TheBloke commited on
Commit
5329e07
1 Parent(s): c0bf34a

Initial GPTQ model commit

Browse files
Files changed (1) hide show
  1. modelling_llama.py +894 -0
modelling_llama.py ADDED
@@ -0,0 +1,894 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 transformers.models.llama.modeling_llama import LlamaConfig
34
+
35
+ logger = logging.get_logger(__name__)
36
+
37
+ _CONFIG_FOR_DOC = "LlamaConfig"
38
+
39
+
40
+ # Copied from transformers.models.bart.modeling_bart._make_causal_mask
41
+ def _make_causal_mask(
42
+ input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
43
+ ):
44
+ """
45
+ Make causal mask used for bi-directional self-attention.
46
+ """
47
+ bsz, tgt_len = input_ids_shape
48
+ mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
49
+ mask_cond = torch.arange(mask.size(-1), device=device)
50
+ mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
51
+ mask = mask.to(dtype)
52
+
53
+ if past_key_values_length > 0:
54
+ mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
55
+ return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
56
+
57
+
58
+ # Copied from transformers.models.bart.modeling_bart._expand_mask
59
+ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
60
+ """
61
+ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
62
+ """
63
+ bsz, src_len = mask.size()
64
+ tgt_len = tgt_len if tgt_len is not None else src_len
65
+
66
+ expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
67
+
68
+ inverted_mask = 1.0 - expanded_mask
69
+
70
+ return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
71
+
72
+
73
+ class LlamaRMSNorm(nn.Module):
74
+ def __init__(self, hidden_size, eps=1e-6):
75
+ """
76
+ LlamaRMSNorm is equivalent to T5LayerNorm
77
+ """
78
+ super().__init__()
79
+ self.weight = nn.Parameter(torch.ones(hidden_size))
80
+ self.variance_epsilon = eps
81
+
82
+ def forward(self, hidden_states):
83
+ input_dtype = hidden_states.dtype
84
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
85
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
86
+
87
+ return (self.weight * hidden_states).to(input_dtype)
88
+
89
+
90
+ class LlamaRotaryEmbedding(torch.nn.Module):
91
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, scale=1, device=None):
92
+ super().__init__()
93
+ inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
94
+ self.register_buffer("inv_freq", inv_freq)
95
+
96
+ # Build here to make `torch.jit.trace` work.
97
+ self.max_seq_len_cached = max_position_embeddings
98
+ t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
99
+
100
+ self.scale = scale
101
+ t *= self.scale
102
+
103
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
104
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
105
+ emb = torch.cat((freqs, freqs), dim=-1)
106
+ dtype = torch.get_default_dtype()
107
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
108
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
109
+
110
+ def forward(self, x, seq_len=None):
111
+ # x: [bs, num_attention_heads, seq_len, head_size]
112
+ # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
113
+ if seq_len > self.max_seq_len_cached:
114
+ self.max_seq_len_cached = seq_len
115
+ t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
116
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
117
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
118
+ emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
119
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(x.dtype), persistent=False)
120
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(x.dtype), persistent=False)
121
+ return (
122
+ self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
123
+ self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
124
+ )
125
+
126
+
127
+ def rotate_half(x):
128
+ """Rotates half the hidden dims of the input."""
129
+ x1 = x[..., : x.shape[-1] // 2]
130
+ x2 = x[..., x.shape[-1] // 2 :]
131
+ return torch.cat((-x2, x1), dim=-1)
132
+
133
+
134
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
135
+ # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
136
+ cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
137
+ sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
138
+ cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
139
+ sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
140
+ q_embed = (q * cos) + (rotate_half(q) * sin)
141
+ k_embed = (k * cos) + (rotate_half(k) * sin)
142
+ return q_embed, k_embed
143
+
144
+
145
+ class LlamaMLP(nn.Module):
146
+ def __init__(
147
+ self,
148
+ hidden_size: int,
149
+ intermediate_size: int,
150
+ hidden_act: str,
151
+ ):
152
+ super().__init__()
153
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
154
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
155
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
156
+ self.act_fn = ACT2FN[hidden_act]
157
+
158
+ def forward(self, x):
159
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
160
+
161
+
162
+ class LlamaAttention(nn.Module):
163
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
164
+
165
+ def __init__(self, config: LlamaConfig):
166
+ super().__init__()
167
+ self.config = config
168
+ self.hidden_size = config.hidden_size
169
+ self.num_heads = config.num_attention_heads
170
+ self.head_dim = self.hidden_size // self.num_heads
171
+ self.max_position_embeddings = config.max_position_embeddings
172
+ self.position_embeddings_scale = 2048 / self.max_position_embeddings
173
+
174
+ if (self.head_dim * self.num_heads) != self.hidden_size:
175
+ raise ValueError(
176
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
177
+ f" and `num_heads`: {self.num_heads})."
178
+ )
179
+ self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
180
+ self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
181
+ self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
182
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
183
+ self.rotary_emb = LlamaRotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings, scale=self.position_embeddings_scale)
184
+
185
+ def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
186
+ return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
187
+
188
+ def forward(
189
+ self,
190
+ hidden_states: torch.Tensor,
191
+ attention_mask: Optional[torch.Tensor] = None,
192
+ position_ids: Optional[torch.LongTensor] = None,
193
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
194
+ output_attentions: bool = False,
195
+ use_cache: bool = False,
196
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
197
+ bsz, q_len, _ = hidden_states.size()
198
+
199
+ query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
200
+ key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
201
+ value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
202
+
203
+ kv_seq_len = key_states.shape[-2]
204
+ if past_key_value is not None:
205
+ kv_seq_len += past_key_value[0].shape[-2]
206
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
207
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
208
+ # [bsz, nh, t, hd]
209
+
210
+ if past_key_value is not None:
211
+ # reuse k, v, self_attention
212
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
213
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
214
+
215
+ past_key_value = (key_states, value_states) if use_cache else None
216
+
217
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
218
+
219
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
220
+ raise ValueError(
221
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
222
+ f" {attn_weights.size()}"
223
+ )
224
+
225
+ if attention_mask is not None:
226
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
227
+ raise ValueError(
228
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
229
+ )
230
+ attn_weights = attn_weights + attention_mask
231
+ attn_weights = torch.max(
232
+ attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min, device=attn_weights.device)
233
+ )
234
+
235
+ # upcast attention to fp32
236
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
237
+ attn_output = torch.matmul(attn_weights, value_states)
238
+
239
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
240
+ raise ValueError(
241
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
242
+ f" {attn_output.size()}"
243
+ )
244
+
245
+ attn_output = attn_output.transpose(1, 2)
246
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
247
+
248
+ attn_output = self.o_proj(attn_output)
249
+
250
+ if not output_attentions:
251
+ attn_weights = None
252
+
253
+ return attn_output, attn_weights, past_key_value
254
+
255
+
256
+ class LlamaDecoderLayer(nn.Module):
257
+ def __init__(self, config: LlamaConfig):
258
+ super().__init__()
259
+ self.hidden_size = config.hidden_size
260
+ self.self_attn = LlamaAttention(config=config)
261
+ self.mlp = LlamaMLP(
262
+ hidden_size=self.hidden_size,
263
+ intermediate_size=config.intermediate_size,
264
+ hidden_act=config.hidden_act,
265
+ )
266
+ self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
267
+ self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
268
+
269
+ def forward(
270
+ self,
271
+ hidden_states: torch.Tensor,
272
+ attention_mask: Optional[torch.Tensor] = None,
273
+ position_ids: Optional[torch.LongTensor] = None,
274
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
275
+ output_attentions: Optional[bool] = False,
276
+ use_cache: Optional[bool] = False,
277
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
278
+ """
279
+ Args:
280
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
281
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
282
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
283
+ output_attentions (`bool`, *optional*):
284
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
285
+ returned tensors for more detail.
286
+ use_cache (`bool`, *optional*):
287
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
288
+ (see `past_key_values`).
289
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
290
+ """
291
+
292
+ residual = hidden_states
293
+
294
+ hidden_states = self.input_layernorm(hidden_states)
295
+
296
+ # Self Attention
297
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
298
+ hidden_states=hidden_states,
299
+ attention_mask=attention_mask,
300
+ position_ids=position_ids,
301
+ past_key_value=past_key_value,
302
+ output_attentions=output_attentions,
303
+ use_cache=use_cache,
304
+ )
305
+ hidden_states = residual + hidden_states
306
+
307
+ # Fully Connected
308
+ residual = hidden_states
309
+ hidden_states = self.post_attention_layernorm(hidden_states)
310
+ hidden_states = self.mlp(hidden_states)
311
+ hidden_states = residual + hidden_states
312
+
313
+ outputs = (hidden_states,)
314
+
315
+ if output_attentions:
316
+ outputs += (self_attn_weights,)
317
+
318
+ if use_cache:
319
+ outputs += (present_key_value,)
320
+
321
+ return outputs
322
+
323
+
324
+ LLAMA_START_DOCSTRING = r"""
325
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
326
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
327
+ etc.)
328
+
329
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
330
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
331
+ and behavior.
332
+
333
+ Parameters:
334
+ config ([`LlamaConfig`]):
335
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
336
+ load the weights associated with the model, only the configuration. Check out the
337
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
338
+ """
339
+
340
+
341
+ @add_start_docstrings(
342
+ "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
343
+ LLAMA_START_DOCSTRING,
344
+ )
345
+ class LlamaPreTrainedModel(PreTrainedModel):
346
+ config_class = LlamaConfig
347
+ base_model_prefix = "model"
348
+ supports_gradient_checkpointing = True
349
+ _no_split_modules = ["LlamaDecoderLayer"]
350
+ _skip_keys_device_placement = "past_key_values"
351
+ _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
352
+
353
+ def _init_weights(self, module):
354
+ std = self.config.initializer_range
355
+ if isinstance(module, nn.Linear):
356
+ module.weight.data.normal_(mean=0.0, std=std)
357
+ if module.bias is not None:
358
+ module.bias.data.zero_()
359
+ elif isinstance(module, nn.Embedding):
360
+ module.weight.data.normal_(mean=0.0, std=std)
361
+ if module.padding_idx is not None:
362
+ module.weight.data[module.padding_idx].zero_()
363
+
364
+ def _set_gradient_checkpointing(self, module, value=False):
365
+ if isinstance(module, LlamaModel):
366
+ module.gradient_checkpointing = value
367
+
368
+
369
+ LLAMA_INPUTS_DOCSTRING = r"""
370
+ Args:
371
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
372
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
373
+ it.
374
+
375
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
376
+ [`PreTrainedTokenizer.__call__`] for details.
377
+
378
+ [What are input IDs?](../glossary#input-ids)
379
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
380
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
381
+
382
+ - 1 for tokens that are **not masked**,
383
+ - 0 for tokens that are **masked**.
384
+
385
+ [What are attention masks?](../glossary#attention-mask)
386
+
387
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
388
+ [`PreTrainedTokenizer.__call__`] for details.
389
+
390
+ If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
391
+ `past_key_values`).
392
+
393
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
394
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
395
+ information on the default strategy.
396
+
397
+ - 1 indicates the head is **not masked**,
398
+ - 0 indicates the head is **masked**.
399
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
400
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
401
+ config.n_positions - 1]`.
402
+
403
+ [What are position IDs?](../glossary#position-ids)
404
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
405
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
406
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
407
+ `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
408
+
409
+ Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
410
+ blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
411
+
412
+ If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
413
+ don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
414
+ `decoder_input_ids` of shape `(batch_size, sequence_length)`.
415
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
416
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
417
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
418
+ model's internal embedding lookup matrix.
419
+ use_cache (`bool`, *optional*):
420
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
421
+ `past_key_values`).
422
+ output_attentions (`bool`, *optional*):
423
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
424
+ tensors for more detail.
425
+ output_hidden_states (`bool`, *optional*):
426
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
427
+ more detail.
428
+ return_dict (`bool`, *optional*):
429
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
430
+ """
431
+
432
+
433
+ @add_start_docstrings(
434
+ "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
435
+ LLAMA_START_DOCSTRING,
436
+ )
437
+ class LlamaModel(LlamaPreTrainedModel):
438
+ """
439
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
440
+
441
+ Args:
442
+ config: LlamaConfig
443
+ """
444
+
445
+ def __init__(self, config: LlamaConfig):
446
+ super().__init__(config)
447
+ self.padding_idx = config.pad_token_id
448
+ self.vocab_size = config.vocab_size
449
+
450
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
451
+ self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
452
+ self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
453
+
454
+ self.gradient_checkpointing = False
455
+ # Initialize weights and apply final processing
456
+ self.post_init()
457
+
458
+ def get_input_embeddings(self):
459
+ return self.embed_tokens
460
+
461
+ def set_input_embeddings(self, value):
462
+ self.embed_tokens = value
463
+
464
+ # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
465
+ def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
466
+ # create causal mask
467
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
468
+ combined_attention_mask = None
469
+ if input_shape[-1] > 1:
470
+ combined_attention_mask = _make_causal_mask(
471
+ input_shape,
472
+ inputs_embeds.dtype,
473
+ device=inputs_embeds.device,
474
+ past_key_values_length=past_key_values_length,
475
+ )
476
+
477
+ if attention_mask is not None:
478
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
479
+ expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
480
+ inputs_embeds.device
481
+ )
482
+ combined_attention_mask = (
483
+ expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
484
+ )
485
+
486
+ return combined_attention_mask
487
+
488
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
489
+ def forward(
490
+ self,
491
+ input_ids: torch.LongTensor = None,
492
+ attention_mask: Optional[torch.Tensor] = None,
493
+ position_ids: Optional[torch.LongTensor] = None,
494
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
495
+ inputs_embeds: Optional[torch.FloatTensor] = None,
496
+ use_cache: Optional[bool] = None,
497
+ output_attentions: Optional[bool] = None,
498
+ output_hidden_states: Optional[bool] = None,
499
+ return_dict: Optional[bool] = None,
500
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
501
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
502
+ output_hidden_states = (
503
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
504
+ )
505
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
506
+
507
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
508
+
509
+ # retrieve input_ids and inputs_embeds
510
+ if input_ids is not None and inputs_embeds is not None:
511
+ raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
512
+ elif input_ids is not None:
513
+ batch_size, seq_length = input_ids.shape
514
+ elif inputs_embeds is not None:
515
+ batch_size, seq_length, _ = inputs_embeds.shape
516
+ else:
517
+ raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
518
+
519
+ seq_length_with_past = seq_length
520
+ past_key_values_length = 0
521
+
522
+ if past_key_values is not None:
523
+ past_key_values_length = past_key_values[0][0].shape[2]
524
+ seq_length_with_past = seq_length_with_past + past_key_values_length
525
+
526
+ if position_ids is None:
527
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
528
+ position_ids = torch.arange(
529
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
530
+ )
531
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
532
+ else:
533
+ position_ids = position_ids.view(-1, seq_length).long()
534
+
535
+ if inputs_embeds is None:
536
+ inputs_embeds = self.embed_tokens(input_ids)
537
+ # embed positions
538
+ if attention_mask is None:
539
+ attention_mask = torch.ones(
540
+ (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
541
+ )
542
+ attention_mask = self._prepare_decoder_attention_mask(
543
+ attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
544
+ )
545
+
546
+ hidden_states = inputs_embeds
547
+
548
+ if self.gradient_checkpointing and self.training:
549
+ if use_cache:
550
+ logger.warning_once(
551
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
552
+ )
553
+ use_cache = False
554
+
555
+ # decoder layers
556
+ all_hidden_states = () if output_hidden_states else None
557
+ all_self_attns = () if output_attentions else None
558
+ next_decoder_cache = () if use_cache else None
559
+
560
+ for idx, decoder_layer in enumerate(self.layers):
561
+ if output_hidden_states:
562
+ all_hidden_states += (hidden_states,)
563
+
564
+ past_key_value = past_key_values[idx] if past_key_values is not None else None
565
+
566
+ if self.gradient_checkpointing and self.training:
567
+
568
+ def create_custom_forward(module):
569
+ def custom_forward(*inputs):
570
+ # None for past_key_value
571
+ return module(*inputs, output_attentions, None)
572
+
573
+ return custom_forward
574
+
575
+ layer_outputs = torch.utils.checkpoint.checkpoint(
576
+ create_custom_forward(decoder_layer),
577
+ hidden_states,
578
+ attention_mask,
579
+ position_ids,
580
+ None,
581
+ )
582
+ else:
583
+ layer_outputs = decoder_layer(
584
+ hidden_states,
585
+ attention_mask=attention_mask,
586
+ position_ids=position_ids,
587
+ past_key_value=past_key_value,
588
+ output_attentions=output_attentions,
589
+ use_cache=use_cache,
590
+ )
591
+
592
+ hidden_states = layer_outputs[0]
593
+
594
+ if use_cache:
595
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
596
+
597
+ if output_attentions:
598
+ all_self_attns += (layer_outputs[1],)
599
+
600
+ hidden_states = self.norm(hidden_states)
601
+
602
+ # add hidden states from the last decoder layer
603
+ if output_hidden_states:
604
+ all_hidden_states += (hidden_states,)
605
+
606
+ next_cache = next_decoder_cache if use_cache else None
607
+ if not return_dict:
608
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
609
+ return BaseModelOutputWithPast(
610
+ last_hidden_state=hidden_states,
611
+ past_key_values=next_cache,
612
+ hidden_states=all_hidden_states,
613
+ attentions=all_self_attns,
614
+ )
615
+
616
+
617
+ class LlamaForCausalLM(LlamaPreTrainedModel):
618
+ _tied_weights_keys = ["lm_head.weight"]
619
+
620
+ def __init__(self, config):
621
+ super().__init__(config)
622
+ self.model = LlamaModel(config)
623
+
624
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
625
+
626
+ # Initialize weights and apply final processing
627
+ self.post_init()
628
+
629
+ def get_input_embeddings(self):
630
+ return self.model.embed_tokens
631
+
632
+ def set_input_embeddings(self, value):
633
+ self.model.embed_tokens = value
634
+
635
+ def get_output_embeddings(self):
636
+ return self.lm_head
637
+
638
+ def set_output_embeddings(self, new_embeddings):
639
+ self.lm_head = new_embeddings
640
+
641
+ def set_decoder(self, decoder):
642
+ self.model = decoder
643
+
644
+ def get_decoder(self):
645
+ return self.model
646
+
647
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
648
+ @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
649
+ def forward(
650
+ self,
651
+ input_ids: torch.LongTensor = None,
652
+ attention_mask: Optional[torch.Tensor] = None,
653
+ position_ids: Optional[torch.LongTensor] = None,
654
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
655
+ inputs_embeds: Optional[torch.FloatTensor] = None,
656
+ labels: Optional[torch.LongTensor] = None,
657
+ use_cache: Optional[bool] = None,
658
+ output_attentions: Optional[bool] = None,
659
+ output_hidden_states: Optional[bool] = None,
660
+ return_dict: Optional[bool] = None,
661
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
662
+ r"""
663
+ Args:
664
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
665
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
666
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
667
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
668
+
669
+ Returns:
670
+
671
+ Example:
672
+
673
+ ```python
674
+ >>> from transformers import AutoTokenizer, LlamaForCausalLM
675
+
676
+ >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
677
+ >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
678
+
679
+ >>> prompt = "Hey, are you conscious? Can you talk to me?"
680
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
681
+
682
+ >>> # Generate
683
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
684
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
685
+ "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
686
+ ```"""
687
+
688
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
689
+ output_hidden_states = (
690
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
691
+ )
692
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
693
+
694
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
695
+ outputs = self.model(
696
+ input_ids=input_ids,
697
+ attention_mask=attention_mask,
698
+ position_ids=position_ids,
699
+ past_key_values=past_key_values,
700
+ inputs_embeds=inputs_embeds,
701
+ use_cache=use_cache,
702
+ output_attentions=output_attentions,
703
+ output_hidden_states=output_hidden_states,
704
+ return_dict=return_dict,
705
+ )
706
+
707
+ hidden_states = outputs[0]
708
+ logits = self.lm_head(hidden_states)
709
+
710
+ loss = None
711
+ if labels is not None:
712
+ # Shift so that tokens < n predict n
713
+ shift_logits = logits[..., :-1, :].contiguous()
714
+ shift_labels = labels[..., 1:].contiguous()
715
+ # Flatten the tokens
716
+ loss_fct = CrossEntropyLoss()
717
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
718
+ shift_labels = shift_labels.view(-1)
719
+ # Enable model parallelism
720
+ shift_labels = shift_labels.to(shift_logits.device)
721
+ loss = loss_fct(shift_logits, shift_labels)
722
+
723
+ if not return_dict:
724
+ output = (logits,) + outputs[1:]
725
+ return (loss,) + output if loss is not None else output
726
+
727
+ return CausalLMOutputWithPast(
728
+ loss=loss,
729
+ logits=logits,
730
+ past_key_values=outputs.past_key_values,
731
+ hidden_states=outputs.hidden_states,
732
+ attentions=outputs.attentions,
733
+ )
734
+
735
+ def prepare_inputs_for_generation(
736
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
737
+ ):
738
+ if past_key_values:
739
+ input_ids = input_ids[:, -1:]
740
+
741
+ position_ids = kwargs.get("position_ids", None)
742
+ if attention_mask is not None and position_ids is None:
743
+ # create position_ids on the fly for batch generation
744
+ position_ids = attention_mask.long().cumsum(-1) - 1
745
+ position_ids.masked_fill_(attention_mask == 0, 1)
746
+ if past_key_values:
747
+ position_ids = position_ids[:, -1].unsqueeze(-1)
748
+
749
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
750
+ if inputs_embeds is not None and past_key_values is None:
751
+ model_inputs = {"inputs_embeds": inputs_embeds}
752
+ else:
753
+ model_inputs = {"input_ids": input_ids}
754
+
755
+ model_inputs.update(
756
+ {
757
+ "position_ids": position_ids,
758
+ "past_key_values": past_key_values,
759
+ "use_cache": kwargs.get("use_cache"),
760
+ "attention_mask": attention_mask,
761
+ }
762
+ )
763
+ return model_inputs
764
+
765
+ @staticmethod
766
+ def _reorder_cache(past_key_values, beam_idx):
767
+ reordered_past = ()
768
+ for layer_past in past_key_values:
769
+ reordered_past += (
770
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
771
+ )
772
+ return reordered_past
773
+
774
+
775
+ @add_start_docstrings(
776
+ """
777
+ The LLaMa Model transformer with a sequence classification head on top (linear layer).
778
+
779
+ [`LlamaForSequenceClassification`] uses the last token in order to do the classification, as other causal models
780
+ (e.g. GPT-2) do.
781
+
782
+ Since it does classification on the last token, it requires to know the position of the last token. If a
783
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
784
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
785
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
786
+ each row of the batch).
787
+ """,
788
+ LLAMA_START_DOCSTRING,
789
+ )
790
+ class LlamaForSequenceClassification(LlamaPreTrainedModel):
791
+ _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
792
+
793
+ def __init__(self, config):
794
+ super().__init__(config)
795
+ self.num_labels = config.num_labels
796
+ self.model = LlamaModel(config)
797
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
798
+
799
+ # Initialize weights and apply final processing
800
+ self.post_init()
801
+
802
+ def get_input_embeddings(self):
803
+ return self.model.embed_tokens
804
+
805
+ def set_input_embeddings(self, value):
806
+ self.model.embed_tokens = value
807
+
808
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
809
+ def forward(
810
+ self,
811
+ input_ids: torch.LongTensor = None,
812
+ attention_mask: Optional[torch.Tensor] = None,
813
+ position_ids: Optional[torch.LongTensor] = None,
814
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
815
+ inputs_embeds: Optional[torch.FloatTensor] = None,
816
+ labels: Optional[torch.LongTensor] = None,
817
+ use_cache: Optional[bool] = None,
818
+ output_attentions: Optional[bool] = None,
819
+ output_hidden_states: Optional[bool] = None,
820
+ return_dict: Optional[bool] = None,
821
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
822
+ r"""
823
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
824
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
825
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
826
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
827
+ """
828
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
829
+
830
+ transformer_outputs = self.model(
831
+ input_ids,
832
+ attention_mask=attention_mask,
833
+ position_ids=position_ids,
834
+ past_key_values=past_key_values,
835
+ inputs_embeds=inputs_embeds,
836
+ use_cache=use_cache,
837
+ output_attentions=output_attentions,
838
+ output_hidden_states=output_hidden_states,
839
+ return_dict=return_dict,
840
+ )
841
+ hidden_states = transformer_outputs[0]
842
+ logits = self.score(hidden_states)
843
+
844
+ if input_ids is not None:
845
+ batch_size = input_ids.shape[0]
846
+ else:
847
+ batch_size = inputs_embeds.shape[0]
848
+
849
+ if self.config.pad_token_id is None and batch_size != 1:
850
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
851
+ if self.config.pad_token_id is None:
852
+ sequence_lengths = -1
853
+ else:
854
+ if input_ids is not None:
855
+ sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
856
+ else:
857
+ sequence_lengths = -1
858
+
859
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
860
+
861
+ loss = None
862
+ if labels is not None:
863
+ labels = labels.to(logits.device)
864
+ if self.config.problem_type is None:
865
+ if self.num_labels == 1:
866
+ self.config.problem_type = "regression"
867
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
868
+ self.config.problem_type = "single_label_classification"
869
+ else:
870
+ self.config.problem_type = "multi_label_classification"
871
+
872
+ if self.config.problem_type == "regression":
873
+ loss_fct = MSELoss()
874
+ if self.num_labels == 1:
875
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
876
+ else:
877
+ loss = loss_fct(pooled_logits, labels)
878
+ elif self.config.problem_type == "single_label_classification":
879
+ loss_fct = CrossEntropyLoss()
880
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
881
+ elif self.config.problem_type == "multi_label_classification":
882
+ loss_fct = BCEWithLogitsLoss()
883
+ loss = loss_fct(pooled_logits, labels)
884
+ if not return_dict:
885
+ output = (pooled_logits,) + transformer_outputs[1:]
886
+ return ((loss,) + output) if loss is not None else output
887
+
888
+ return SequenceClassifierOutputWithPast(
889
+ loss=loss,
890
+ logits=pooled_logits,
891
+ past_key_values=transformer_outputs.past_key_values,
892
+ hidden_states=transformer_outputs.hidden_states,
893
+ attentions=transformer_outputs.attentions,
894
+ )