig
505
+ load_tf_weights = load_tf_weights_in_btlm
506
+ base_model_prefix = "transformer"
507
+ is_parallelizable = True
508
+ supports_gradient_checkpointing = True
509
+ _no_split_modules = ["BTLMBlock"]
510
+ _skip_keys_device_placement = "past_key_values"
511
+
512
+ def __init__(self, *inputs, **kwargs):
513
+ super().__init__(*inputs, **kwargs)
514
+
515
+ def _init_weights(self, module):
516
+ """Initialize the weights."""
517
+ mup_init_scale = math.sqrt(self.config.width_scale)
518
+ if isinstance(module, (nn.Linear, Conv1D)):
519
+ # Slightly different from the TF version which uses truncated_normal for initialization
520
+ # cf https://github.com/pytorch/pytorch/pull/5617
521
+ module.weight.data.normal_(mean=0.0, std=(self.config.initializer_range * mup_init_scale))
522
+ if module.bias is not None:
523
+ module.bias.data.zero_()
524
+ elif isinstance(module, nn.Embedding):
525
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
526
+ if module.padding_idx is not None:
527
+ module.weight.data[module.padding_idx].zero_()
528
+ elif isinstance(module, nn.LayerNorm):
529
+ module.bias.data.zero_()
530
+ module.weight.data.fill_(1.0)
531
+
532
+ # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
533
+ # > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
534
+ # > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
535
+ # > -- GPT-2 :: https://openai.com/blog/better-language-models/
536
+ #
537
+ # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
538
+ for name, p in module.named_parameters():
539
+ if name == "c_proj.weight":
540
+ # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
541
+ stddev = self.config.initializer_range * mup_init_scale / math.sqrt(2 * self.config.n_layer)
542
+ p.data.normal_(mean=0.0, std=stddev)
543
+
544
+ def _set_gradient_checkpointing(self, module, value=False):
545
+ if isinstance(module, BTLMModel):
546
+ module.gradient_checkpointing = value
547
+
548
+
549
+ BTLM_START_DOCSTRING = r"""
550
+
551
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
552
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
553
+ etc.)
554
+
555
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
556
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
557
+ and behavior.
558
+
559
+ Parameters:
560
+ config ([`BTLMConfig`]): Model configuration class with all the parameters of the model.
561
+ Initializing with a config file does not load the weights associated with the model, only the
562
+ configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
563
+ """
564
+
565
+ BTLM_INPUTS_DOCSTRING = r"""
566
+ Args:
567
+ input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
568
+ `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
569
+ `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
570
+ sequence tokens in the vocabulary.
571
+
572
+ If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
573
+ `input_ids`.
574
+
575
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
576
+ [`PreTrainedTokenizer.__call__`] for details.
577
+
578
+ [What are input IDs?](../glossary#input-ids)
579
+ past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.n_layers`):
580
+ Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
581
+ `past_key_values` output below). Can be used to speed up sequential decoding. The `input_ids` which have
582
+ their past given to this model should not be passed as `input_ids` as they have already been computed.
583
+ attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
584
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
585
+
586
+ - 1 for tokens that are **not masked**,
587
+ - 0 for tokens that are **masked**.
588
+
589
+ If `past_key_values` is used, `attention_mask` needs to contain the masking strategy that was used for
590
+ `past_key_values`. In other words, the `attention_mask` always has to have the length:
591
+ `len(past_key_values) + len(input_ids)`
592
+
593
+ [What are attention masks?](../glossary#attention-mask)
594
+ token_type_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
595
+ Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
596
+ 1]`:
597
+
598
+ - 0 corresponds to a *sentence A* token,
599
+ - 1 corresponds to a *sentence B* token.
600
+
601
+ [What are token type IDs?](../glossary#token-type-ids)
602
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
603
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
604
+ config.max_position_embeddings - 1]`.
605
+
606
+ [What are position IDs?](../glossary#position-ids)
607
+ head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
608
+ Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
609
+
610
+ - 1 indicates the head is **not masked**,
611
+ - 0 indicates the head is **masked**.
612
+
613
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
614
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
615
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
616
+ model's internal embedding lookup matrix.
617
+
618
+ If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see
619
+ `past_key_values`).
620
+ use_cache (`bool`, *optional*):
621
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
622
+ `past_key_values`).
623
+ output_attentions (`bool`, *optional*):
624
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
625
+ tensors for more detail.
626
+ output_hidden_states (`bool`, *optional*):
627
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
628
+ more detail.
629
+ return_dict (`bool`, *optional*):
630
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
631
+ """
632
+ PARALLELIZE_DOCSTRING = r"""
633
+ This is an experimental feature and is a subject to change at a moment's notice.
634
+
635
+ Uses a device map to distribute attention modules of the model across several devices. If no device map is given,
636
+ it will evenly distribute blocks across all devices.
637
+
638
+ Args:
639
+ device_map (`Dict[int, list]`, optional, defaults to None):
640
+ A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
641
+ automatically mapped to the first device (for esoteric reasons). That means that the first device should
642
+ have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
643
+ following number of attention modules:
644
+
645
+ - gpt2: 12
646
+ - gpt2-medium: 24
647
+ - gpt2-large: 36
648
+ - gpt2-xl: 48
649
+
650
+ Example:
651
+
652
+ ```python
653
+ # Here is an example of a device map on a machine with 4 GPUs using gpt2-xl, which has a total of 48 attention modules:
654
+ model = GPT2LMHeadModel.from_pretrained("gpt2-xl")
655
+ device_map = {
656
+ 0: [0, 1, 2, 3, 4, 5, 6, 7, 8],
657
+ 1: [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21],
658
+ 2: [22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
659
+ 3: [35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47],
660
+ }
661
+ model.parallelize(device_map)
662
+ ```
663
+ """
664
+ DEPARALLELIZE_DOCSTRING = r"""
665
+ Moves the model to cpu from a model parallel state.
666
+
667
+ Example:
668
+
669
+ ```python
670
+ # On a 4 GPU machine with gpt2-large:
671
+ model = GPT2LMHeadModel.from_pretrained("gpt2-large")
672
+ device_map = {
673
+ 0: [0, 1, 2, 3, 4, 5, 6, 7],
674
+ 1: [8, 9, 10, 11, 12, 13, 14, 15],
675
+ 2: [16, 17, 18, 19, 20, 21, 22, 23],
676
+ 3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35],
677
+ }
678
+ model.parallelize(device_map) # Splits the model across several devices
679
+ model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
680
+ ```
681
+ """
682
+
683
+
684
+ @add_start_docstrings(
685
+ "The bare BTLM Model transformer outputting raw hidden-states without any specific head on top.",
686
+ BTLM_START_DOCSTRING,
687
+ )
688
+ class BTLMModel(BTLMPreTrainedModel):
689
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
690
+ _keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias"]
691
+
692
+ def __init__(self, config):
693
+ super().__init__(config)
694
+
695
+ self.embed_dim = config.hidden_size
696
+
697
+ self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
698
+ self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim) if config.position_embedding_type != "alibi" else None
699
+ self.embeddings_scale = config.embeddings_scale
700
+
701
+ self.drop = nn.Dropout(config.embd_pdrop)
702
+ self.h = nn.ModuleList([BTLMBlock(config, layer_idx=i) for i in range(config.num_hidden_layers)])
703
+ self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
704
+
705
+ self.relative_pe = AlibiPositionEmbeddingLayer(config.num_attention_heads) if config.position_embedding_type == "alibi" else None
706
+
707
+ # Model parallel
708
+ self.model_parallel = False
709
+ self.device_map = None
710
+ self.gradient_checkpointing = False
711
+
712
+ # Initialize weights and apply final processing
713
+ self.post_init()
714
+
715
+ @add_start_docstrings(PARALLELIZE_DOCSTRING)
716
+ def parallelize(self, device_map=None):
717
+ # Check validity of device_map
718
+ warnings.warn(
719
+ "`BTLMModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your"
720
+ " model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
721
+ " `device_map` but it needs to be a dictionary module_name to device, so for instance {'h.0': 0, 'h.1': 1,"
722
+ " ...}",
723
+ FutureWarning,
724
+ )
725
+ self.device_map = (
726
+ get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map
727
+ )
728
+ assert_device_map(self.device_map, len(self.h))
729
+ self.model_parallel = True
730
+ self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
731
+ self.last_device = "cuda:" + str(max(self.device_map.keys()))
732
+ self.wte = self.wte.to(self.first_device)
733
+ if self.wpe is not None:
734
+ self.wpe = self.wpe.to(self.first_device)
735
+ # Load onto devices
736
+ for k, v in self.device_map.items():
737
+ for block in v:
738
+ cuda_device = "cuda:" + str(k)
739
+ self.h[block] = self.h[block].to(cuda_device)
740
+ # ln_f to last
741
+ self.ln_f = self.ln_f.to(self.last_device)
742
+
743
+ @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
744
+ def deparallelize(self):
745
+ warnings.warn(
746
+ "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
747
+ FutureWarning,
748
+ )
749
+ self.model_parallel = False
750
+ self.device_map = None
751
+ self.first_device = "cpu"
752
+ self.last_device = "cpu"
753
+ self.wte = self.wte.to("cpu")
754
+ if self.wpe is not None:
755
+ self.wpe = self.wpe.to("cpu")
756
+ for index in range(len(self.h)):
757
+ self.h[index] = self.h[index].to("cpu")
758
+ self.ln_f = self.ln_f.to("cpu")
759
+ torch.cuda.empty_cache()
760
+
761
+ def get_input_embeddings(self):
762
+ return self.wte
763
+
764
+ def set_input_embeddings(self, new_embeddings):
765
+ self.wte = new_embeddings
766
+
767
+ def _prune_heads(self, heads_to_prune):
768
+ """
769
+ Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
770
+ """
771
+ for layer, heads in heads_to_prune.items():
772
+ self.h[layer].attn.prune_heads(heads)
773
+
774
+ @add_start_docstrings_to_model_forward(BTLM_INPUTS_DOCSTRING)
775
+ @add_code_sample_docstrings(
776
+ checkpoint=_CHECKPOINT_FOR_DOC,
777
+ output_type=BaseModelOutputWithPastAndCrossAttentions,
778
+ config_class=_CONFIG_FOR_DOC,
779
+ )
780
+ def forward(
781
+ self,
782
+ input_ids: Optional[torch.LongTensor] = None,
783
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
784
+ attention_mask: Optional[torch.FloatTensor] = None,
785
+ token_type_ids: Optional[torch.LongTensor] = None,
786
+ position_ids: Optional[torch.LongTensor] = None,
787
+ head_mask: Optional[torch.FloatTensor] = None,
788
+ inputs_embeds: Optional[torch.FloatTensor] = None,
789
+ encoder_hidden_states: Optional[torch.Tensor] = None,
790
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
791
+ use_cache: Optional[bool] = None,
792
+ output_attentions: Optional[bool] = None,
793
+ output_hidden_states: Optional[bool] = None,
794
+ return_dict: Optional[bool] = None,
795
+ ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
796
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
797
+ output_hidden_states = (
798
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
799
+ )
800
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
801
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
802
+
803
+ if input_ids is not None and inputs_embeds is not None:
804
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
805
+ elif input_ids is not None:
806
+ input_shape = input_ids.size()
807
+ input_ids = input_ids.view(-1, input_shape[-1])
808
+ batch_size = input_ids.shape[0]
809
+ elif inputs_embeds is not None:
810
+ input_shape = inputs_embeds.size()[:-1]
811
+ batch_size = inputs_embeds.shape[0]
812
+ else:
813
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
814
+
815
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
816
+
817
+ if token_type_ids is not None:
818
+ token_type_ids = token_type_ids.view(-1, input_shape[-1])
819
+ if position_ids is not None:
820
+ position_ids = position_ids.view(-1, input_shape[-1])
821
+
822
+ if past_key_values is None:
823
+ past_length = 0
824
+ past_key_values = tuple([None] * len(self.h))
825
+ else:
826
+ past_length = past_key_values[0][0].size(-2)
827
+ if position_ids is None:
828
+ position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
829
+ position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
830
+
831
+ # BTLMAttention mask.
832
+ if attention_mask is not None:
833
+ if batch_size <= 0:
834
+ raise ValueError("batch_size has to be defined and > 0")
835
+ attention_mask = attention_mask.view(batch_size, -1)
836
+ # We create a 3D attention mask from a 2D tensor mask.
837
+ # Sizes are [batch_size, 1, 1, to_seq_length]
838
+ # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
839
+ # this attention mask is more simple than the triangular masking of causal attention
840
+ # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
841
+ attention_mask = attention_mask[:, None, None, :]
842
+
843
+ # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
844
+ # masked positions, this operation will create a tensor which is 0.0 for
845
+ # positions we want to attend and the dtype's smallest value for masked positions.
846
+ # Since we are adding it to the raw scores before the softmax, this is
847
+ # effectively the same as removing these entirely.
848
+ attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility
849
+ attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
850
+
851
+ # If a 2D or 3D attention mask is provided for the cross-attention
852
+ # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
853
+ if self.config.add_cross_attention and encoder_hidden_states is not None:
854
+ encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
855
+ encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
856
+ if encoder_attention_mask is None:
857
+ encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
858
+ encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
859
+ else:
860
+ encoder_attention_mask = None
861
+
862
+ # Prepare head mask if needed
863
+ # 1.0 in head_mask indicate we keep the head
864
+ # attention_probs has shape bsz x n_heads x N x N
865
+ # head_mask has shape n_layer x batch x n_heads x N x N
866
+ head_mask = self.get_head_mask(head_mask, self.config.n_layer)
867
+
868
+ if inputs_embeds is None:
869
+ inputs_embeds = self.wte(input_ids)
870
+ if self.wpe is not None:
871
+ position_embeds = self.wpe(position_ids)
872
+ hidden_states = inputs_embeds + position_embeds
873
+ else:
874
+ hidden_states = inputs_embeds
875
+ hidden_states *= torch.tensor(
876
+ float(self.embeddings_scale), dtype=hidden_states.dtype, device=hidden_states.device
877
+ )
878
+
879
+ if token_type_ids is not None:
880
+ token_type_embeds = self.wte(token_type_ids)
881
+ hidden_states = hidden_states + token_type_embeds
882
+
883
+ hidden_states = self.drop(hidden_states)
884
+
885
+ if self.relative_pe is not None:
886
+ length = input_ids.shape[1]
887
+ cached_kv_length = 0
888
+ cached_kv = past_key_values[0]
889
+ if cached_kv is not None:
890
+ cached_kv_length = cached_kv[0].shape[-2]
891
+ position_bias = self.relative_pe(length, length, cached_kv_length)
892
+ else:
893
+ position_bias = None
894
+
895
+ output_shape = input_shape + (hidden_states.size(-1),)
896
+
897
+ if self.gradient_checkpointing and self.training:
898
+ if use_cache:
899
+ logger.warning_once(
900
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
901
+ )
902
+ use_cache = False
903
+
904
+ presents = () if use_cache else None
905
+ all_self_attentions = () if output_attentions else None
906
+ all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
907
+ all_hidden_states = () if output_hidden_states else None
908
+ for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
909
+ # Model parallel
910
+ if self.model_parallel:
911
+ torch.cuda.set_device(hidden_states.device)
912
+ # Ensure layer_past is on same device as hidden_states (might not be correct)
913
+ if layer_past is not None:
914
+ layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
915
+ # Ensure that attention_mask is always on the same device as hidden_states
916
+ if attention_mask is not None:
917
+ attention_mask = attention_mask.to(hidden_states.device)
918
+ if isinstance(head_mask, torch.Tensor):
919
+ head_mask = head_mask.to(hidden_states.device)
920
+ if output_hidden_states:
921
+ all_hidden_states = all_hidden_states + (hidden_states,)
922
+
923
+ if self.gradient_checkpointing and self.training:
924
+
925
+ def create_custom_forward(module):
926
+ def custom_forward(*inputs):
927
+ # None for past_key_value
928
+ return module(*inputs, use_cache, output_attentions)
929
+
930
+ return custom_forward
931
+
932
+ outputs = torch.utils.checkpoint.checkpoint(
933
+ create_custom_forward(block),
934
+ hidden_states,
935
+ None,
936
+ attention_mask,
937
+ head_mask[i],
938
+ encoder_hidden_states,
939
+ encoder_attention_mask,
940
+ )
941
+ else:
942
+ outputs = block(
943
+ hidden_states,
944
+ layer_past=layer_past,
945
+ attention_mask=attention_mask,
946
+ head_mask=head_mask[i],
947
+ encoder_hidden_states=encoder_hidden_states,
948
+ encoder_attention_mask=encoder_attention_mask,
949
+ use_cache=use_cache,
950
+ output_attentions=output_attentions,
951
+ position_bias=position_bias,
952
+ )
953
+
954
+ hidden_states = outputs[0]
955
+ if use_cache is True:
956
+ presents = presents + (outputs[1],)
957
+
958
+ if output_attentions:
959
+ all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
960
+ if self.config.add_cross_attention:
961
+ all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)
962
+
963
+ # Model Parallel: If it's the last layer for that device, put things on the next device
964
+ if self.model_parallel:
965
+ for k, v in self.device_map.items():
966
+ if i == v[-1] and "cuda:" + str(k) != self.last_device:
967
+ hidden_states = hidden_states.to("cuda:" + str(k + 1))
968
+
969
+ hidden_states = self.ln_f(hidden_states)
970
+
971
+ hidden_states = hidden_states.view(output_shape)
972
+ # Add last hidden state
973
+ if output_hidden_states:
974
+ all_hidden_states = all_hidden_states + (hidden_states,)
975
+
976
+ if not return_dict:
977
+ return tuple(
978
+ v
979
+ for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions]
980
+ if v is not None
981
+ )
982
+
983
+ return BaseModelOutputWithPastAndCrossAttentions(
984
+ last_hidden_state=hidden_states,
985
+ past_key_values=presents,
986
+ hidden_states=all_hidden_states,
987
+ attentions=all_self_attentions,
988
+ cross_attentions=all_cross_attentions,
989
+ )
990
+
991
+
992
+ @add_start_docstrings(
993
+ """
994
+ The BTLM Model transformer with a language modeling head on top (linear layer with weights tied to the input
995
+ embeddings).
996
+ """,
997
+ BTLM_START_DOCSTRING,
998
+ )
999
+ class BTLMLMHeadModel(BTLMPreTrainedModel):
1000
+ _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
1001
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias"]
1002
+
1003
+ def __init__(self, config):
1004
+ super().__init__(config)
1005
+ self.transformer = BTLMModel(config)
1006
+ self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
1007
+ self.output_logits_scale = config.width_scale
1008
+
1009
+ # Model parallel
1010
+ self.model_parallel = False
1011
+ self.device_map = None
1012
+
1013
+ # Initialize weights and apply final processing
1014
+ self.post_init()
1015
+
1016
+ @add_start_docstrings(PARALLELIZE_DOCSTRING)
1017
+ def parallelize(self, device_map=None):
1018
+ warnings.warn(
1019
+ "`BTLMLMHeadModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load"
1020
+ " your model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
1021
+ " `device_map` but it needs to be a dictionary module_name to device, so for instance {'transformer.h.0':"
1022
+ " 0, 'transformer.h.1': 1, ...}",
1023
+ FutureWarning,
1024
+ )
1025
+ self.device_map = (
1026
+ get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
1027
+ if device_map is None
1028
+ else device_map
1029
+ )
1030
+ assert_device_map(self.device_map, len(self.transformer.h))
1031
+ self.transformer.parallelize(self.device_map)
1032
+ self.lm_head = self.lm_head.to(self.transformer.first_device)
1033
+ self.model_parallel = True
1034
+
1035
+ @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
1036
+ def deparallelize(self):
1037
+ warnings.warn(
1038
+ "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
1039
+ FutureWarning,
1040
+ )
1041
+ self.transformer.deparallelize()
1042
+ self.transformer = self.transformer.to("cpu")
1043
+ self.lm_head = self.lm_head.to("cpu")
1044
+ self.model_parallel = False
1045
+ torch.cuda.empty_cache()
1046
+
1047
+ def get_output_embeddings(self):
1048
+ return self.lm_head
1049
+
1050
+ def set_output_embeddings(self, new_embeddings):
1051
+ self.lm_head = new_embeddings
1052
+
1053
+ def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
1054
+ token_type_ids = kwargs.get("token_type_ids", None)
1055
+ # only last token for inputs_ids if past is defined in kwargs
1056
+ if past_key_values:
1057
+ input_ids = input_ids[:, -1].unsqueeze(-1)
1058
+ if token_type_ids is not None:
1059
+ token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
1060
+
1061
+ attention_mask = kwargs.get("attention_mask", None)
1062
+ position_ids = kwargs.get("position_ids", None)
1063
+
1064
+ if attention_mask is not None and position_ids is None:
1065
+ # create position_ids on the fly for batch generation
1066
+ position_ids = attention_mask.long().cumsum(-1) - 1
1067
+ position_ids.masked_fill_(attention_mask == 0, 1)
1068
+ if past_key_values:
1069
+ position_ids = position_ids[:, -1].unsqueeze(-1)
1070
+ else:
1071
+ position_ids = None
1072
+
1073
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1074
+ if inputs_embeds is not None and past_key_values is None:
1075
+ model_inputs = {"inputs_embeds": inputs_embeds}
1076
+ else:
1077
+ model_inputs = {"input_ids": input_ids}
1078
+
1079
+ model_inputs.update(
1080
+ {
1081
+ "past_key_values": past_key_values,
1082
+ "use_cache": kwargs.get("use_cache"),
1083
+ "position_ids": position_ids,
1084
+ "attention_mask": attention_mask,
1085
+ "token_type_ids": token_type_ids,
1086
+ }
1087
+ )
1088
+ return model_inputs
1089
+
1090
+ @add_start_docstrings_to_model_forward(BTLM_INPUTS_DOCSTRING)
1091
+ @add_code_sample_docstrings(
1092
+ checkpoint=_CHECKPOINT_FOR_DOC,
1093
+ output_type=CausalLMOutputWithCrossAttentions,
1094
+ config_class=_CONFIG_FOR_DOC,
1095
+ )
1096
+ def forward(
1097
+ self,
1098
+ input_ids: Optional[torch.LongTensor] = None,
1099
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1100
+ attention_mask: Optional[torch.FloatTensor] = None,
1101
+ token_type_ids: Optional[torch.LongTensor] = None,
1102
+ position_ids: Optional[torch.LongTensor] = None,
1103
+ head_mask: Optional[torch.FloatTensor] = None,
1104
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1105
+ encoder_hidden_states: Optional[torch.Tensor] = None,
1106
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
1107
+ labels: Optional[torch.LongTensor] = None,
1108
+ use_cache: Optional[bool] = None,
1109
+ output_attentions: Optional[bool] = None,
1110
+ output_hidden_states: Optional[bool] = None,
1111
+ return_dict: Optional[bool] = None,
1112
+ ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
1113
+ r"""
1114
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1115
+ Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
1116
+ `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
1117
+ are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
1118
+ """
1119
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1120
+
1121
+ transformer_outputs = self.transformer(
1122
+ input_ids,
1123
+ past_key_values=past_key_values,
1124
+ attention_mask=attention_mask,
1125
+ token_type_ids=token_type_ids,
1126
+ position_ids=position_ids,
1127
+ head_mask=head_mask,
1128
+ inputs_embeds=inputs_embeds,
1129
+ encoder_hidden_states=encoder_hidden_states,
1130
+ encoder_attention_mask=encoder_attention_mask,
1131
+ use_cache=use_cache,
1132
+ output_attentions=output_attentions,
1133
+ output_hidden_states=output_hidden_states,
1134
+ return_dict=return_dict,
1135
+ )
1136
+ hidden_states = transformer_outputs[0]
1137
+
1138
+ # Set device for model parallelism
1139
+ if self.model_parallel:
1140
+ torch.cuda.set_device(self.transformer.first_device)
1141
+ hidden_states = hidden_states.to(self.lm_head.weight.device)
1142
+
1143
+ lm_logits = self.lm_head(hidden_states)
1144
+ lm_logits *= torch.tensor(
1145
+ float(self.output_logits_scale), dtype=lm_logits.dtype, device=lm_logits.device
1146
+ )
1147
+
1148
+ loss = None
1149
+ if labels is not None:
1150
+ # move labels to correct device to enable model parallelism
1151
+ labels = labels.to(lm_logits.device)
1152
+ # Shift so that tokens < n predict n
1153
+ shift_logits = lm_logits[..., :-1, :].contiguous()
1154
+ shift_labels = labels[..., 1:].contiguous()
1155
+ # Flatten the tokens
1156
+ loss_fct = CrossEntropyLoss()
1157
+ loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
1158
+
1159
+ if not return_dict:
1160
+ output = (lm_logits,) + transformer_outputs[1:]
1161
+ return ((loss,) + output) if loss is not None else output
1162
+
1163
+ return CausalLMOutputWithCrossAttentions(
1164
+ loss=loss,
1165
+ logits=lm_logits,
1166
+ past_key_values=transformer_outputs.past_key_values,
1167
+ hidden_states=transformer_outputs.hidden_states,
1168
+ attentions=transformer_outputs.attentions,
1169
+ cross_attentions=transformer_outputs.cross_attentions,
1170
+ )
1171
+
1172
+ @staticmethod
1173
+ def _reorder_cache(
1174
+ past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
1175
+ ) -> Tuple[Tuple[torch.Tensor]]:
1176
+ """
1177
+ This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
1178
+ [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
1179
+ beam_idx at every generation step.
1180
+ """
1181
+ return tuple(
1182
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
1183
+ for layer_past in past_key_values
1184
+ )
1185
+
1186
+
1187
+ @add_start_docstrings(
1188
+ """
1189
+ The BTLM Model transformer with a sequence classification head on top (linear layer).
1190
+
1191
+ [`BTLMForSequenceClassification`] uses the last token in order to do the classification, as other causal models
1192
+ (e.g. GPT-1) do.
1193
+
1194
+ Since it does classification on the last token, it requires to know the position of the last token. If a
1195
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
1196
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
1197
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
1198
+ each row of the batch).
1199
+ """,
1200
+ BTLM_START_DOCSTRING,
1201
+ )
1202
+ class BTLMForSequenceClassification(BTLMPreTrainedModel):
1203
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
1204
+ _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head.weight"]
1205
+
1206
+ def __init__(self, config):
1207
+ super().__init__(config)
1208
+ self.num_labels = config.num_labels
1209
+ self.transformer = BTLMModel(config)
1210
+ self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)
1211
+ self.output_logits_scale = config.width_scale
1212
+
1213
+ # Model parallel
1214
+ self.model_parallel = False
1215
+ self.device_map = None
1216
+
1217
+ # Initialize weights and apply final processing
1218
+ self.post_init()
1219
+
1220
+ @add_start_docstrings_to_model_forward(BTLM_INPUTS_DOCSTRING)
1221
+ @add_code_sample_docstrings(
1222
+ checkpoint="microsoft/DialogRPT-updown",
1223
+ output_type=SequenceClassifierOutputWithPast,
1224
+ config_class=_CONFIG_FOR_DOC,
1225
+ )
1226
+ def forward(
1227
+ self,
1228
+ input_ids: Optional[torch.LongTensor] = None,
1229
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1230
+ attention_mask: Optional[torch.FloatTensor] = None,
1231
+ token_type_ids: Optional[torch.LongTensor] = None,
1232
+ position_ids: Optional[torch.LongTensor] = None,
1233
+ head_mask: Optional[torch.FloatTensor] = None,
1234
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1235
+ labels: Optional[torch.LongTensor] = None,
1236
+ use_cache: Optional[bool] = None,
1237
+ output_attentions: Optional[bool] = None,
1238
+ output_hidden_states: Optional[bool] = None,
1239
+ return_dict: Optional[bool] = None,
1240
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1241
+ r"""
1242
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1243
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1244
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1245
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1246
+ """
1247
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1248
+
1249
+ transformer_outputs = self.transformer(
1250
+ input_ids,
1251
+ past_key_values=past_key_values,
1252
+ attention_mask=attention_mask,
1253
+ token_type_ids=token_type_ids,
1254
+ position_ids=position_ids,
1255
+ head_mask=head_mask,
1256
+ inputs_embeds=inputs_embeds,
1257
+ use_cache=use_cache,
1258
+ output_attentions=output_attentions,
1259
+ output_hidden_states=output_hidden_states,
1260
+ return_dict=return_dict,
1261
+ )
1262
+ hidden_states = transformer_outputs[0]
1263
+ logits = self.score(hidden_states)
1264
+ logits *= torch.tensor(
1265
+ float(self.output_logits_scale), dtype=logits.dtype, device=logits.device
1266
+ )
1267
+
1268
+ if input_ids is not None:
1269
+ batch_size, sequence_length = input_ids.shape[:2]
1270
+ else:
1271
+ batch_size, sequence_length = inputs_embeds.shape[:2]
1272
+
1273
+ assert (
1274
+ self.config.pad_token_id is not None or batch_size == 1
1275
+ ), "Cannot handle batch sizes > 1 if no padding token is defined."
1276
+ if self.config.pad_token_id is None:
1277
+ sequence_lengths = -1
1278
+ else:
1279
+ if input_ids is not None:
1280
+ sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
1281
+ else:
1282
+ sequence_lengths = -1
1283
+ logger.warning(
1284
+ f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
1285
+ "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
1286
+ )
1287
+
1288
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1289
+
1290
+ loss = None
1291
+ if labels is not None:
1292
+ if self.config.problem_type is None:
1293
+ if self.num_labels == 1:
1294
+ self.config.problem_type = "regression"
1295
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1296
+ self.config.problem_type = "single_label_classification"
1297
+ else:
1298
+ self.config.problem_type = "multi_label_classification"
1299
+
1300
+ if self.config.problem_type == "regression":
1301
+ loss_fct = MSELoss()
1302
+ if self.num_labels == 1:
1303
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1304
+ else:
1305
+ loss = loss_fct(pooled_logits, labels)
1306
+ elif self.config.problem_type == "single_label_classification":
1307
+ loss_fct = CrossEntropyLoss()
1308
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1309
+ elif self.config.problem_type == "multi_label_classification":
1310
+ loss_fct = BCEWithLogitsLoss()
1311
+ loss = loss_fct(pooled_logits, labels)
1312
+ if not return_dict:
1313
+ output = (pooled_logits,) + transformer_outputs[1:]
1314
+ return ((loss,) + output) if loss is not None else output
1315
+
1316
+ return SequenceClassifierOutputWithPast(
1317
+ loss=loss,
1318
+ logits=pooled_logits,
1319
+ past_key_values=transformer_outputs.past_key_values,
1320
+ hidden_states=transformer_outputs.hidden_states,
1321
+ attentions=transformer_outputs.attentions,
1322
+ )
1323
+
1324
+
1325
+ @add_start_docstrings(
1326
+ """
1327
+ BTLM Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
1328
+ Named-Entity-Recognition (NER) tasks.
1329
+ """,
1330
+ BTLM_START_DOCSTRING,
1331
+ )
1332
+ class BTLMForTokenClassification(BTLMPreTrainedModel):
1333
+ def __init__(self, config):
1334
+ super().__init__(config)
1335
+ self.num_labels = config.num_labels
1336
+
1337
+ self.transformer = BTLMModel(config)
1338
+ if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
1339
+ classifier_dropout = config.classifier_dropout
1340
+ elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
1341
+ classifier_dropout = config.hidden_dropout
1342
+ else:
1343
+ classifier_dropout = 0.1
1344
+ self.dropout = nn.Dropout(classifier_dropout)
1345
+ self.classifier = nn.Linear(config.hidden_size, config.num_labels)
1346
+ self.output_logits_scale = config.width_scale
1347
+
1348
+ # Model parallel
1349
+ self.model_parallel = False
1350
+ self.device_map = None
1351
+
1352
+ # Initialize weights and apply final processing
1353
+ self.post_init()
1354
+
1355
+ @add_start_docstrings_to_model_forward(BTLM_INPUTS_DOCSTRING)
1356
+ # fmt: off
1357
+ @add_code_sample_docstrings(
1358
+ checkpoint="brad1141/gpt2-finetuned-comp2",
1359
+ output_type=TokenClassifierOutput,
1360
+ config_class=_CONFIG_FOR_DOC,
1361
+ expected_loss=0.25,
1362
+ expected_output=["Lead", "Lead", "Lead", "Position", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead"],
1363
+ )
1364
+ # fmt: on
1365
+ def forward(
1366
+ self,
1367
+ input_ids: Optional[torch.LongTensor] = None,
1368
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1369
+ attention_mask: Optional[torch.FloatTensor] = None,
1370
+ token_type_ids: Optional[torch.LongTensor] = None,
1371
+ position_ids: Optional[torch.LongTensor] = None,
1372
+ head_mask: Optional[torch.FloatTensor] = None,
1373
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1374
+ labels: Optional[torch.LongTensor] = None,
1375
+ use_cache: Optional[bool] = None,
1376
+ output_attentions: Optional[bool] = None,
1377
+ output_hidden_states: Optional[bool] = None,
1378
+ return_dict: Optional[bool] = None,
1379
+ ) -> Union[Tuple, TokenClassifierOutput]:
1380
+ r"""
1381
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1382
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1383
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1384
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1385
+ """
1386
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1387
+
1388
+ transformer_outputs = self.transformer(
1389
+ input_ids,
1390
+ past_key_values=past_key_values,
1391
+ attention_mask=attention_mask,
1392
+ token_type_ids=token_type_ids,
1393
+ position_ids=position_ids,
1394
+ head_mask=head_mask,
1395
+ inputs_embeds=inputs_embeds,
1396
+ use_cache=use_cache,
1397
+ output_attentions=output_attentions,
1398
+ output_hidden_states=output_hidden_states,
1399
+ return_dict=return_dict,
1400
+ )
1401
+
1402
+ hidden_states = transformer_outputs[0]
1403
+ hidden_states = self.dropout(hidden_states)
1404
+ logits = self.classifier(hidden_states)
1405
+ logits *= torch.tensor(
1406
+ float(self.output_logits_scale), dtype=logits.dtype, device=logits.device
1407
+ )
1408
+
1409
+ loss = None
1410
+ if labels is not None:
1411
+ labels = labels.to(logits.device)
1412
+ loss_fct = CrossEntropyLoss()
1413
+ loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
1414
+
1415
+ if not return_dict:
1416
+ output = (logits,) + transformer_outputs[2:]
1417
+ return ((loss,) + output) if loss is not None else output
1418
+
1419
+ return TokenClassifierOutput(
1420
+ loss=loss,
1421
+ logits=logits,
1422
+ hidden_states=transformer_outputs.hidden_states,
1423
+ attentions=transformer_outputs.attentions,
1424
+ )
1425
+
1426
+
1427
+ @add_start_docstrings(
1428
+ """
1429
+ The BTLM Model transformer with a span classification head on top for extractive question-answering tasks like
1430
+ SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
1431
+ """,
1432
+ BTLM_START_DOCSTRING,
1433
+ )
1434
+ class BTLMForQuestionAnswering(BTLMPreTrainedModel):
1435
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
1436
+ _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias", r"lm_head.weight"]
1437
+
1438
+ def __init__(self, config):
1439
+ super().__init__(config)
1440
+ self.num_labels = config.num_labels
1441
+ self.transformer = BTLMModel(config)
1442
+ self.qa_outputs = nn.Linear(config.hidden_size, 2)
1443
+ self.output_logits_scale = config.width_scale
1444
+
1445
+ # Model parallel
1446
+ self.model_parallel = False
1447
+ self.device_map = None
1448
+ self.gradient_checkpointing = False
1449
+
1450
+ # Initialize weights and apply final processing
1451
+ self.post_init()
1452
+
1453
+ def forward(
1454
+ self,
1455
+ input_ids: Optional[torch.LongTensor] = None,
1456
+ attention_mask: Optional[torch.FloatTensor] = None,
1457
+ token_type_ids: Optional[torch.LongTensor] = None,
1458
+ position_ids: Optional[torch.LongTensor] = None,
1459
+ head_mask: Optional[torch.FloatTensor] = None,
1460
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1461
+ start_positions: Optional[torch.LongTensor] = None,
1462
+ end_positions: Optional[torch.LongTensor] = None,
1463
+ output_attentions: Optional[bool] = None,
1464
+ output_hidden_states: Optional[bool] = None,
1465
+ return_dict: Optional[bool] = None,
1466
+ ) -> Union[Tuple, QuestionAnsweringModelOutput]:
1467
+ r"""
1468
+ start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1469
+ Labels for position (index) of the start of the labelled span for computing the token classification loss.
1470
+ Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
1471
+ are not taken into account for computing the loss.
1472
+ end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1473
+ Labels for position (index) of the end of the labelled span for computing the token classification loss.
1474
+ Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
1475
+ are not taken into account for computing the loss.
1476
+ """
1477
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1478
+
1479
+ outputs = self.transformer(
1480
+ input_ids,
1481
+ attention_mask=attention_mask,
1482
+ token_type_ids=token_type_ids,
1483
+ position_ids=position_ids,
1484
+ head_mask=head_mask,
1485
+ inputs_embeds=inputs_embeds,
1486
+ output_attentions=output_attentions,
1487
+ output_hidden_states=output_hidden_states,
1488
+ return_dict=return_dict,
1489
+ )
1490
+
1491
+ sequence_output = outputs[0]
1492
+
1493
+ logits = self.qa_outputs(sequence_output)
1494
+ logits *= torch.tensor(
1495
+ float(self.output_logits_scale), dtype=logits.dtype, device=logits.device
1496
+ )
1497
+ start_logits, end_logits = logits.split(1, dim=-1)
1498
+ start_logits = start_logits.squeeze(-1).contiguous()
1499
+ end_logits = end_logits.squeeze(-1).contiguous()
1500
+
1501
+ total_loss = None
1502
+ if start_positions is not None and end_positions is not None:
1503
+ # If we are on multi-GPU, split add a dimension
1504
+ if len(start_positions.size()) > 1:
1505
+ start_positions = start_positions.squeeze(-1).to(start_logits.device)
1506
+ if len(end_positions.size()) > 1:
1507
+ end_positions = end_positions.squeeze(-1).to(end_logits.device)
1508
+ # sometimes the start/end positions are outside our model inputs, we ignore these terms
1509
+ ignored_index = start_logits.size(1)
1510
+ start_positions = start_positions.clamp(0, ignored_index)
1511
+ end_positions = end_positions.clamp(0, ignored_index)
1512
+
1513
+ loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
1514
+ start_loss = loss_fct(start_logits, start_positions)
1515
+ end_loss = loss_fct(end_logits, end_positions)
1516
+ total_loss = (start_loss + end_loss) / 2
1517
+
1518
+ if not return_dict:
1519
+ output = (start_logits, end_logits) + outputs[2:]
1520
+ return ((total_loss,) + output) if total_loss is not None else output
1521
+
1522
+ return QuestionAnsweringModelOutput(
1523
+ loss=total_loss,
1524
+ start_logits=start_logits,
1525
+ end_logits=end_logits,
1526
+ hidden_states=outputs.hidden_states,
1527
+ attentions=outputs.attentions,
1528
+ )
params_train_to_hf.json ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "attn_pdrop": 0.0,
3
+ "scale_attn_weights": true,
4
+ "resid_pdrop": 0.0,
5
+ "embeddings_scale": 14.6,
6
+ "n_inner": 5461,
7
+ "n_embd": 2048,
8
+ "layer_norm_epsilon": 1e-05,
9
+ "n_positions": 2048,
10
+ "activation_function": "swiglu",
11
+ "n_head": 16,
12
+ "n_layer": 24,
13
+ "width_scale": 0.2775,
14
+ "position_embedding_type": "alibi",
15
+ "scale_qk_dot_by_d": true,
16
+ "tie_word_embeddings": true,
17
+ "vocab_size": 50277,
18
+ "embd_pdrop": 0.0
19
+ }
pytorch_model-00001-of-00001.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b7b5a35de43b7d7be36805ec8607e55c06e0534c15f044913132cf5685d6d886
3
+ size 5759005258
pytorch_model.bin.index.json ADDED
@@ -0,0 +1,396 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 5758877792
4
+ },
5
+ "weight_map": {
6
+ "transformer.wte.weight": "pytorch_model-00001-of-00001.bin",
7
+ "transformer.relative_pe.slopes": "pytorch_model-00001-of-00001.bin",
8
+ "transformer.h.0.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
9
+ "transformer.h.0.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
10
+ "transformer.h.0.attn.bias": "pytorch_model-00001-of-00001.bin",
11
+ "transformer.h.0.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
12
+ "transformer.h.0.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
13
+ "transformer.h.0.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
14
+ "transformer.h.0.ln_1.weight": "pytorch_model-00001-of-00001.bin",
15
+ "transformer.h.0.ln_1.bias": "pytorch_model-00001-of-00001.bin",
16
+ "transformer.h.0.ln_2.weight": "pytorch_model-00001-of-00001.bin",
17
+ "transformer.h.0.ln_2.bias": "pytorch_model-00001-of-00001.bin",
18
+ "transformer.h.0.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
19
+ "transformer.h.0.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
20
+ "transformer.h.0.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
21
+ "transformer.h.0.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
22
+ "transformer.h.0.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
23
+ "transformer.h.0.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
24
+ "transformer.h.1.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
25
+ "transformer.h.1.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
26
+ "transformer.h.1.attn.bias": "pytorch_model-00001-of-00001.bin",
27
+ "transformer.h.1.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
28
+ "transformer.h.1.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
29
+ "transformer.h.1.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
30
+ "transformer.h.1.ln_1.weight": "pytorch_model-00001-of-00001.bin",
31
+ "transformer.h.1.ln_1.bias": "pytorch_model-00001-of-00001.bin",
32
+ "transformer.h.1.ln_2.weight": "pytorch_model-00001-of-00001.bin",
33
+ "transformer.h.1.ln_2.bias": "pytorch_model-00001-of-00001.bin",
34
+ "transformer.h.1.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
35
+ "transformer.h.1.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
36
+ "transformer.h.1.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
37
+ "transformer.h.1.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
38
+ "transformer.h.1.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
39
+ "transformer.h.1.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
40
+ "transformer.h.2.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
41
+ "transformer.h.2.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
42
+ "transformer.h.2.attn.bias": "pytorch_model-00001-of-00001.bin",
43
+ "transformer.h.2.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
44
+ "transformer.h.2.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
45
+ "transformer.h.2.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
46
+ "transformer.h.2.ln_1.weight": "pytorch_model-00001-of-00001.bin",
47
+ "transformer.h.2.ln_1.bias": "pytorch_model-00001-of-00001.bin",
48
+ "transformer.h.2.ln_2.weight": "pytorch_model-00001-of-00001.bin",
49
+ "transformer.h.2.ln_2.bias": "pytorch_model-00001-of-00001.bin",
50
+ "transformer.h.2.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
51
+ "transformer.h.2.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
52
+ "transformer.h.2.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
53
+ "transformer.h.2.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
54
+ "transformer.h.2.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
55
+ "transformer.h.2.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
56
+ "transformer.h.3.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
57
+ "transformer.h.3.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
58
+ "transformer.h.3.attn.bias": "pytorch_model-00001-of-00001.bin",
59
+ "transformer.h.3.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
60
+ "transformer.h.3.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
61
+ "transformer.h.3.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
62
+ "transformer.h.3.ln_1.weight": "pytorch_model-00001-of-00001.bin",
63
+ "transformer.h.3.ln_1.bias": "pytorch_model-00001-of-00001.bin",
64
+ "transformer.h.3.ln_2.weight": "pytorch_model-00001-of-00001.bin",
65
+ "transformer.h.3.ln_2.bias": "pytorch_model-00001-of-00001.bin",
66
+ "transformer.h.3.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
67
+ "transformer.h.3.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
68
+ "transformer.h.3.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
69
+ "transformer.h.3.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
70
+ "transformer.h.3.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
71
+ "transformer.h.3.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
72
+ "transformer.h.4.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
73
+ "transformer.h.4.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
74
+ "transformer.h.4.attn.bias": "pytorch_model-00001-of-00001.bin",
75
+ "transformer.h.4.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
76
+ "transformer.h.4.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
77
+ "transformer.h.4.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
78
+ "transformer.h.4.ln_1.weight": "pytorch_model-00001-of-00001.bin",
79
+ "transformer.h.4.ln_1.bias": "pytorch_model-00001-of-00001.bin",
80
+ "transformer.h.4.ln_2.weight": "pytorch_model-00001-of-00001.bin",
81
+ "transformer.h.4.ln_2.bias": "pytorch_model-00001-of-00001.bin",
82
+ "transformer.h.4.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
83
+ "transformer.h.4.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
84
+ "transformer.h.4.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
85
+ "transformer.h.4.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
86
+ "transformer.h.4.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
87
+ "transformer.h.4.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
88
+ "transformer.h.5.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
89
+ "transformer.h.5.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
90
+ "transformer.h.5.attn.bias": "pytorch_model-00001-of-00001.bin",
91
+ "transformer.h.5.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
92
+ "transformer.h.5.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
93
+ "transformer.h.5.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
94
+ "transformer.h.5.ln_1.weight": "pytorch_model-00001-of-00001.bin",
95
+ "transformer.h.5.ln_1.bias": "pytorch_model-00001-of-00001.bin",
96
+ "transformer.h.5.ln_2.weight": "pytorch_model-00001-of-00001.bin",
97
+ "transformer.h.5.ln_2.bias": "pytorch_model-00001-of-00001.bin",
98
+ "transformer.h.5.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
99
+ "transformer.h.5.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
100
+ "transformer.h.5.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
101
+ "transformer.h.5.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
102
+ "transformer.h.5.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
103
+ "transformer.h.5.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
104
+ "transformer.h.6.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
105
+ "transformer.h.6.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
106
+ "transformer.h.6.attn.bias": "pytorch_model-00001-of-00001.bin",
107
+ "transformer.h.6.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
108
+ "transformer.h.6.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
109
+ "transformer.h.6.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
110
+ "transformer.h.6.ln_1.weight": "pytorch_model-00001-of-00001.bin",
111
+ "transformer.h.6.ln_1.bias": "pytorch_model-00001-of-00001.bin",
112
+ "transformer.h.6.ln_2.weight": "pytorch_model-00001-of-00001.bin",
113
+ "transformer.h.6.ln_2.bias": "pytorch_model-00001-of-00001.bin",
114
+ "transformer.h.6.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
115
+ "transformer.h.6.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
116
+ "transformer.h.6.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
117
+ "transformer.h.6.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
118
+ "transformer.h.6.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
119
+ "transformer.h.6.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
120
+ "transformer.h.7.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
121
+ "transformer.h.7.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
122
+ "transformer.h.7.attn.bias": "pytorch_model-00001-of-00001.bin",
123
+ "transformer.h.7.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
124
+ "transformer.h.7.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
125
+ "transformer.h.7.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
126
+ "transformer.h.7.ln_1.weight": "pytorch_model-00001-of-00001.bin",
127
+ "transformer.h.7.ln_1.bias": "pytorch_model-00001-of-00001.bin",
128
+ "transformer.h.7.ln_2.weight": "pytorch_model-00001-of-00001.bin",
129
+ "transformer.h.7.ln_2.bias": "pytorch_model-00001-of-00001.bin",
130
+ "transformer.h.7.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
131
+ "transformer.h.7.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
132
+ "transformer.h.7.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
133
+ "transformer.h.7.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
134
+ "transformer.h.7.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
135
+ "transformer.h.7.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
136
+ "transformer.h.8.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
137
+ "transformer.h.8.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
138
+ "transformer.h.8.attn.bias": "pytorch_model-00001-of-00001.bin",
139
+ "transformer.h.8.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
140
+ "transformer.h.8.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
141
+ "transformer.h.8.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
142
+ "transformer.h.8.ln_1.weight": "pytorch_model-00001-of-00001.bin",
143
+ "transformer.h.8.ln_1.bias": "pytorch_model-00001-of-00001.bin",
144
+ "transformer.h.8.ln_2.weight": "pytorch_model-00001-of-00001.bin",
145
+ "transformer.h.8.ln_2.bias": "pytorch_model-00001-of-00001.bin",
146
+ "transformer.h.8.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
147
+ "transformer.h.8.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
148
+ "transformer.h.8.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
149
+ "transformer.h.8.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
150
+ "transformer.h.8.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
151
+ "transformer.h.8.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
152
+ "transformer.h.9.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
153
+ "transformer.h.9.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
154
+ "transformer.h.9.attn.bias": "pytorch_model-00001-of-00001.bin",
155
+ "transformer.h.9.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
156
+ "transformer.h.9.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
157
+ "transformer.h.9.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
158
+ "transformer.h.9.ln_1.weight": "pytorch_model-00001-of-00001.bin",
159
+ "transformer.h.9.ln_1.bias": "pytorch_model-00001-of-00001.bin",
160
+ "transformer.h.9.ln_2.weight": "pytorch_model-00001-of-00001.bin",
161
+ "transformer.h.9.ln_2.bias": "pytorch_model-00001-of-00001.bin",
162
+ "transformer.h.9.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
163
+ "transformer.h.9.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
164
+ "transformer.h.9.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
165
+ "transformer.h.9.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
166
+ "transformer.h.9.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
167
+ "transformer.h.9.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
168
+ "transformer.h.10.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
169
+ "transformer.h.10.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
170
+ "transformer.h.10.attn.bias": "pytorch_model-00001-of-00001.bin",
171
+ "transformer.h.10.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
172
+ "transformer.h.10.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
173
+ "transformer.h.10.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
174
+ "transformer.h.10.ln_1.weight": "pytorch_model-00001-of-00001.bin",
175
+ "transformer.h.10.ln_1.bias": "pytorch_model-00001-of-00001.bin",
176
+ "transformer.h.10.ln_2.weight": "pytorch_model-00001-of-00001.bin",
177
+ "transformer.h.10.ln_2.bias": "pytorch_model-00001-of-00001.bin",
178
+ "transformer.h.10.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
179
+ "transformer.h.10.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
180
+ "transformer.h.10.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
181
+ "transformer.h.10.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
182
+ "transformer.h.10.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
183
+ "transformer.h.10.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
184
+ "transformer.h.11.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
185
+ "transformer.h.11.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
186
+ "transformer.h.11.attn.bias": "pytorch_model-00001-of-00001.bin",
187
+ "transformer.h.11.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
188
+ "transformer.h.11.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
189
+ "transformer.h.11.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
190
+ "transformer.h.11.ln_1.weight": "pytorch_model-00001-of-00001.bin",
191
+ "transformer.h.11.ln_1.bias": "pytorch_model-00001-of-00001.bin",
192
+ "transformer.h.11.ln_2.weight": "pytorch_model-00001-of-00001.bin",
193
+ "transformer.h.11.ln_2.bias": "pytorch_model-00001-of-00001.bin",
194
+ "transformer.h.11.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
195
+ "transformer.h.11.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
196
+ "transformer.h.11.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
197
+ "transformer.h.11.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
198
+ "transformer.h.11.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
199
+ "transformer.h.11.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
200
+ "transformer.h.12.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
201
+ "transformer.h.12.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
202
+ "transformer.h.12.attn.bias": "pytorch_model-00001-of-00001.bin",
203
+ "transformer.h.12.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
204
+ "transformer.h.12.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
205
+ "transformer.h.12.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
206
+ "transformer.h.12.ln_1.weight": "pytorch_model-00001-of-00001.bin",
207
+ "transformer.h.12.ln_1.bias": "pytorch_model-00001-of-00001.bin",
208
+ "transformer.h.12.ln_2.weight": "pytorch_model-00001-of-00001.bin",
209
+ "transformer.h.12.ln_2.bias": "pytorch_model-00001-of-00001.bin",
210
+ "transformer.h.12.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
211
+ "transformer.h.12.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
212
+ "transformer.h.12.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
213
+ "transformer.h.12.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
214
+ "transformer.h.12.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
215
+ "transformer.h.12.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
216
+ "transformer.h.13.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
217
+ "transformer.h.13.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
218
+ "transformer.h.13.attn.bias": "pytorch_model-00001-of-00001.bin",
219
+ "transformer.h.13.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
220
+ "transformer.h.13.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
221
+ "transformer.h.13.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
222
+ "transformer.h.13.ln_1.weight": "pytorch_model-00001-of-00001.bin",
223
+ "transformer.h.13.ln_1.bias": "pytorch_model-00001-of-00001.bin",
224
+ "transformer.h.13.ln_2.weight": "pytorch_model-00001-of-00001.bin",
225
+ "transformer.h.13.ln_2.bias": "pytorch_model-00001-of-00001.bin",
226
+ "transformer.h.13.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
227
+ "transformer.h.13.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
228
+ "transformer.h.13.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
229
+ "transformer.h.13.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
230
+ "transformer.h.13.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
231
+ "transformer.h.13.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
232
+ "transformer.h.14.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
233
+ "transformer.h.14.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
234
+ "transformer.h.14.attn.bias": "pytorch_model-00001-of-00001.bin",
235
+ "transformer.h.14.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
236
+ "transformer.h.14.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
237
+ "transformer.h.14.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
238
+ "transformer.h.14.ln_1.weight": "pytorch_model-00001-of-00001.bin",
239
+ "transformer.h.14.ln_1.bias": "pytorch_model-00001-of-00001.bin",
240
+ "transformer.h.14.ln_2.weight": "pytorch_model-00001-of-00001.bin",
241
+ "transformer.h.14.ln_2.bias": "pytorch_model-00001-of-00001.bin",
242
+ "transformer.h.14.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
243
+ "transformer.h.14.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
244
+ "transformer.h.14.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
245
+ "transformer.h.14.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
246
+ "transformer.h.14.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
247
+ "transformer.h.14.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
248
+ "transformer.h.15.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
249
+ "transformer.h.15.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
250
+ "transformer.h.15.attn.bias": "pytorch_model-00001-of-00001.bin",
251
+ "transformer.h.15.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
252
+ "transformer.h.15.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
253
+ "transformer.h.15.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
254
+ "transformer.h.15.ln_1.weight": "pytorch_model-00001-of-00001.bin",
255
+ "transformer.h.15.ln_1.bias": "pytorch_model-00001-of-00001.bin",
256
+ "transformer.h.15.ln_2.weight": "pytorch_model-00001-of-00001.bin",
257
+ "transformer.h.15.ln_2.bias": "pytorch_model-00001-of-00001.bin",
258
+ "transformer.h.15.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
259
+ "transformer.h.15.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
260
+ "transformer.h.15.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
261
+ "transformer.h.15.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
262
+ "transformer.h.15.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
263
+ "transformer.h.15.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
264
+ "transformer.h.16.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
265
+ "transformer.h.16.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
266
+ "transformer.h.16.attn.bias": "pytorch_model-00001-of-00001.bin",
267
+ "transformer.h.16.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
268
+ "transformer.h.16.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
269
+ "transformer.h.16.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
270
+ "transformer.h.16.ln_1.weight": "pytorch_model-00001-of-00001.bin",
271
+ "transformer.h.16.ln_1.bias": "pytorch_model-00001-of-00001.bin",
272
+ "transformer.h.16.ln_2.weight": "pytorch_model-00001-of-00001.bin",
273
+ "transformer.h.16.ln_2.bias": "pytorch_model-00001-of-00001.bin",
274
+ "transformer.h.16.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
275
+ "transformer.h.16.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
276
+ "transformer.h.16.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
277
+ "transformer.h.16.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
278
+ "transformer.h.16.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
279
+ "transformer.h.16.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
280
+ "transformer.h.17.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
281
+ "transformer.h.17.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
282
+ "transformer.h.17.attn.bias": "pytorch_model-00001-of-00001.bin",
283
+ "transformer.h.17.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
284
+ "transformer.h.17.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
285
+ "transformer.h.17.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
286
+ "transformer.h.17.ln_1.weight": "pytorch_model-00001-of-00001.bin",
287
+ "transformer.h.17.ln_1.bias": "pytorch_model-00001-of-00001.bin",
288
+ "transformer.h.17.ln_2.weight": "pytorch_model-00001-of-00001.bin",
289
+ "transformer.h.17.ln_2.bias": "pytorch_model-00001-of-00001.bin",
290
+ "transformer.h.17.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
291
+ "transformer.h.17.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
292
+ "transformer.h.17.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
293
+ "transformer.h.17.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
294
+ "transformer.h.17.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
295
+ "transformer.h.17.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
296
+ "transformer.h.18.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
297
+ "transformer.h.18.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
298
+ "transformer.h.18.attn.bias": "pytorch_model-00001-of-00001.bin",
299
+ "transformer.h.18.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
300
+ "transformer.h.18.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
301
+ "transformer.h.18.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
302
+ "transformer.h.18.ln_1.weight": "pytorch_model-00001-of-00001.bin",
303
+ "transformer.h.18.ln_1.bias": "pytorch_model-00001-of-00001.bin",
304
+ "transformer.h.18.ln_2.weight": "pytorch_model-00001-of-00001.bin",
305
+ "transformer.h.18.ln_2.bias": "pytorch_model-00001-of-00001.bin",
306
+ "transformer.h.18.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
307
+ "transformer.h.18.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
308
+ "transformer.h.18.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
309
+ "transformer.h.18.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
310
+ "transformer.h.18.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
311
+ "transformer.h.18.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
312
+ "transformer.h.19.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
313
+ "transformer.h.19.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
314
+ "transformer.h.19.attn.bias": "pytorch_model-00001-of-00001.bin",
315
+ "transformer.h.19.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
316
+ "transformer.h.19.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
317
+ "transformer.h.19.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
318
+ "transformer.h.19.ln_1.weight": "pytorch_model-00001-of-00001.bin",
319
+ "transformer.h.19.ln_1.bias": "pytorch_model-00001-of-00001.bin",
320
+ "transformer.h.19.ln_2.weight": "pytorch_model-00001-of-00001.bin",
321
+ "transformer.h.19.ln_2.bias": "pytorch_model-00001-of-00001.bin",
322
+ "transformer.h.19.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
323
+ "transformer.h.19.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
324
+ "transformer.h.19.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
325
+ "transformer.h.19.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
326
+ "transformer.h.19.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
327
+ "transformer.h.19.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
328
+ "transformer.h.20.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
329
+ "transformer.h.20.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
330
+ "transformer.h.20.attn.bias": "pytorch_model-00001-of-00001.bin",
331
+ "transformer.h.20.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
332
+ "transformer.h.20.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
333
+ "transformer.h.20.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
334
+ "transformer.h.20.ln_1.weight": "pytorch_model-00001-of-00001.bin",
335
+ "transformer.h.20.ln_1.bias": "pytorch_model-00001-of-00001.bin",
336
+ "transformer.h.20.ln_2.weight": "pytorch_model-00001-of-00001.bin",
337
+ "transformer.h.20.ln_2.bias": "pytorch_model-00001-of-00001.bin",
338
+ "transformer.h.20.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
339
+ "transformer.h.20.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
340
+ "transformer.h.20.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
341
+ "transformer.h.20.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
342
+ "transformer.h.20.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
343
+ "transformer.h.20.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
344
+ "transformer.h.21.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
345
+ "transformer.h.21.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
346
+ "transformer.h.21.attn.bias": "pytorch_model-00001-of-00001.bin",
347
+ "transformer.h.21.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
348
+ "transformer.h.21.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
349
+ "transformer.h.21.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
350
+ "transformer.h.21.ln_1.weight": "pytorch_model-00001-of-00001.bin",
351
+ "transformer.h.21.ln_1.bias": "pytorch_model-00001-of-00001.bin",
352
+ "transformer.h.21.ln_2.weight": "pytorch_model-00001-of-00001.bin",
353
+ "transformer.h.21.ln_2.bias": "pytorch_model-00001-of-00001.bin",
354
+ "transformer.h.21.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
355
+ "transformer.h.21.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
356
+ "transformer.h.21.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
357
+ "transformer.h.21.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
358
+ "transformer.h.21.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
359
+ "transformer.h.21.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
360
+ "transformer.h.22.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
361
+ "transformer.h.22.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
362
+ "transformer.h.22.attn.bias": "pytorch_model-00001-of-00001.bin",
363
+ "transformer.h.22.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
364
+ "transformer.h.22.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
365
+ "transformer.h.22.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
366
+ "transformer.h.22.ln_1.weight": "pytorch_model-00001-of-00001.bin",
367
+ "transformer.h.22.ln_1.bias": "pytorch_model-00001-of-00001.bin",
368
+ "transformer.h.22.ln_2.weight": "pytorch_model-00001-of-00001.bin",
369
+ "transformer.h.22.ln_2.bias": "pytorch_model-00001-of-00001.bin",
370
+ "transformer.h.22.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
371
+ "transformer.h.22.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
372
+ "transformer.h.22.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
373
+ "transformer.h.22.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
374
+ "transformer.h.22.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
375
+ "transformer.h.22.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
376
+ "transformer.h.23.attn.c_attn.weight": "pytorch_model-00001-of-00001.bin",
377
+ "transformer.h.23.attn.c_attn.bias": "pytorch_model-00001-of-00001.bin",
378
+ "transformer.h.23.attn.bias": "pytorch_model-00001-of-00001.bin",
379
+ "transformer.h.23.attn.masked_bias": "pytorch_model-00001-of-00001.bin",
380
+ "transformer.h.23.attn.c_proj.weight": "pytorch_model-00001-of-00001.bin",
381
+ "transformer.h.23.attn.c_proj.bias": "pytorch_model-00001-of-00001.bin",
382
+ "transformer.h.23.ln_1.weight": "pytorch_model-00001-of-00001.bin",
383
+ "transformer.h.23.ln_1.bias": "pytorch_model-00001-of-00001.bin",
384
+ "transformer.h.23.ln_2.weight": "pytorch_model-00001-of-00001.bin",
385
+ "transformer.h.23.ln_2.bias": "pytorch_model-00001-of-00001.bin",
386
+ "transformer.h.23.mlp.c_fc.weight": "pytorch_model-00001-of-00001.bin",
387
+ "transformer.h.23.mlp.c_fc.bias": "pytorch_model-00001-of-00001.bin",
388
+ "transformer.h.23.mlp.c_fc2.weight": "pytorch_model-00001-of-00001.bin",
389
+ "transformer.h.23.mlp.c_fc2.bias": "pytorch_model-00001-of-00001.bin",
390
+ "transformer.h.23.mlp.c_proj.weight": "pytorch_model-00001-of-00001.bin",
391
+ "transformer.h.23.mlp.c_proj.bias": "pytorch_model-00001-of-00001.bin",
392
+ "transformer.ln_f.weight": "pytorch_model-00001-of-00001.bin",
393
+ "transformer.ln_f.bias": "pytorch_model-00001-of-00001.bin",
394
+ "lm_head.weight": "pytorch_model-00001-of-00001.bin"
395
+ }
396
+ }
register_btlm.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoConfig, AutoModel, AutoModelForCausalLM, AutoTokenizer
2
+ from transformers.tokenization_utils_fast import PreTrainedTokenizerFast
3
+ from configuration_btlm import BTLMConfig
4
+ from modeling_btlm import BTLMModel, BTLMLMHeadModel
5
+
6
+ AutoConfig.register("btlm", BTLMConfig)
7
+ AutoModel.register(BTLMConfig, BTLMModel)
8
+ AutoModelForCausalLM.register(BTLMConfig, BTLMLMHeadModel)
9
+ AutoTokenizer.register(BTLMConfig, fast_tokenizer_class=PreTrainedTokenizerFast)
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff