Guanzheng commited on
Commit
91f64f9
1 Parent(s): f062d4b

Upload 4 files

Browse files
Files changed (4) hide show
  1. clex_layer.py +141 -0
  2. config.py +33 -0
  3. configuration_clex.py +148 -0
  4. modeling_llama.py +1008 -0
clex_layer.py ADDED
@@ -0,0 +1,141 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ from torchdiffeq import odeint
4
+
5
+
6
+
7
+ import math
8
+
9
+ class ODELinear(nn.Module):
10
+ def __init__(
11
+ self,
12
+ dim: int,
13
+ factor,
14
+ **kwargs
15
+ ):
16
+ super().__init__()
17
+ self.ode_up_proj = nn.Parameter(torch.empty(dim//2, factor*dim).to(torch.float32))
18
+ self.ode_down_proj = nn.Parameter(torch.empty(factor*dim, dim//2).to(torch.float32))
19
+ self.dim = dim
20
+ self.act = torch.nn.SiLU()
21
+ self.reset_parameters()
22
+
23
+ def reset_parameters(self):
24
+ nn.init.kaiming_uniform_(self.ode_up_proj, a=math.sqrt(5))
25
+ nn.init.zeros_(self.ode_down_proj)
26
+
27
+ def get_time_embedding(self, t, base=10000, device='cuda', dtype=torch.float32):
28
+ if t < 1:
29
+ alpha = 1
30
+ else:
31
+ alpha = 2*t-1
32
+ ntk_base = base * alpha ** (self.dim / (self.dim-2))
33
+ ntk_inv_freq = 1.0 / (ntk_base ** (torch.arange(0, self.dim, 2, dtype=torch.float32).to(device) / self.dim))
34
+ index = torch.arange(0, self.dim, 2, dtype=torch.float32).to(device)
35
+ delta_ntk_freq = -2*index/(self.dim-2) * 1 / (base ** (index/self.dim) * (alpha ** (index/(self.dim-2) + 1)))
36
+ return delta_ntk_freq.to(device, dtype=dtype), ntk_inv_freq.to(device, dtype=dtype)
37
+
38
+ def forward(self, t, x: torch.Tensor):
39
+ delta_time, time = self.get_time_embedding(t, device=x.device, dtype=x.dtype)
40
+ x = x + torch.log(time)
41
+ time_embed = delta_time / time
42
+ delta_inv_freq = self.act(x @ self.ode_up_proj.float()) @ self.ode_down_proj.float() + time_embed
43
+ return delta_inv_freq
44
+
45
+
46
+
47
+ class LlamaCLEXScalingRotaryEmbedding(nn.Module):
48
+
49
+ def __init__(self, dim, max_position_embeddings=2048, rope_scaling=None, base=10000, device=None) -> None:
50
+ super().__init__()
51
+
52
+ self.max_t = rope_scaling["max_factor"]
53
+ self.dim = dim
54
+ self.max_position_embeddings = max_position_embeddings
55
+ self.base = base
56
+ inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
57
+ self.register_buffer("inv_freq", inv_freq)
58
+
59
+ self.proj_func = ODELinear(dim, rope_scaling["param_factor"])
60
+ self.rope_cached = None
61
+ self.max_t_cached = 0
62
+ self.freq_cached = None
63
+ self.time_dt = 0.01
64
+ self.ode_args = {
65
+ "method": "rk4",
66
+ "options": {"step_size": self.time_dt},
67
+ }
68
+
69
+ def sample_random_times(self, max_t, device):
70
+ return torch.randint(2, max_t, (1,), dtype = torch.long, device=device)
71
+
72
+ def get_random_position_ids(self, n=2048, max=8192):
73
+ positions = torch.randperm(max)[:n].sort().values
74
+ # positions = positions.to(device=device)
75
+ return positions
76
+
77
+
78
+ def get_continuous_freq(self, time_grid, ex_positions, device):
79
+ solution = odeint(
80
+ self.proj_func, torch.log(self.inv_freq.to(device, dtype=torch.float32)), time_grid, **self.ode_args
81
+ )
82
+ if time_grid.size(0) == 2:
83
+ training
84
+ scale_inv_freq = torch.exp(solution[1])
85
+ # print(time_grid[1].tolist(), torch.sum(scale_inv_freq).tolist(), torch.sum(self.proj_func.ode_down_proj).tolist())
86
+ freqs = torch.outer(ex_positions.float().squeeze(), scale_inv_freq)
87
+ else:
88
+ scale_inv_freq = torch.exp(solution)
89
+ # freqs = torch.einsum('i, kl -> kil', ex_positions, scale_inv_freq)
90
+ return scale_inv_freq
91
+ embed = torch.cat((freqs,freqs), dim=-1)
92
+ return embed
93
+
94
+
95
+
96
+ def forward(self, device, dtype, seq_len, do_train=False):
97
+ device = self.proj_func.ode_up_proj.device
98
+ scale_factor = seq_len // self.max_position_embeddings
99
+ if do_train:
100
+ t_val = self.sample_random_times(self.max_t+1, device)[0]
101
+ import math
102
+ sampled_position_ids = self.get_random_position_ids(n=seq_len-2, max=seq_len*t_val-2).float()
103
+ ex_positions = torch.cat([
104
+ torch.tensor([0]),
105
+ (sampled_position_ids + 1) / scale_factor,
106
+ torch.tensor([seq_len*t_val//scale_factor-1])]
107
+ ).to(device, dtype=torch.float32)
108
+ else:
109
+ t_val = scale_factor if seq_len%self.max_position_embeddings == 0.0 else scale_factor + 1
110
+ t_val = t_val if t_val <= self.max_t else self.max_t
111
+ ex_positions = torch.arange(0, self.max_position_embeddings * t_val, dtype=torch.float32).to(device)
112
+
113
+
114
+
115
+ if t_val == 1.0:
116
+ scale_inv_freq = self.inv_freq.to(device)
117
+ freqs = torch.outer(ex_positions.float().squeeze(), scale_inv_freq)
118
+ embed = torch.cat((freqs,freqs), dim=-1)
119
+ cos, sin = embed.cos()[None, None, :, :], embed.sin()[None, None, :, :]
120
+ elif do_train:
121
+ time_grid = torch.tensor([1.0, t_val]).float().to(device)
122
+ embed = self.get_continuous_freq(time_grid, ex_positions, device)
123
+ cos, sin = embed.cos()[None, None, :, :], embed.sin()[None, None, :, :]
124
+ else:
125
+ if t_val > self.max_t_cached:
126
+ if self.freq_cached is None:
127
+ time_grid = torch.arange(1.0, self.max_t, dtype=torch.float32).to(device)
128
+ self.freq_cached = self.get_continuous_freq(time_grid, ex_positions, device)
129
+ scale_inv_freq = self.freq_cached[int(t_val-1.0)]
130
+ freqs = torch.outer(ex_positions.float().squeeze(), scale_inv_freq)
131
+ embed = torch.cat((freqs,freqs), dim=-1)
132
+ self.rope_cached = torch.cat((embed.cos()[None, None, None, :, :], embed.sin()[None, None, None, :, :]), dim=0)
133
+ self.max_t_cached = t_val
134
+ cos, sin = self.rope_cached
135
+
136
+ return torch.cat(
137
+ (cos[None, :, :, :seq_len, ...].to(dtype=dtype),
138
+ sin[None, :, :, :seq_len, ...].to(dtype=dtype)),
139
+ dim=0
140
+ )
141
+
config.py ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "LlamaForCausalLM"
4
+ ],
5
+ "auto_map": {
6
+ "AutoConfig": "configuration_clex.CLEXLlamaConfig",
7
+ "AutoModelForCausalLM": "modeling_llama.LlamaForCausalLM"
8
+ },
9
+ "bos_token_id": 1,
10
+ "eos_token_id": 2,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 4096,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 11008,
15
+ "max_position_embeddings": 4096,
16
+ "model_type": "llama",
17
+ "num_attention_heads": 32,
18
+ "num_hidden_layers": 32,
19
+ "num_key_value_heads": 32,
20
+ "pad_token_id": 0,
21
+ "pretraining_tp": 1,
22
+ "rms_norm_eps": 1e-05,
23
+ "tie_word_embeddings": false,
24
+ "use_cache": true,
25
+ "vocab_size": 32000,
26
+ "log_scale": false,
27
+ "use_flashattn": true,
28
+ "rope_scaling": {
29
+ "type": "clex",
30
+ "max_factor": 16,
31
+ "param_factor": 1,
32
+ }
33
+ }
configuration_clex.py ADDED
@@ -0,0 +1,148 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+ """ LLaMA model configuration"""
21
+
22
+ from transformers.configuration_utils import PretrainedConfig
23
+ from transformers.utils import logging
24
+ from transformers import LlamaConfig
25
+
26
+
27
+ logger = logging.get_logger(__name__)
28
+
29
+ LLAMA_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
30
+
31
+
32
+ class CLEXLlamaConfig(LlamaConfig):
33
+ r"""
34
+ This is the configuration class to store the configuration of a [`LlamaModel`]. It is used to instantiate an LLaMA
35
+ model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
36
+ defaults will yield a similar configuration to that of the LLaMA-7B.
37
+
38
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
39
+ documentation from [`PretrainedConfig`] for more information.
40
+
41
+
42
+ Args:
43
+ vocab_size (`int`, *optional*, defaults to 32000):
44
+ Vocabulary size of the LLaMA model. Defines the number of different tokens that can be represented by the
45
+ `inputs_ids` passed when calling [`LlamaModel`]
46
+ hidden_size (`int`, *optional*, defaults to 4096):
47
+ Dimension of the hidden representations.
48
+ intermediate_size (`int`, *optional*, defaults to 11008):
49
+ Dimension of the MLP representations.
50
+ num_hidden_layers (`int`, *optional*, defaults to 32):
51
+ Number of hidden layers in the Transformer encoder.
52
+ num_attention_heads (`int`, *optional*, defaults to 32):
53
+ Number of attention heads for each attention layer in the Transformer encoder.
54
+ num_key_value_heads (`int`, *optional*):
55
+ This is the number of key_value heads that should be used to implement Grouped Query Attention. If
56
+ `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
57
+ `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
58
+ converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
59
+ by meanpooling all the original heads within that group. For more details checkout [this
60
+ paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
61
+ `num_attention_heads`.
62
+ pretraining_tp (`int`, *optional*, defaults to `1`):
63
+ Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
64
+ document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
65
+ necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
66
+ issue](https://github.com/pytorch/pytorch/issues/76232).
67
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
68
+ The non-linear activation function (function or string) in the decoder.
69
+ max_position_embeddings (`int`, *optional*, defaults to 2048):
70
+ The maximum sequence length that this model might ever be used with. Typically set this to something large
71
+ just in case (e.g., 512 or 1024 or 2048).
72
+ initializer_range (`float`, *optional*, defaults to 0.02):
73
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
74
+ rms_norm_eps (`float`, *optional*, defaults to 1e-12):
75
+ The epsilon used by the rms normalization layers.
76
+ use_cache (`bool`, *optional*, defaults to `True`):
77
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
78
+ relevant if `config.is_decoder=True`.
79
+ tie_word_embeddings(`bool`, *optional*, defaults to `False`):
80
+ Whether to tie weight embeddings
81
+ rope_scaling (`Dict`, *optional*):
82
+ Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports three scaling
83
+ strategies: linear and dynamic. Their scaling factor must be an float greater than 1. The expected format
84
+ is `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
85
+ `max_position_embeddings` to the expected new maximum. See the following thread for more information on how
86
+ these scaling strategies behave:
87
+ https://www.reddit.com/r/LocalLLaMA/comments/14mrgpr/dynamically_scaled_rope_further_increases/. This is an
88
+ experimental feature, subject to breaking API changes in future versions.
89
+
90
+ Example:
91
+
92
+ ```python
93
+ >>> from transformers import LlamaModel, LlamaConfig
94
+
95
+ >>> # Initializing a LLaMA llama-7b style configuration
96
+ >>> configuration = LlamaConfig()
97
+
98
+ >>> # Initializing a model from the llama-7b style configuration
99
+ >>> model = LlamaModel(configuration)
100
+
101
+ >>> # Accessing the model configuration
102
+ >>> configuration = model.config
103
+ ```"""
104
+ model_type = "llama"
105
+ keys_to_ignore_at_inference = ["past_key_values"]
106
+
107
+ def __init__(
108
+ self,
109
+ rope_scaling=None,
110
+ use_flashattn=True,
111
+ log_scale=True,
112
+ **kwargs,
113
+ ):
114
+ super().__init__(
115
+ **kwargs,
116
+ )
117
+ self.use_flashattn = use_flashattn
118
+ self.log_scale = log_scale
119
+ self.rope_theta = 10000
120
+ self.max_position_embeddings = 4096
121
+ self.data_length = 4096
122
+ self.rope_scaling = rope_scaling
123
+ self._rope_scaling_validation()
124
+
125
+
126
+ def _rope_scaling_validation(self):
127
+ """
128
+ Validate the `rope_scaling` configuration.
129
+ """
130
+ if self.rope_scaling is None:
131
+ return
132
+
133
+ # if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:
134
+ # raise ValueError(
135
+ # "`rope_scaling` must be a dictionary with with two fields, `name` and `factor`, "
136
+ # f"got {self.rope_scaling}"
137
+ # )
138
+ rope_scaling_type = self.rope_scaling.get("type", None)
139
+ rope_scaling_max_factor = self.rope_scaling.get("max_factor", None)
140
+ rope_scaling_param_factor = self.rope_scaling.get("param_factor", None)
141
+ if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic", "clex"]:
142
+ raise ValueError(
143
+ f"`rope_scaling`'s name field must be one of ['linear', 'dynamic'], got {rope_scaling_type}"
144
+ )
145
+ # if rope_scaling_max_factor is None or not isinstance(rope_scaling_max_factor, float) or rope_scaling_max_factor <= 1.0:
146
+ # raise ValueError(f"`rope_scaling`'s factor field must be an float > 1, got {rope_scaling_max_factor}")
147
+ # if rope_scaling_param_factor is None or not isinstance(rope_scaling_param_factor, float) or rope_scaling_param_factor <= 1.0:
148
+ # raise ValueError(f"`rope_scaling`'s factor field must be an float > 1, got {rope_scaling_param_factor}")
modeling_llama.py ADDED
@@ -0,0 +1,1008 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+ """ PyTorch LLaMA model."""
21
+ import math
22
+ from typing import List, Optional, Tuple, Union
23
+
24
+ import torch
25
+ import torch.utils.checkpoint
26
+ from torch import nn
27
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
28
+
29
+ from transformers.activations import ACT2FN
30
+ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
31
+ from transformers.modeling_utils import PreTrainedModel
32
+ from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
33
+ from .configuration_clex import CLEXLlamaConfig
34
+ from .clex_layer import LlamaCLEXScalingRotaryEmbedding
35
+ from einops import rearrange
36
+ import importlib.metadata
37
+ import importlib.util
38
+
39
+
40
+ logger = logging.get_logger(__name__)
41
+
42
+ def _is_package_available(pkg_name: str, return_version: bool = False) -> Union[Tuple[bool, str], bool]:
43
+ # Check we're not importing a "pkg_name" directory somewhere but the actual library by trying to grab the version
44
+ package_exists = importlib.util.find_spec(pkg_name) is not None
45
+ package_version = "N/A"
46
+ if package_exists:
47
+ try:
48
+ package_version = importlib.metadata.version(pkg_name)
49
+ package_exists = True
50
+ except importlib.metadata.PackageNotFoundError:
51
+ package_exists = False
52
+ logger.info(f"Detected {pkg_name} version {package_version}")
53
+ if return_version:
54
+ return package_exists, package_version
55
+ else:
56
+ return package_exists
57
+
58
+ def is_flash_attn_available():
59
+ if not _is_package_available("torch", return_version=True):
60
+ return False
61
+
62
+ # Let's add an extra check to see if cuda is available
63
+ import torch
64
+
65
+ return _is_package_available("flash_attn") and torch.cuda.is_available()
66
+
67
+ if is_flash_attn_available():
68
+ from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func, flash_attn_qkvpacked_func, flash_attn_with_kvcache
69
+ # from flash_attn.flash_attn_interface import flash_attn_unpadded_qkvpacked_func
70
+ from flash_attn.bert_padding import unpad_input, pad_input
71
+
72
+
73
+
74
+
75
+ _CONFIG_FOR_DOC = "CLEXLlamaConfig"
76
+
77
+
78
+
79
+
80
+
81
+ # Copied from transformers.models.bart.modeling_bart._make_causal_mask
82
+ def _make_causal_mask(
83
+ input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
84
+ ):
85
+ """
86
+ Make causal mask used for bi-directional self-attention.
87
+ """
88
+ bsz, tgt_len = input_ids_shape
89
+ mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
90
+ mask_cond = torch.arange(mask.size(-1), device=device)
91
+ mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
92
+ mask = mask.to(dtype)
93
+
94
+ if past_key_values_length > 0:
95
+ mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
96
+ return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
97
+
98
+
99
+ # Copied from transformers.models.bart.modeling_bart._expand_mask
100
+ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
101
+ """
102
+ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
103
+ """
104
+ bsz, src_len = mask.size()
105
+ tgt_len = tgt_len if tgt_len is not None else src_len
106
+
107
+ expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
108
+
109
+ inverted_mask = 1.0 - expanded_mask
110
+
111
+ return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
112
+
113
+
114
+ class LlamaRMSNorm(nn.Module):
115
+ def __init__(self, hidden_size, eps=1e-6):
116
+ """
117
+ LlamaRMSNorm is equivalent to T5LayerNorm
118
+ """
119
+ super().__init__()
120
+ self.weight = nn.Parameter(torch.ones(hidden_size))
121
+ self.variance_epsilon = eps
122
+
123
+ def forward(self, hidden_states):
124
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
125
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
126
+
127
+ # convert into half-precision if necessary
128
+ if self.weight.dtype in [torch.float16, torch.bfloat16]:
129
+ hidden_states = hidden_states.to(self.weight.dtype)
130
+
131
+ return self.weight * hidden_states
132
+
133
+
134
+ class LlamaRotaryEmbedding(torch.nn.Module):
135
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
136
+ super().__init__()
137
+ inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
138
+ self.register_buffer("inv_freq", inv_freq)
139
+
140
+ # Build here to make `torch.jit.trace` work.
141
+ self.max_seq_len_cached = max_position_embeddings
142
+ t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
143
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
144
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
145
+ emb = torch.cat((freqs, freqs), dim=-1)
146
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
147
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
148
+
149
+ def forward(self, x, seq_len=None):
150
+ # x: [bs, num_attention_heads, seq_len, head_size]
151
+ # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
152
+ if seq_len > self.max_seq_len_cached:
153
+ self.max_seq_len_cached = seq_len
154
+ t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
155
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
156
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
157
+ emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
158
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
159
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
160
+ return (
161
+ self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
162
+ self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
163
+ )
164
+
165
+
166
+ def rotate_half(x):
167
+ """Rotates half the hidden dims of the input."""
168
+ x1 = x[..., : x.shape[-1] // 2]
169
+ x2 = x[..., x.shape[-1] // 2 :]
170
+ return torch.cat((-x2, x1), dim=-1)
171
+
172
+
173
+ def apply_rotary_pos_emb(q, k, cos, sin, q_len, position_ids):
174
+ # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
175
+ cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
176
+ sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
177
+ cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
178
+ sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
179
+ q_embed = (q * cos[:, :, -q_len:, :]) + (rotate_half(q) * sin[:, :, -q_len:, :])
180
+ k_embed = (k * cos) + (rotate_half(k) * sin)
181
+ return q_embed, k_embed
182
+
183
+
184
+ class LlamaMLP(nn.Module):
185
+ def __init__(
186
+ self,
187
+ hidden_size: int,
188
+ intermediate_size: int,
189
+ hidden_act: str,
190
+ ):
191
+ super().__init__()
192
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
193
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
194
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
195
+ self.act_fn = ACT2FN[hidden_act]
196
+
197
+ def forward(self, x):
198
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
199
+
200
+
201
+ class LlamaAttention(nn.Module):
202
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
203
+
204
+ def __init__(self, config: CLEXLlamaConfig):
205
+ super().__init__()
206
+ self.config = config
207
+ self.hidden_size = config.hidden_size
208
+ self.num_heads = config.num_attention_heads
209
+ self.head_dim = self.hidden_size // self.num_heads
210
+ self.max_position_embeddings = config.max_position_embeddings
211
+ self.log_scale = config.log_scale
212
+ if (self.head_dim * self.num_heads) != self.hidden_size:
213
+ raise ValueError(
214
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
215
+ f" and `num_heads`: {self.num_heads})."
216
+ )
217
+ self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
218
+ self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
219
+ self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
220
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
221
+ self.rotary_emb = LlamaRotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
222
+
223
+ def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
224
+ return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
225
+
226
+ def flash_attn_forward(
227
+ self,
228
+ qkv: torch.Tensor,
229
+ key_padding_mask: Optional[torch.Tensor] = None,
230
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
231
+ """Input shape: Batch x Time x Channel
232
+
233
+ attention_mask: [bsz, q_len]
234
+ """
235
+
236
+ bsz, q_len, *_ = qkv.size()
237
+
238
+ if key_padding_mask is None:
239
+ # qkv = rearrange(qkv, "b s ... -> (b s) ...")
240
+ max_s = q_len
241
+ cu_q_lens = torch.arange(
242
+ 0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device=qkv.device
243
+ )
244
+ output = flash_attn_qkvpacked_func(
245
+ qkv, 0.0, softmax_scale=None, causal=True
246
+ )
247
+ else:
248
+ nheads = qkv.shape[-2]
249
+ x = rearrange(qkv, "b s three h d -> b s (three h d)")
250
+ x_unpad, indices, cu_q_lens, max_s = unpad_input(x, key_padding_mask)
251
+ x_unpad = rearrange(
252
+ x_unpad, "nnz (three h d) -> nnz three h d", three=3, h=nheads
253
+ )
254
+ output_unpad = flash_attn_varlen_qkvpacked_func(
255
+ x_unpad, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
256
+ )
257
+ output = rearrange(
258
+ pad_input(
259
+ rearrange(output_unpad, "nnz h d -> nnz (h d)"), indices, bsz, q_len
260
+ ),
261
+ "b s (h d) -> b s h d",
262
+ h=nheads,
263
+ )
264
+ return self.o_proj(rearrange(output, "b s h d -> b s (h d)"))
265
+
266
+ def forward(
267
+ self,
268
+ hidden_states: torch.Tensor,
269
+ attention_mask: Optional[torch.Tensor] = None,
270
+ position_ids: Optional[torch.LongTensor] = None,
271
+ pack_cos_sin = None,
272
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
273
+ output_attentions: bool = False,
274
+ use_cache: bool = False,
275
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
276
+ bsz, q_len, _ = hidden_states.size()
277
+
278
+ query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
279
+ key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
280
+ value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
281
+
282
+ kv_seq_len = key_states.shape[-2]
283
+
284
+ if past_key_value is not None:
285
+ kv_seq_len += past_key_value[0].shape[-2]
286
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
287
+
288
+ if pack_cos_sin is not None:
289
+ cos, sin = pack_cos_sin.to(query_states.device)
290
+ else:
291
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
292
+ key_position_ids = torch.arange(kv_seq_len, dtype=torch.long, device=position_ids.device).unsqueeze(0).view(-1, kv_seq_len)
293
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, q_len, key_position_ids)
294
+
295
+ if past_key_value is not None:
296
+ # reuse k, v, self_attention
297
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
298
+
299
+ past_key_value = (key_states, value_states) if use_cache else None
300
+
301
+ use_flashatn = self.config.use_flashattn and is_flash_attn_available()
302
+
303
+ if self.log_scale:
304
+ log_n = torch.log(torch.tensor(kv_seq_len*1.0)).to(query_states.device, dtype=query_states.dtype) / \
305
+ torch.log(torch.tensor(self.config.max_position_embeddings)).to(query_states.device, dtype=query_states.dtype)
306
+ query_states = query_states * log_n
307
+ if query_states.shape[-2] == 1 or query_states.shape[-2] != key_states.shape[-2] or use_flashatn:
308
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
309
+
310
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
311
+ raise ValueError(
312
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
313
+ f" {attn_weights.size()}"
314
+ )
315
+
316
+ if attention_mask is not None:
317
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
318
+ raise ValueError(
319
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
320
+ )
321
+ attn_weights = attn_weights + attention_mask
322
+ attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
323
+
324
+ # upcast attention to fp32
325
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
326
+ attn_output = torch.matmul(attn_weights, value_states)
327
+
328
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
329
+ raise ValueError(
330
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
331
+ f" {attn_output.size()}"
332
+ )
333
+
334
+ attn_output = attn_output.transpose(1, 2)
335
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
336
+
337
+ attn_output = self.o_proj(attn_output)
338
+
339
+ if not output_attentions:
340
+ attn_weights = None
341
+
342
+ return attn_output, attn_weights, past_key_value
343
+ # use flash attention
344
+ elif past_key_value is not None:
345
+ output = flash_attn_with_kvcache(
346
+ query_states.transpose(1, 2),
347
+ key_states.transpose(1, 2),
348
+ value_states.transpose(1, 2),
349
+ cache_seqlens=kv_seq_len,
350
+ causal=True,
351
+ )
352
+ attn_output = self.o_proj(rearrange(output, "b s h d -> b s (h d)"))
353
+ else:
354
+ qkv = torch.stack(
355
+ [query_states, key_states, value_states], dim=2
356
+ ) # [bsz, nh, 3, q_len, hd]
357
+ qkv = qkv.transpose(1, 3) # [bsz, q_len, 3, nh, hd]
358
+ attn_output = self.flash_attn_forward(qkv)
359
+ return attn_output, None, past_key_value
360
+
361
+
362
+ class LlamaDecoderLayer(nn.Module):
363
+ def __init__(self, config: CLEXLlamaConfig):
364
+ super().__init__()
365
+ self.hidden_size = config.hidden_size
366
+ self.self_attn = LlamaAttention(config=config)
367
+ self.mlp = LlamaMLP(
368
+ hidden_size=self.hidden_size,
369
+ intermediate_size=config.intermediate_size,
370
+ hidden_act=config.hidden_act,
371
+ )
372
+ self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
373
+ self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
374
+
375
+ def forward(
376
+ self,
377
+ hidden_states: torch.Tensor,
378
+ attention_mask: Optional[torch.Tensor] = None,
379
+ position_ids: Optional[torch.LongTensor] = None,
380
+ pack_cos_sin=None,
381
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
382
+ output_attentions: Optional[bool] = False,
383
+ use_cache: Optional[bool] = False,
384
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
385
+ """
386
+ Args:
387
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
388
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
389
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
390
+ output_attentions (`bool`, *optional*):
391
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
392
+ returned tensors for more detail.
393
+ use_cache (`bool`, *optional*):
394
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
395
+ (see `past_key_values`).
396
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
397
+ """
398
+
399
+ residual = hidden_states
400
+
401
+ hidden_states = self.input_layernorm(hidden_states)
402
+
403
+ # Self Attention
404
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
405
+ hidden_states=hidden_states,
406
+ attention_mask=attention_mask,
407
+ position_ids=position_ids,
408
+ pack_cos_sin=pack_cos_sin,
409
+ past_key_value=past_key_value,
410
+ output_attentions=output_attentions,
411
+ use_cache=use_cache,
412
+ )
413
+ hidden_states = residual + hidden_states
414
+
415
+ # Fully Connected
416
+ residual = hidden_states
417
+ hidden_states = self.post_attention_layernorm(hidden_states)
418
+ hidden_states = self.mlp(hidden_states)
419
+ hidden_states = residual + hidden_states
420
+
421
+ outputs = (hidden_states,)
422
+
423
+ if output_attentions:
424
+ outputs += (self_attn_weights,)
425
+
426
+ if use_cache:
427
+ outputs += (present_key_value,)
428
+
429
+ return outputs
430
+
431
+
432
+ LLAMA_START_DOCSTRING = r"""
433
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
434
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
435
+ etc.)
436
+
437
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
438
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
439
+ and behavior.
440
+
441
+ Parameters:
442
+ config ([`CLEXLlamaConfig`]):
443
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
444
+ load the weights associated with the model, only the configuration. Check out the
445
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
446
+ """
447
+
448
+
449
+ @add_start_docstrings(
450
+ "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
451
+ LLAMA_START_DOCSTRING,
452
+ )
453
+ class LlamaPreTrainedModel(PreTrainedModel):
454
+ config_class = CLEXLlamaConfig
455
+ base_model_prefix = "model"
456
+ supports_gradient_checkpointing = True
457
+ _no_split_modules = ["LlamaDecoderLayer"]
458
+ _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
459
+ _keep_in_fp32_modules = ["model.clex_layer.proj_func.ode_up_proj", "model.clex_layer.proj_func.ode_down_proj", "model.clex_layer.inv_freq"]
460
+
461
+ def _init_weights(self, module):
462
+ std = self.config.initializer_range
463
+ if isinstance(module, nn.Linear):
464
+ module.weight.data.normal_(mean=0.0, std=std)
465
+ if module.bias is not None:
466
+ module.bias.data.zero_()
467
+ elif isinstance(module, nn.Embedding):
468
+ module.weight.data.normal_(mean=0.0, std=std)
469
+ if module.padding_idx is not None:
470
+ module.weight.data[module.padding_idx].zero_()
471
+
472
+ def _set_gradient_checkpointing(self, module, value=False):
473
+ if isinstance(module, LlamaModel):
474
+ module.gradient_checkpointing = value
475
+
476
+
477
+ LLAMA_INPUTS_DOCSTRING = r"""
478
+ Args:
479
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
480
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
481
+ it.
482
+
483
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
484
+ [`PreTrainedTokenizer.__call__`] for details.
485
+
486
+ [What are input IDs?](../glossary#input-ids)
487
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
488
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
489
+
490
+ - 1 for tokens that are **not masked**,
491
+ - 0 for tokens that are **masked**.
492
+
493
+ [What are attention masks?](../glossary#attention-mask)
494
+
495
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
496
+ [`PreTrainedTokenizer.__call__`] for details.
497
+
498
+ If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
499
+ `past_key_values`).
500
+
501
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
502
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
503
+ information on the default strategy.
504
+
505
+ - 1 indicates the head is **not masked**,
506
+ - 0 indicates the head is **masked**.
507
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
508
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
509
+ config.n_positions - 1]`.
510
+
511
+ [What are position IDs?](../glossary#position-ids)
512
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
513
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
514
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
515
+ `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
516
+
517
+ Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
518
+ blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
519
+
520
+ If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
521
+ don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
522
+ `decoder_input_ids` of shape `(batch_size, sequence_length)`.
523
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
524
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
525
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
526
+ model's internal embedding lookup matrix.
527
+ use_cache (`bool`, *optional*):
528
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
529
+ `past_key_values`).
530
+ output_attentions (`bool`, *optional*):
531
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
532
+ tensors for more detail.
533
+ output_hidden_states (`bool`, *optional*):
534
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
535
+ more detail.
536
+ return_dict (`bool`, *optional*):
537
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
538
+ """
539
+
540
+ from torchdiffeq import odeint
541
+ from CLEX.clex_layer import ODELinear
542
+ @add_start_docstrings(
543
+ "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
544
+ LLAMA_START_DOCSTRING,
545
+ )
546
+ class LlamaModel(LlamaPreTrainedModel):
547
+ """
548
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
549
+
550
+ Args:
551
+ config: CLEXLlamaConfig
552
+ """
553
+
554
+ def __init__(self, config: CLEXLlamaConfig):
555
+ super().__init__(config)
556
+ self.padding_idx = config.pad_token_id
557
+ self.vocab_size = config.vocab_size
558
+
559
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
560
+ self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
561
+ self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
562
+ head_dim = config.hidden_size // config.num_attention_heads
563
+ if config.rope_scaling["type"] == "clex":
564
+ self.clex_layer = LlamaCLEXScalingRotaryEmbedding(head_dim, config.max_position_embeddings, config.rope_scaling)
565
+ self.gradient_checkpointing = False
566
+ # Initialize weights and apply final processing
567
+ self.post_init()
568
+
569
+
570
+ def get_input_embeddings(self):
571
+ return self.embed_tokens
572
+
573
+ def set_input_embeddings(self, value):
574
+ self.embed_tokens = value
575
+
576
+ # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
577
+ def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
578
+ # create causal mask
579
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
580
+ combined_attention_mask = None
581
+ if input_shape[-1] > 1:
582
+ combined_attention_mask = _make_causal_mask(
583
+ input_shape,
584
+ inputs_embeds.dtype,
585
+ device=inputs_embeds.device,
586
+ past_key_values_length=past_key_values_length,
587
+ )
588
+
589
+ if attention_mask is not None:
590
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
591
+ expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
592
+ inputs_embeds.device
593
+ )
594
+ combined_attention_mask = (
595
+ expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
596
+ )
597
+
598
+ return combined_attention_mask
599
+
600
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
601
+ def forward(
602
+ self,
603
+ input_ids: torch.LongTensor = None,
604
+ attention_mask: Optional[torch.Tensor] = None,
605
+ position_ids: Optional[torch.LongTensor] = None,
606
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
607
+ inputs_embeds: Optional[torch.FloatTensor] = None,
608
+ use_cache: Optional[bool] = None,
609
+ output_attentions: Optional[bool] = None,
610
+ output_hidden_states: Optional[bool] = None,
611
+ return_dict: Optional[bool] = None,
612
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
613
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
614
+ output_hidden_states = (
615
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
616
+ )
617
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
618
+
619
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
620
+
621
+ # retrieve input_ids and inputs_embeds
622
+ if input_ids is not None and inputs_embeds is not None:
623
+ raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
624
+ elif input_ids is not None:
625
+ batch_size, seq_length = input_ids.shape
626
+ elif inputs_embeds is not None:
627
+ batch_size, seq_length, _ = inputs_embeds.shape
628
+ else:
629
+ raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
630
+
631
+ seq_length_with_past = seq_length
632
+ past_key_values_length = 0
633
+
634
+ if past_key_values is not None:
635
+ past_key_values_length = past_key_values[0][0].shape[2]
636
+ seq_length_with_past = seq_length_with_past + past_key_values_length
637
+
638
+ if position_ids is None:
639
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
640
+ position_ids = torch.arange(
641
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
642
+ )
643
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
644
+ else:
645
+ position_ids = position_ids.view(-1, seq_length).long()
646
+
647
+ if inputs_embeds is None:
648
+ inputs_embeds = self.embed_tokens(input_ids)
649
+ # embed positions
650
+ # if attention_mask is None:
651
+ # attention_mask = torch.ones(
652
+ # (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
653
+ # )
654
+ # attention_mask = self._prepare_decoder_attention_mask(
655
+ # attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
656
+ # )
657
+ attention_mask = None
658
+
659
+
660
+ hidden_states = inputs_embeds
661
+
662
+ if self.gradient_checkpointing and self.training:
663
+ if use_cache:
664
+ logger.warning_once(
665
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
666
+ )
667
+ use_cache = False
668
+
669
+ # decoder layers
670
+ all_hidden_states = () if output_hidden_states else None
671
+ all_self_attns = () if output_attentions else None
672
+ next_decoder_cache = () if use_cache else None
673
+
674
+ pack_cos_sin = None
675
+ if self.config.rope_scaling["type"] == "clex":
676
+ pack_cos_sin = self.clex_layer(inputs_embeds.device, inputs_embeds.dtype, seq_length_with_past, self.training)
677
+
678
+ for idx, decoder_layer in enumerate(self.layers):
679
+ if output_hidden_states:
680
+ all_hidden_states += (hidden_states,)
681
+
682
+ past_key_value = past_key_values[idx] if past_key_values is not None else None
683
+
684
+ if self.gradient_checkpointing and self.training:
685
+
686
+ def create_custom_forward(module):
687
+ def custom_forward(*inputs):
688
+ # None for past_key_value
689
+ return module(*inputs, output_attentions, None)
690
+
691
+ return custom_forward
692
+
693
+ layer_outputs = torch.utils.checkpoint.checkpoint(
694
+ create_custom_forward(decoder_layer),
695
+ hidden_states,
696
+ attention_mask,
697
+ position_ids,
698
+ pack_cos_sin,
699
+ None,
700
+ )
701
+ else:
702
+ layer_outputs = decoder_layer(
703
+ hidden_states,
704
+ attention_mask=attention_mask,
705
+ position_ids=position_ids,
706
+ pack_cos_sin=pack_cos_sin,
707
+ past_key_value=past_key_value,
708
+ output_attentions=output_attentions,
709
+ use_cache=use_cache,
710
+ )
711
+
712
+ hidden_states = layer_outputs[0]
713
+
714
+ if use_cache:
715
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
716
+
717
+ if output_attentions:
718
+ all_self_attns += (layer_outputs[1],)
719
+
720
+ hidden_states = self.norm(hidden_states)
721
+
722
+ # add hidden states from the last decoder layer
723
+ if output_hidden_states:
724
+ all_hidden_states += (hidden_states,)
725
+
726
+ next_cache = next_decoder_cache if use_cache else None
727
+ if not return_dict:
728
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
729
+ return BaseModelOutputWithPast(
730
+ last_hidden_state=hidden_states,
731
+ past_key_values=next_cache,
732
+ hidden_states=all_hidden_states,
733
+ attentions=all_self_attns,
734
+ )
735
+
736
+
737
+ class LlamaForCausalLM(LlamaPreTrainedModel):
738
+ def __init__(self, config):
739
+ super().__init__(config)
740
+ self.model = LlamaModel(config)
741
+
742
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
743
+
744
+ # Initialize weights and apply final processing
745
+ self.post_init()
746
+
747
+ def get_input_embeddings(self):
748
+ return self.model.embed_tokens
749
+
750
+ def set_input_embeddings(self, value):
751
+ self.model.embed_tokens = value
752
+
753
+ def get_output_embeddings(self):
754
+ return self.lm_head
755
+
756
+ def set_output_embeddings(self, new_embeddings):
757
+ self.lm_head = new_embeddings
758
+
759
+ def set_decoder(self, decoder):
760
+ self.model = decoder
761
+
762
+ def get_decoder(self):
763
+ return self.model
764
+
765
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
766
+ @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
767
+ def forward(
768
+ self,
769
+ input_ids: torch.LongTensor = None,
770
+ attention_mask: Optional[torch.Tensor] = None,
771
+ position_ids: Optional[torch.LongTensor] = None,
772
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
773
+ inputs_embeds: Optional[torch.FloatTensor] = None,
774
+ labels: Optional[torch.LongTensor] = None,
775
+ use_cache: Optional[bool] = None,
776
+ output_attentions: Optional[bool] = None,
777
+ output_hidden_states: Optional[bool] = None,
778
+ return_dict: Optional[bool] = None,
779
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
780
+ r"""
781
+ Args:
782
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
783
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
784
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
785
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
786
+
787
+ Returns:
788
+
789
+ Example:
790
+
791
+ ```python
792
+ >>> from transformers import AutoTokenizer, LlamaForCausalLM
793
+
794
+ >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
795
+ >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
796
+
797
+ >>> prompt = "Hey, are you consciours? Can you talk to me?"
798
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
799
+
800
+ >>> # Generate
801
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
802
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
803
+ "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
804
+ ```"""
805
+
806
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
807
+ output_hidden_states = (
808
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
809
+ )
810
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
811
+
812
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
813
+ outputs = self.model(
814
+ input_ids=input_ids,
815
+ attention_mask=attention_mask,
816
+ position_ids=position_ids,
817
+ past_key_values=past_key_values,
818
+ inputs_embeds=inputs_embeds,
819
+ use_cache=use_cache,
820
+ output_attentions=output_attentions,
821
+ output_hidden_states=output_hidden_states,
822
+ return_dict=return_dict,
823
+ )
824
+
825
+ hidden_states = outputs[0]
826
+ logits = self.lm_head(hidden_states)
827
+
828
+ loss = None
829
+ if labels is not None:
830
+ # Shift so that tokens < n predict n
831
+ shift_logits = logits[..., :-1, :].contiguous()
832
+ shift_labels = labels[..., 1:].contiguous()
833
+ # Flatten the tokens
834
+ loss_fct = CrossEntropyLoss()
835
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
836
+ shift_labels = shift_labels.view(-1)
837
+ # Enable model parallelism
838
+ shift_labels = shift_labels.to(shift_logits.device)
839
+ loss = loss_fct(shift_logits, shift_labels)
840
+ if not return_dict:
841
+ output = (logits,) + outputs[1:]
842
+ return (loss,) + output if loss is not None else output
843
+ return CausalLMOutputWithPast(
844
+ loss=loss,
845
+ logits=logits,
846
+ past_key_values=outputs.past_key_values,
847
+ hidden_states=outputs.hidden_states,
848
+ attentions=outputs.attentions,
849
+ )
850
+
851
+ def prepare_inputs_for_generation(
852
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
853
+ ):
854
+ if past_key_values:
855
+ input_ids = input_ids[:, -1:]
856
+
857
+ position_ids = kwargs.get("position_ids", None)
858
+ if attention_mask is not None and position_ids is None:
859
+ # create position_ids on the fly for batch generation
860
+ position_ids = attention_mask.long().cumsum(-1) - 1
861
+ position_ids.masked_fill_(attention_mask == 0, 1)
862
+ if past_key_values:
863
+ position_ids = position_ids[:, -1].unsqueeze(-1)
864
+
865
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
866
+ if inputs_embeds is not None and past_key_values is None:
867
+ model_inputs = {"inputs_embeds": inputs_embeds}
868
+ else:
869
+ model_inputs = {"input_ids": input_ids}
870
+
871
+ model_inputs.update(
872
+ {
873
+ "position_ids": position_ids,
874
+ "past_key_values": past_key_values,
875
+ "use_cache": kwargs.get("use_cache"),
876
+ "attention_mask": attention_mask,
877
+ }
878
+ )
879
+ return model_inputs
880
+
881
+ @staticmethod
882
+ def _reorder_cache(past_key_values, beam_idx):
883
+ reordered_past = ()
884
+ for layer_past in past_key_values:
885
+ reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
886
+ return reordered_past
887
+
888
+
889
+ @add_start_docstrings(
890
+ """
891
+ The LLaMa Model transformer with a sequence classification head on top (linear layer).
892
+
893
+ [`LlamaForSequenceClassification`] uses the last token in order to do the classification, as other causal models
894
+ (e.g. GPT-2) do.
895
+
896
+ Since it does classification on the last token, it requires to know the position of the last token. If a
897
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
898
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
899
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
900
+ each row of the batch).
901
+ """,
902
+ LLAMA_START_DOCSTRING,
903
+ )
904
+ class LlamaForSequenceClassification(LlamaPreTrainedModel):
905
+ _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
906
+
907
+ def __init__(self, config):
908
+ super().__init__(config)
909
+ self.num_labels = config.num_labels
910
+ self.model = LlamaModel(config)
911
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
912
+
913
+ # Initialize weights and apply final processing
914
+ self.post_init()
915
+
916
+ def get_input_embeddings(self):
917
+ return self.model.embed_tokens
918
+
919
+ def set_input_embeddings(self, value):
920
+ self.model.embed_tokens = value
921
+
922
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
923
+ def forward(
924
+ self,
925
+ input_ids: torch.LongTensor = None,
926
+ attention_mask: Optional[torch.Tensor] = None,
927
+ position_ids: Optional[torch.LongTensor] = None,
928
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
929
+ inputs_embeds: Optional[torch.FloatTensor] = None,
930
+ labels: Optional[torch.LongTensor] = None,
931
+ use_cache: Optional[bool] = None,
932
+ output_attentions: Optional[bool] = None,
933
+ output_hidden_states: Optional[bool] = None,
934
+ return_dict: Optional[bool] = None,
935
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
936
+ r"""
937
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
938
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
939
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
940
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
941
+ """
942
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
943
+
944
+ transformer_outputs = self.model(
945
+ input_ids,
946
+ attention_mask=attention_mask,
947
+ position_ids=position_ids,
948
+ past_key_values=past_key_values,
949
+ inputs_embeds=inputs_embeds,
950
+ use_cache=use_cache,
951
+ output_attentions=output_attentions,
952
+ output_hidden_states=output_hidden_states,
953
+ return_dict=return_dict,
954
+ )
955
+ hidden_states = transformer_outputs[0]
956
+ logits = self.score(hidden_states)
957
+
958
+ if input_ids is not None:
959
+ batch_size = input_ids.shape[0]
960
+ else:
961
+ batch_size = inputs_embeds.shape[0]
962
+
963
+ if self.config.pad_token_id is None and batch_size != 1:
964
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
965
+ if self.config.pad_token_id is None:
966
+ sequence_lengths = -1
967
+ else:
968
+ if input_ids is not None:
969
+ sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
970
+ else:
971
+ sequence_lengths = -1
972
+
973
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
974
+
975
+ loss = None
976
+ if labels is not None:
977
+ labels = labels.to(logits.device)
978
+ if self.config.problem_type is None:
979
+ if self.num_labels == 1:
980
+ self.config.problem_type = "regression"
981
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
982
+ self.config.problem_type = "single_label_classification"
983
+ else:
984
+ self.config.problem_type = "multi_label_classification"
985
+
986
+ if self.config.problem_type == "regression":
987
+ loss_fct = MSELoss()
988
+ if self.num_labels == 1:
989
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
990
+ else:
991
+ loss = loss_fct(pooled_logits, labels)
992
+ elif self.config.problem_type == "single_label_classification":
993
+ loss_fct = CrossEntropyLoss()
994
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
995
+ elif self.config.problem_type == "multi_label_classification":
996
+ loss_fct = BCEWithLogitsLoss()
997
+ loss = loss_fct(pooled_logits, labels)
998
+ if not return_dict:
999
+ output = (pooled_logits,) + transformer_outputs[1:]
1000
+ return ((loss,) + output) if loss is not None else output
1001
+
1002
+ return SequenceClassifierOutputWithPast(
1003
+ loss=loss,
1004
+ logits=pooled_logits,
1005
+ past_key_values=transformer_outputs.past_key_values,
1006
+ hidden_states=transformer_outputs.hidden_states,
1007
+ attentions=transformer_outputs.attentions,
1008
+ )