TheBloke commited on
Commit
e033630
1 Parent(s): 09d9e02

Initial GPTQ model commit

Browse files
Files changed (1) hide show
  1. modeling_llama_together_yarn.py +1183 -0
modeling_llama_together_yarn.py ADDED
@@ -0,0 +1,1183 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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.nn.functional as F
26
+ import torch.utils.checkpoint
27
+ from torch import nn
28
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
29
+
30
+ from transformers.activations import ACT2FN
31
+ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
32
+ from transformers.modeling_utils import PreTrainedModel
33
+ from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
34
+ from .configuration_llama import LlamaConfig
35
+
36
+
37
+ try:
38
+ from flash_attn.flash_attn_interface import (
39
+ flash_attn_func,
40
+ flash_attn_kvpacked_func,
41
+ flash_attn_qkvpacked_func,
42
+ flash_attn_varlen_kvpacked_func,
43
+ )
44
+ from flash_attn.bert_padding import unpad_input, pad_input
45
+ flash_attn_v2_installed = True
46
+ print('>>>> Flash Attention installed')
47
+ except ImportError:
48
+ flash_attn_v2_installed = False
49
+ raise ImportError('Please install Flash Attention: `pip install flash-attn --no-build-isolation`')
50
+
51
+ try:
52
+ from flash_attn.layers.rotary import apply_rotary_emb_func
53
+ flash_rope_installed = True
54
+ print('>>>> Flash RoPE installed')
55
+ except ImportError:
56
+ flash_rope_installed = False
57
+ raise ImportError('Please install RoPE kernels: `pip install git+https://github.com/HazyResearch/flash-attention.git#subdirectory=csrc/rotary`')
58
+
59
+
60
+ logger = logging.get_logger(__name__)
61
+
62
+ _CONFIG_FOR_DOC = "LlamaConfig"
63
+
64
+
65
+ #@torch.jit.script
66
+ def rmsnorm_func(hidden_states, weight, variance_epsilon):
67
+ input_dtype = hidden_states.dtype
68
+ hidden_states = hidden_states.to(torch.float32)
69
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
70
+ hidden_states = hidden_states * torch.rsqrt(variance + variance_epsilon)
71
+ return (weight * hidden_states).to(input_dtype)
72
+
73
+
74
+ class LlamaRMSNorm(nn.Module):
75
+ def __init__(self, hidden_size, eps=1e-6):
76
+ """
77
+ LlamaRMSNorm is equivalent to T5LayerNorm
78
+ """
79
+ super().__init__()
80
+ self.weight = nn.Parameter(torch.ones(hidden_size))
81
+ self.register_buffer(
82
+ "variance_epsilon",
83
+ torch.tensor(eps),
84
+ persistent=False,
85
+ )
86
+
87
+ def forward(self, hidden_states):
88
+ return rmsnorm_func(hidden_states, self.weight, self.variance_epsilon)
89
+
90
+
91
+ # Inverse dim formula to find dim based on number of rotations
92
+ def _yarn_find_correction_dim(num_rotations, dim, base=10000, max_position_embeddings=2048):
93
+ return (dim * math.log(max_position_embeddings/(num_rotations * 2 * math.pi)))/(2 * math.log(base))
94
+
95
+ # Find dim range bounds based on rotations
96
+ def _yarn_find_correction_range(low_rot, high_rot, dim, base=10000, max_position_embeddings=2048):
97
+ low = math.floor(_yarn_find_correction_dim(
98
+ low_rot, dim, base, max_position_embeddings))
99
+ high = math.ceil(_yarn_find_correction_dim(
100
+ high_rot, dim, base, max_position_embeddings))
101
+ return max(low, 0), min(high, dim-1) # Clamp values just in case
102
+
103
+ def _yarn_linear_ramp_mask(min, max, dim):
104
+ if min == max:
105
+ max += 0.001 # Prevent singularity
106
+
107
+ linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
108
+ ramp_func = torch.clamp(linear_func, 0, 1)
109
+ return ramp_func
110
+
111
+ def _yarn_get_mscale(scale=1):
112
+ if scale <= 1:
113
+ return 1.0
114
+ return 0.1 * math.log(scale) + 1.0
115
+
116
+ class FlashYaRNRotaryEmbedding(torch.nn.Module):
117
+ """
118
+ The rotary position embeddings from RoFormer_ (Su et. al).
119
+ A crucial insight from the method is that the query and keys are
120
+ transformed by rotation matrices which depend on the relative positions.
121
+
122
+ Other implementations are available in the Rotary Transformer repo_ and in
123
+ GPT-NeoX_, GPT-NeoX was an inspiration
124
+
125
+ .. _RoFormer: https://arxiv.org/abs/2104.09864
126
+ .. _repo: https://github.com/ZhuiyiTechnology/roformer
127
+ .. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
128
+
129
+ This implements the YaRN extension method.
130
+ """
131
+
132
+ def __init__(self, dim: int, base=10000.0, interleaved=False,
133
+ scaling_factor=1.0, pos_idx_in_fp32=True,
134
+ max_position_embeddings=2048,
135
+ original_max_position_embeddings=2048, extrapolation_factor=1,
136
+ attn_factor=1, beta_fast=32, beta_slow=1,
137
+ dynamic=False, finetuned=False, device=None):
138
+ """
139
+ interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
140
+ of 1st half and 2nd half (GPT-NeoX style).
141
+ pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
142
+ otherwise they might be in lower precision.
143
+ This option was added because previously (before 2023-07-02), when we construct
144
+ the position indices, we use the dtype of self.inv_freq. In most cases this would
145
+ be fp32, but if the model is trained in pure bf16 (not mixed precision), then
146
+ self.inv_freq would be bf16, and the position indices are also in bf16.
147
+ Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
148
+ embeddings for some positions will coincide.
149
+ To maintain compatibility with models previously trained in pure bf16,
150
+ we add this option.
151
+ scaling_factor: RotaryEmbedding extended with YaRN scaling.
152
+ """
153
+ super().__init__()
154
+
155
+ self.dim = dim
156
+ self.base = float(base)
157
+ self.interleaved = interleaved
158
+ self.scaling_factor = scaling_factor
159
+ self.max_position_embeddings = max_position_embeddings
160
+ self.original_max_position_embeddings = original_max_position_embeddings if original_max_position_embeddings else max_position_embeddings
161
+ self.extrapolation_factor = extrapolation_factor
162
+ self.attn_factor = attn_factor
163
+ self.beta_fast = beta_fast
164
+ self.beta_slow = beta_slow
165
+ self.pos_idx_in_fp32 = pos_idx_in_fp32
166
+ self.mscale = float(_yarn_get_mscale(self.scaling_factor) * attn_factor) # Get n-d magnitude scaling corrected for interpolation
167
+ self.dynamic = dynamic
168
+ self.finetuned = finetuned
169
+
170
+ # Generate and save the inverse frequency buffer (non trainable)
171
+ if not dynamic:
172
+ self._compute_inv_freq(scaling_factor, device)
173
+
174
+ self._seq_len_cached = 0
175
+ self._cos_cached = None
176
+ self._sin_cached = None
177
+
178
+ def _compute_inv_freq(self, scaling_factor, device=None):
179
+ pos_freqs = self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim)
180
+ inv_freq_extrapolation = 1.0 / pos_freqs
181
+ inv_freq_interpolation = 1.0 / (scaling_factor * pos_freqs)
182
+
183
+ low, high = _yarn_find_correction_range(self.beta_fast, self.beta_slow, self.dim, self.base, self.original_max_position_embeddings)
184
+ inv_freq_mask = (1 - _yarn_linear_ramp_mask(low, high, self.dim // 2).float().to(device)) * self.extrapolation_factor # Get n-d rotational scaling corrected for extrapolation
185
+ inv_freq = inv_freq_interpolation * (1 - inv_freq_mask) + inv_freq_extrapolation * inv_freq_mask
186
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
187
+
188
+ def _compute_inv_freq_original(self, device=None):
189
+ inv_freq = 1 / (self.base ** (torch.arange(0, self.dim, 2, device=device,
190
+ dtype=torch.float32) / self.dim))
191
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
192
+
193
+ def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
194
+ # Reset the tables if the sequence length has changed,
195
+ # if we're on a new device (possibly due to tracing for instance),
196
+ # or if we're switching from inference mode to training
197
+ if (seqlen > self._seq_len_cached or self._cos_cached.device != device
198
+ or self._cos_cached.dtype != dtype
199
+ or (self.training and self._cos_cached.is_inference())):
200
+ self._seq_len_cached = seqlen
201
+
202
+ if self.dynamic:
203
+ scaling_factor = None
204
+ if seqlen <= self.max_position_embeddings:
205
+ if self.finetuned:
206
+ scaling_factor = self.scaling_factor
207
+ else:
208
+ scaling_factor = seqlen / self.original_max_position_embeddings
209
+ if scaling_factor:
210
+ self._compute_inv_freq(scaling_factor, device)
211
+ self.mscale = float(_yarn_get_mscale(scaling_factor) * self.attn_factor)
212
+ else:
213
+ self._compute_inv_freq_original(device)
214
+
215
+ # We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
216
+ # And the output of arange can be quite large, so bf16 would lose a lot of precision.
217
+ # However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
218
+ if self.pos_idx_in_fp32:
219
+ t = torch.arange(seqlen, device=device, dtype=torch.float32)
220
+ # We want fp32 here as well since inv_freq will be multiplied with t, and the output
221
+ # will be large. Having it in bf16 will lose a lot of precision and cause the
222
+ # cos & sin output to change significantly.
223
+ # We want to recompute self.inv_freq if it was not loaded in fp32
224
+ if self.inv_freq.dtype != torch.float32:
225
+ inv_freq = self.inv_freq.to(torch.float32)
226
+ else:
227
+ inv_freq = self.inv_freq
228
+ else:
229
+ t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
230
+ inv_freq = self.inv_freq
231
+ # Don't do einsum, it converts fp32 to fp16 under AMP
232
+ # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
233
+ freqs = torch.outer(t, inv_freq)
234
+ self._cos_cached = (torch.cos(freqs) * self.mscale).to(dtype)
235
+ self._sin_cached = (torch.sin(freqs) * self.mscale).to(dtype)
236
+
237
+
238
+ def forward(self, q: torch.Tensor, k: torch.Tensor, seqlen_offset: int = 0) -> Tuple[torch.Tensor, torch.Tensor]:
239
+ """
240
+ q: (batch, seqlen, nheads, headdim)
241
+ k: (batch, seqlen, nheads, headdim)
242
+ seqlen_offset: can be used in generation where the qkv being passed in is only the last
243
+ token in the batch.
244
+ """
245
+ self._update_cos_sin_cache(q.shape[1] + seqlen_offset, device=q.device, dtype=q.dtype)
246
+ return apply_rotary_emb_func(
247
+ q, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:],
248
+ self.interleaved, True # inplace=True
249
+ ), apply_rotary_emb_func(
250
+ k, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:],
251
+ self.interleaved, True # inplace=True
252
+ )
253
+
254
+
255
+ class FlashRotaryEmbedding(torch.nn.Module):
256
+ """
257
+ The rotary position embeddings from RoFormer_ (Su et. al).
258
+ A crucial insight from the method is that the query and keys are
259
+ transformed by rotation matrices which depend on the relative positions.
260
+ Other implementations are available in the Rotary Transformer repo_ and in
261
+ GPT-NeoX_, GPT-NeoX was an inspiration
262
+ .. _RoFormer: https://arxiv.org/abs/2104.09864
263
+ .. _repo: https://github.com/ZhuiyiTechnology/roformer
264
+ .. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
265
+ If scale_base is not None, this implements XPos (Sun et al., https://arxiv.org/abs/2212.10554).
266
+ A recommended value for scale_base is 512: https://github.com/HazyResearch/flash-attention/issues/96
267
+ Reference: https://github.com/sunyt32/torchscale/blob/main/torchscale/component/xpos_relative_position.py
268
+ """
269
+
270
+ def __init__(self, dim: int, base=10000.0, interleaved=False, scale_base=None,
271
+ scaling_factor=1.0, pos_idx_in_fp32=True, device=None):
272
+ """
273
+ interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
274
+ of 1st half and 2nd half (GPT-NeoX style).
275
+ pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
276
+ otherwise they might be in lower precision.
277
+ This option was added because previously (before 2023-07-02), when we construct
278
+ the position indices, we use the dtype of self.inv_freq. In most cases this would
279
+ be fp32, but if the model is trained in pure bf16 (not mixed precision), then
280
+ self.inv_freq would be bf16, and the position indices are also in bf16.
281
+ Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
282
+ embeddings for some positions will coincide.
283
+ To maintain compatibility with models previously trained in pure bf16,
284
+ we add this option.
285
+ scaling_factor: RotaryEmbedding extended with linear scaling.
286
+ """
287
+ super().__init__()
288
+ self.dim = dim
289
+ self.base = float(base)
290
+ self.pos_idx_in_fp32 = pos_idx_in_fp32
291
+ # Generate and save the inverse frequency buffer (non trainable)
292
+ inv_freq = self._compute_inv_freq(device)
293
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
294
+ self.interleaved = interleaved
295
+ self.scale_base = scale_base
296
+ self.scaling_factor = scaling_factor
297
+ scale = ((torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim)
298
+ / (1.4 * dim) if scale_base is not None else None)
299
+ self.register_buffer("scale", scale)
300
+
301
+ self._seq_len_cached = 0
302
+ self._cos_cached = None
303
+ self._sin_cached = None
304
+ self._cos_k_cached = None
305
+ self._sin_k_cached = None
306
+
307
+ def _compute_inv_freq(self, device=None):
308
+ return 1 / (self.base ** (torch.arange(0, self.dim, 2, device=device,
309
+ dtype=torch.float32) / self.dim))
310
+
311
+
312
+ def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
313
+ # Reset the tables if the sequence length has changed,
314
+ # if we're on a new device (possibly due to tracing for instance),
315
+ # or if we're switching from inference mode to training
316
+ if (seqlen > self._seq_len_cached or self._cos_cached.device != device
317
+ or self._cos_cached.dtype != dtype
318
+ or (self.training and self._cos_cached.is_inference())):
319
+ self._seq_len_cached = seqlen
320
+ # We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
321
+ # And the output of arange can be quite large, so bf16 would lose a lot of precision.
322
+ # However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
323
+ if self.pos_idx_in_fp32:
324
+ t = torch.arange(seqlen, device=device, dtype=torch.float32)
325
+ t /= self.scaling_factor
326
+ # We want fp32 here as well since inv_freq will be multiplied with t, and the output
327
+ # will be large. Having it in bf16 will lose a lot of precision and cause the
328
+ # cos & sin output to change significantly.
329
+ # We want to recompute self.inv_freq if it was not loaded in fp32
330
+ if self.inv_freq.dtype != torch.float32:
331
+ inv_freq = self.inv_freq.to(torch.float32)
332
+ else:
333
+ inv_freq = self.inv_freq
334
+ else:
335
+ t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
336
+ t /= self.scaling_factor
337
+ inv_freq = self.inv_freq
338
+ # Don't do einsum, it converts fp32 to fp16 under AMP
339
+ # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
340
+ freqs = torch.outer(t, inv_freq)
341
+ if self.scale is None:
342
+ self._cos_cached = torch.cos(freqs).to(dtype)
343
+ self._sin_cached = torch.sin(freqs).to(dtype)
344
+ else:
345
+ power = ((torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device)
346
+ - seqlen // 2) / self.scale_base)
347
+ scale = self.scale.to(device=power.device) ** power.unsqueeze(-1)
348
+ # We want the multiplication by scale to happen in fp32
349
+ self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
350
+ self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
351
+ self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
352
+ self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
353
+
354
+ def forward(self, q: torch.Tensor, k: torch.Tensor, seqlen_offset: int = 0) -> Tuple[torch.Tensor, torch.Tensor]:
355
+ """
356
+ q: (batch, seqlen, nheads, headdim)
357
+ k: (batch, seqlen, nheads, headdim)
358
+ seqlen_offset: can be used in generation where the qkv being passed in is only the last
359
+ token in the batch.
360
+ """
361
+ self._update_cos_sin_cache(q.shape[1] + seqlen_offset, device=q.device, dtype=q.dtype)
362
+ if self.scale is None:
363
+ return apply_rotary_emb_func(
364
+ q, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:],
365
+ self.interleaved, True # inplace=True
366
+ ), apply_rotary_emb_func(
367
+ k, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:],
368
+ self.interleaved, True # inplace=True
369
+ )
370
+ else:
371
+ assert False
372
+
373
+ class LlamaMLP(nn.Module):
374
+ def __init__(self, config):
375
+ super().__init__()
376
+ self.config = config
377
+ self.hidden_size = config.hidden_size
378
+ self.intermediate_size = config.intermediate_size
379
+ self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
380
+ self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
381
+ self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
382
+ self.act_fn = ACT2FN[config.hidden_act]
383
+
384
+ def forward(self, x):
385
+ if self.config.pretraining_tp > 1:
386
+ slice = self.intermediate_size // self.config.pretraining_tp
387
+ gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
388
+ up_proj_slices = self.up_proj.weight.split(slice, dim=0)
389
+ down_proj_slices = self.down_proj.weight.split(slice, dim=1)
390
+
391
+ gate_proj = torch.cat(
392
+ [F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
393
+ )
394
+ up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
395
+
396
+ intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
397
+ down_proj = [
398
+ F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
399
+ ]
400
+ down_proj = sum(down_proj)
401
+ else:
402
+ down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
403
+
404
+ return down_proj
405
+
406
+ @torch.jit.script
407
+ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
408
+ """
409
+ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
410
+ num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
411
+ """
412
+ batch, slen, _, num_key_value_heads, head_dim = hidden_states.shape
413
+ if n_rep == 1:
414
+ return hidden_states
415
+ hidden_states = hidden_states[:, :, :, :, None, :].expand(batch, slen, 2, num_key_value_heads, n_rep, head_dim)
416
+ return hidden_states.reshape(batch, slen, 2, num_key_value_heads * n_rep, head_dim)
417
+
418
+
419
+ class LlamaAttention(nn.Module):
420
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
421
+
422
+ def __init__(self, config: LlamaConfig):
423
+ super().__init__()
424
+ self.config = config
425
+ self.hidden_size = config.hidden_size
426
+ self.num_heads = config.num_attention_heads
427
+ self.head_dim = self.hidden_size // self.num_heads
428
+ self.num_key_value_heads = config.num_key_value_heads
429
+ self.num_key_value_groups = self.num_heads // self.num_key_value_heads
430
+ self.max_position_embeddings = config.max_position_embeddings
431
+
432
+ if (self.head_dim * self.num_heads) != self.hidden_size:
433
+ raise ValueError(
434
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
435
+ f" and `num_heads`: {self.num_heads})."
436
+ )
437
+ self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
438
+ self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
439
+ self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
440
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
441
+
442
+ self.register_buffer(
443
+ "norm_factor",
444
+ torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype()),
445
+ persistent=False,
446
+ )
447
+
448
+ if self.config.rope_scaling is None:
449
+ scaling_type = "linear"
450
+ scaling_factor = 1.0
451
+ else:
452
+ scaling_type = self.config.rope_scaling["type"]
453
+ scaling_factor = self.config.rope_scaling["factor"]
454
+ if scaling_type == "yarn" or scaling_type == "dynamic-yarn":
455
+ original_max_position_embeddings = self.config.rope_scaling["original_max_position_embeddings"]
456
+
457
+ self.rotary_emb = FlashYaRNRotaryEmbedding(
458
+ self.head_dim, base=10000, interleaved=False, scaling_factor=scaling_factor,
459
+ max_position_embeddings=self.max_position_embeddings,
460
+ original_max_position_embeddings=original_max_position_embeddings,
461
+ dynamic=scaling_type.startswith("dynamic"), finetuned=self.config.rope_scaling.get("finetuned", False)
462
+ )
463
+ elif scaling_type == "linear":
464
+ self.rotary_emb = FlashRotaryEmbedding(
465
+ self.head_dim, base=10000, interleaved=False, scaling_factor=scaling_factor,
466
+ )
467
+ else:
468
+ raise RuntimeError(f"Unknown scaling type {scaling_type}")
469
+
470
+ def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
471
+ return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
472
+
473
+ def forward(
474
+ self,
475
+ hidden_states: torch.Tensor,
476
+ attention_mask: Optional[torch.Tensor] = None,
477
+ position_ids: Optional[torch.LongTensor] = None,
478
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
479
+ output_attentions: bool = False,
480
+ use_cache: bool = False,
481
+ is_padded_inputs: Optional[bool] = False,
482
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
483
+ bsz, q_len, h_size = hidden_states.size()
484
+
485
+ has_layer_past = past_key_value is not None
486
+
487
+ if has_layer_past:
488
+ past_kv = past_key_value[0]
489
+ past_len = past_key_value[1]
490
+ else:
491
+ past_len = 0
492
+
493
+ if self.config.pretraining_tp > 1:
494
+ key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
495
+ query_slices = self.q_proj.weight.split(
496
+ (self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
497
+ )
498
+ key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
499
+ value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
500
+
501
+ q = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
502
+ q = torch.cat(q, dim=-1)
503
+
504
+ k = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
505
+ k = torch.cat(k, dim=-1)
506
+
507
+ v = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
508
+ v = torch.cat(v, dim=-1)
509
+
510
+ else:
511
+ q = self.q_proj(hidden_states)
512
+ k = self.k_proj(hidden_states)
513
+ v = self.v_proj(hidden_states)
514
+
515
+ q = q.view(bsz, q_len, self.num_heads, self.head_dim)
516
+ k = k.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
517
+ v = v.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
518
+
519
+ q, k = self.rotary_emb(q, k, past_len)
520
+
521
+ kv = torch.stack([k, v], 2)
522
+ kv = repeat_kv(kv, self.num_key_value_groups)
523
+
524
+ # Cache QKV values
525
+ if has_layer_past:
526
+ new_len = past_len+q.size(1)
527
+ if new_len > past_kv.size(1):
528
+ past_kv = torch.cat([past_kv, torch.empty(bsz, 256, 2, kv.size(3), kv.size(4), dtype=kv.dtype, device=kv.device)], 1)
529
+ past_kv[:, past_len:new_len] = kv
530
+ kv = past_kv[:, :new_len]
531
+ else:
532
+ past_kv = kv
533
+
534
+ past_key_value = (past_kv, past_len+q.size(1)) if use_cache else None
535
+
536
+ if is_padded_inputs:
537
+
538
+ # varlen, ignore padding tokens, efficient for large batch with many paddings
539
+
540
+ assert attention_mask is not None
541
+
542
+ unpadded_kv, indices_k, cu_seqlens_k, max_seqlen_k = unpad_input(kv, attention_mask)
543
+ unpadded_q, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, attention_mask[:, -q.size(1):])
544
+ attn_outputs = flash_attn_varlen_kvpacked_func(
545
+ unpadded_q, unpadded_kv, cu_seqlens_q, cu_seqlens_k,
546
+ max_seqlen_q, max_seqlen_k,
547
+ dropout_p=0.0, softmax_scale=1.0/self.norm_factor,
548
+ causal=(not has_layer_past), return_attn_probs=output_attentions
549
+ )
550
+
551
+ attn_output = attn_outputs[0] if output_attentions else attn_outputs
552
+ attn_output = pad_input(
553
+ attn_output, indices_q, bsz, q_len
554
+ ).reshape(bsz, q_len, h_size)
555
+ attn_weights = attn_outputs[2] if output_attentions else None
556
+
557
+ else:
558
+
559
+ # no padding tokens, more efficient
560
+
561
+ attn_outputs = flash_attn_kvpacked_func(
562
+ q, kv, dropout_p=0.0, softmax_scale=1.0/self.norm_factor, causal=(not has_layer_past), return_attn_probs=output_attentions)
563
+
564
+ attn_output = attn_outputs[0] if output_attentions else attn_outputs
565
+ attn_output = attn_output.reshape(bsz, q_len, h_size)
566
+ attn_weights = attn_outputs[2] if output_attentions else None
567
+
568
+ if self.config.pretraining_tp > 1:
569
+ attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
570
+ o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
571
+ attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
572
+ else:
573
+ attn_output = self.o_proj(attn_output)
574
+
575
+ if not output_attentions:
576
+ attn_weights = None
577
+
578
+ return attn_output, attn_weights, past_key_value
579
+
580
+
581
+ class LlamaDecoderLayer(nn.Module):
582
+ def __init__(self, config: LlamaConfig):
583
+ super().__init__()
584
+ self.hidden_size = config.hidden_size
585
+ self.self_attn = LlamaAttention(config=config)
586
+ self.mlp = LlamaMLP(config)
587
+ self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
588
+ self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
589
+
590
+ def forward(
591
+ self,
592
+ hidden_states: torch.Tensor,
593
+ attention_mask: Optional[torch.Tensor] = None,
594
+ position_ids: Optional[torch.LongTensor] = None,
595
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
596
+ is_padded_inputs: Optional[bool] = False,
597
+ output_attentions: Optional[bool] = False,
598
+ use_cache: Optional[bool] = False,
599
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
600
+ """
601
+ Args:
602
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
603
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
604
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
605
+ output_attentions (`bool`, *optional*):
606
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
607
+ returned tensors for more detail.
608
+ use_cache (`bool`, *optional*):
609
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
610
+ (see `past_key_values`).
611
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
612
+ """
613
+
614
+ residual = hidden_states
615
+
616
+ hidden_states = self.input_layernorm(hidden_states)
617
+
618
+ # Self Attention
619
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
620
+ hidden_states=hidden_states,
621
+ attention_mask=attention_mask,
622
+ position_ids=position_ids,
623
+ past_key_value=past_key_value,
624
+ output_attentions=output_attentions,
625
+ use_cache=use_cache,
626
+ is_padded_inputs=is_padded_inputs,
627
+ )
628
+ hidden_states = residual + hidden_states
629
+
630
+ # Fully Connected
631
+ residual = hidden_states
632
+ hidden_states = self.post_attention_layernorm(hidden_states)
633
+ hidden_states = self.mlp(hidden_states)
634
+ hidden_states = residual + hidden_states
635
+
636
+ outputs = (hidden_states,)
637
+
638
+ if output_attentions:
639
+ outputs += (self_attn_weights,)
640
+
641
+ if use_cache:
642
+ outputs += (present_key_value,)
643
+
644
+ return outputs
645
+
646
+
647
+ LLAMA_START_DOCSTRING = r"""
648
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
649
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
650
+ etc.)
651
+
652
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
653
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
654
+ and behavior.
655
+
656
+ Parameters:
657
+ config ([`LlamaConfig`]):
658
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
659
+ load the weights associated with the model, only the configuration. Check out the
660
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
661
+ """
662
+
663
+
664
+ @add_start_docstrings(
665
+ "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
666
+ LLAMA_START_DOCSTRING,
667
+ )
668
+ class LlamaPreTrainedModel(PreTrainedModel):
669
+ config_class = LlamaConfig
670
+ base_model_prefix = "model"
671
+ supports_gradient_checkpointing = True
672
+ _no_split_modules = ["LlamaDecoderLayer"]
673
+ _skip_keys_device_placement = "past_key_values"
674
+
675
+ def _init_weights(self, module):
676
+ std = self.config.initializer_range
677
+ if isinstance(module, nn.Linear):
678
+ module.weight.data.normal_(mean=0.0, std=std)
679
+ if module.bias is not None:
680
+ module.bias.data.zero_()
681
+ elif isinstance(module, nn.Embedding):
682
+ module.weight.data.normal_(mean=0.0, std=std)
683
+ if module.padding_idx is not None:
684
+ module.weight.data[module.padding_idx].zero_()
685
+
686
+ def _set_gradient_checkpointing(self, module, value=False):
687
+ if isinstance(module, LlamaModel):
688
+ module.gradient_checkpointing = value
689
+
690
+
691
+ LLAMA_INPUTS_DOCSTRING = r"""
692
+ Args:
693
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
694
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
695
+ it.
696
+
697
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
698
+ [`PreTrainedTokenizer.__call__`] for details.
699
+
700
+ [What are input IDs?](../glossary#input-ids)
701
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
702
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
703
+
704
+ - 1 for tokens that are **not masked**,
705
+ - 0 for tokens that are **masked**.
706
+
707
+ [What are attention masks?](../glossary#attention-mask)
708
+
709
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
710
+ [`PreTrainedTokenizer.__call__`] for details.
711
+
712
+ If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
713
+ `past_key_values`).
714
+
715
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
716
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
717
+ information on the default strategy.
718
+
719
+ - 1 indicates the head is **not masked**,
720
+ - 0 indicates the head is **masked**.
721
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
722
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
723
+ config.n_positions - 1]`.
724
+
725
+ [What are position IDs?](../glossary#position-ids)
726
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
727
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
728
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
729
+ `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
730
+
731
+ Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
732
+ blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
733
+
734
+ If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
735
+ don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
736
+ `decoder_input_ids` of shape `(batch_size, sequence_length)`.
737
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
738
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
739
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
740
+ model's internal embedding lookup matrix.
741
+ use_cache (`bool`, *optional*):
742
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
743
+ `past_key_values`).
744
+ output_attentions (`bool`, *optional*):
745
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
746
+ tensors for more detail.
747
+ output_hidden_states (`bool`, *optional*):
748
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
749
+ more detail.
750
+ return_dict (`bool`, *optional*):
751
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
752
+ """
753
+
754
+
755
+ @add_start_docstrings(
756
+ "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
757
+ LLAMA_START_DOCSTRING,
758
+ )
759
+ class LlamaModel(LlamaPreTrainedModel):
760
+ """
761
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
762
+
763
+ Args:
764
+ config: LlamaConfig
765
+ """
766
+
767
+ def __init__(self, config: LlamaConfig):
768
+ super().__init__(config)
769
+ self.padding_idx = config.pad_token_id
770
+ self.vocab_size = config.vocab_size
771
+
772
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
773
+ self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
774
+ self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
775
+
776
+ self.gradient_checkpointing = False
777
+ # Initialize weights and apply final processing
778
+ self.post_init()
779
+
780
+ def get_input_embeddings(self):
781
+ return self.embed_tokens
782
+
783
+ def set_input_embeddings(self, value):
784
+ self.embed_tokens = value
785
+
786
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
787
+ def forward(
788
+ self,
789
+ input_ids: torch.LongTensor = None,
790
+ attention_mask: Optional[torch.Tensor] = None,
791
+ position_ids: Optional[torch.LongTensor] = None,
792
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
793
+ inputs_embeds: Optional[torch.FloatTensor] = None,
794
+ use_cache: Optional[bool] = None,
795
+ output_attentions: Optional[bool] = None,
796
+ output_hidden_states: Optional[bool] = None,
797
+ return_dict: Optional[bool] = None,
798
+ is_padded_inputs: Optional[bool] = False,
799
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
800
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
801
+ output_hidden_states = (
802
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
803
+ )
804
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
805
+
806
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
807
+
808
+ # retrieve input_ids and inputs_embeds
809
+ if input_ids is not None and inputs_embeds is not None:
810
+ raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
811
+ elif input_ids is not None:
812
+ batch_size, seq_length = input_ids.shape
813
+ elif inputs_embeds is not None:
814
+ batch_size, seq_length, _ = inputs_embeds.shape
815
+ else:
816
+ raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
817
+
818
+ seq_length_with_past = seq_length
819
+ past_key_values_length = 0
820
+
821
+ if past_key_values is not None:
822
+ past_key_values_length = past_key_values[0][0].shape[2]
823
+ seq_length_with_past = seq_length_with_past + past_key_values_length
824
+
825
+ position_ids = None
826
+
827
+ if inputs_embeds is None:
828
+ inputs_embeds = self.embed_tokens(input_ids)
829
+
830
+ hidden_states = inputs_embeds
831
+
832
+ if self.gradient_checkpointing and self.training:
833
+ if use_cache:
834
+ logger.warning_once(
835
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
836
+ )
837
+ use_cache = False
838
+
839
+ # decoder layers
840
+ all_hidden_states = () if output_hidden_states else None
841
+ all_self_attns = () if output_attentions else None
842
+ next_decoder_cache = () if use_cache else None
843
+
844
+ for idx, decoder_layer in enumerate(self.layers):
845
+ if output_hidden_states:
846
+ all_hidden_states += (hidden_states,)
847
+
848
+ past_key_value = past_key_values[idx] if past_key_values is not None else None
849
+
850
+ if self.gradient_checkpointing and self.training:
851
+
852
+ def create_custom_forward(module):
853
+ def custom_forward(*inputs):
854
+ # None for past_key_value
855
+ return module(*inputs, output_attentions, None)
856
+
857
+ return custom_forward
858
+
859
+ layer_outputs = torch.utils.checkpoint.checkpoint(
860
+ create_custom_forward(decoder_layer),
861
+ hidden_states,
862
+ attention_mask,
863
+ position_ids,
864
+ None,
865
+ is_padded_inputs
866
+ )
867
+ else:
868
+ layer_outputs = decoder_layer(
869
+ hidden_states,
870
+ attention_mask=attention_mask,
871
+ position_ids=position_ids,
872
+ past_key_value=past_key_value,
873
+ output_attentions=output_attentions,
874
+ use_cache=use_cache,
875
+ is_padded_inputs=is_padded_inputs,
876
+ )
877
+
878
+ hidden_states = layer_outputs[0]
879
+
880
+ if use_cache:
881
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
882
+
883
+ if output_attentions:
884
+ all_self_attns += (layer_outputs[1],)
885
+
886
+ hidden_states = self.norm(hidden_states)
887
+
888
+ # add hidden states from the last decoder layer
889
+ if output_hidden_states:
890
+ all_hidden_states += (hidden_states,)
891
+
892
+ next_cache = next_decoder_cache if use_cache else None
893
+ if not return_dict:
894
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
895
+ return BaseModelOutputWithPast(
896
+ last_hidden_state=hidden_states,
897
+ past_key_values=next_cache,
898
+ hidden_states=all_hidden_states,
899
+ attentions=all_self_attns,
900
+ )
901
+
902
+
903
+ class LlamaForCausalLM(LlamaPreTrainedModel):
904
+ _tied_weights_keys = ["lm_head.weight"]
905
+
906
+ def __init__(self, config):
907
+ super().__init__(config)
908
+ self.model = LlamaModel(config)
909
+ self.vocab_size = config.vocab_size
910
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
911
+
912
+ # Initialize weights and apply final processing
913
+ self.post_init()
914
+
915
+ def get_input_embeddings(self):
916
+ return self.model.embed_tokens
917
+
918
+ def set_input_embeddings(self, value):
919
+ self.model.embed_tokens = value
920
+
921
+ def get_output_embeddings(self):
922
+ return self.lm_head
923
+
924
+ def set_output_embeddings(self, new_embeddings):
925
+ self.lm_head = new_embeddings
926
+
927
+ def set_decoder(self, decoder):
928
+ self.model = decoder
929
+
930
+ def get_decoder(self):
931
+ return self.model
932
+
933
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
934
+ @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
935
+ def forward(
936
+ self,
937
+ input_ids: torch.LongTensor = None,
938
+ attention_mask: Optional[torch.Tensor] = None,
939
+ position_ids: Optional[torch.LongTensor] = None,
940
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
941
+ inputs_embeds: Optional[torch.FloatTensor] = None,
942
+ labels: Optional[torch.LongTensor] = None,
943
+ use_cache: Optional[bool] = None,
944
+ output_attentions: Optional[bool] = None,
945
+ output_hidden_states: Optional[bool] = None,
946
+ return_dict: Optional[bool] = None,
947
+ is_padded_inputs: Optional[bool] = None,
948
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
949
+ r"""
950
+ Args:
951
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
952
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
953
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
954
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
955
+
956
+ Returns:
957
+
958
+ Example:
959
+
960
+ ```python
961
+ >>> from transformers import AutoTokenizer, LlamaForCausalLM
962
+
963
+ >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
964
+ >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
965
+
966
+ >>> prompt = "Hey, are you conscious? Can you talk to me?"
967
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
968
+
969
+ >>> # Generate
970
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
971
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
972
+ "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
973
+ ```"""
974
+
975
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
976
+ output_hidden_states = (
977
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
978
+ )
979
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
980
+
981
+ is_padded_inputs = ((attention_mask is not None) and (not attention_mask.all().item()))
982
+
983
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
984
+ outputs = self.model(
985
+ input_ids=input_ids,
986
+ attention_mask=attention_mask,
987
+ position_ids=position_ids,
988
+ past_key_values=past_key_values,
989
+ inputs_embeds=inputs_embeds,
990
+ use_cache=use_cache,
991
+ output_attentions=output_attentions,
992
+ output_hidden_states=output_hidden_states,
993
+ return_dict=return_dict,
994
+ is_padded_inputs=is_padded_inputs,
995
+ )
996
+
997
+ hidden_states = outputs[0]
998
+ if self.config.pretraining_tp > 1:
999
+ lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.config.pretraining_tp, dim=0)
1000
+ logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.config.pretraining_tp)]
1001
+ logits = torch.cat(logits, dim=-1)
1002
+ else:
1003
+ logits = self.lm_head(hidden_states)
1004
+ logits = logits.float()
1005
+
1006
+ loss = None
1007
+ if labels is not None:
1008
+ # Shift so that tokens < n predict n
1009
+ shift_logits = logits[..., :-1, :].contiguous()
1010
+ shift_labels = labels[..., 1:].contiguous()
1011
+ # Flatten the tokens
1012
+ loss_fct = CrossEntropyLoss()
1013
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
1014
+ shift_labels = shift_labels.view(-1)
1015
+ # Enable model parallelism
1016
+ shift_labels = shift_labels.to(shift_logits.device)
1017
+ loss = loss_fct(shift_logits, shift_labels)
1018
+
1019
+ if not return_dict:
1020
+ output = (logits,) + outputs[1:]
1021
+ return (loss,) + output if loss is not None else output
1022
+
1023
+ return CausalLMOutputWithPast(
1024
+ loss=loss,
1025
+ logits=logits,
1026
+ past_key_values=outputs.past_key_values,
1027
+ hidden_states=outputs.hidden_states,
1028
+ attentions=outputs.attentions,
1029
+ )
1030
+
1031
+ def prepare_inputs_for_generation(
1032
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
1033
+ ):
1034
+ if past_key_values:
1035
+ input_ids = input_ids[:, -1:]
1036
+
1037
+ position_ids = kwargs.get("position_ids", None)
1038
+
1039
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1040
+ if inputs_embeds is not None and past_key_values is None:
1041
+ model_inputs = {"inputs_embeds": inputs_embeds}
1042
+ else:
1043
+ model_inputs = {"input_ids": input_ids}
1044
+
1045
+ model_inputs.update(
1046
+ {
1047
+ "position_ids": position_ids,
1048
+ "past_key_values": past_key_values,
1049
+ "use_cache": kwargs.get("use_cache"),
1050
+ "attention_mask": attention_mask,
1051
+ "is_padded_inputs": ((attention_mask is not None) and (not attention_mask.all().item()))
1052
+ }
1053
+ )
1054
+ return model_inputs
1055
+
1056
+ @staticmethod
1057
+ def _reorder_cache(past_key_values, beam_idx):
1058
+ reordered_past = ()
1059
+ for layer_past in past_key_values:
1060
+ reordered_past += (
1061
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
1062
+ )
1063
+ return reordered_past
1064
+
1065
+
1066
+ @add_start_docstrings(
1067
+ """
1068
+ The LLaMa Model transformer with a sequence classification head on top (linear layer).
1069
+
1070
+ [`LlamaForSequenceClassification`] uses the last token in order to do the classification, as other causal models
1071
+ (e.g. GPT-2) do.
1072
+
1073
+ Since it does classification on the last token, it requires to know the position of the last token. If a
1074
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
1075
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
1076
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
1077
+ each row of the batch).
1078
+ """,
1079
+ LLAMA_START_DOCSTRING,
1080
+ )
1081
+ class LlamaForSequenceClassification(LlamaPreTrainedModel):
1082
+ def __init__(self, config):
1083
+ super().__init__(config)
1084
+ self.num_labels = config.num_labels
1085
+ self.model = LlamaModel(config)
1086
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
1087
+
1088
+ # Initialize weights and apply final processing
1089
+ self.post_init()
1090
+
1091
+ def get_input_embeddings(self):
1092
+ return self.model.embed_tokens
1093
+
1094
+ def set_input_embeddings(self, value):
1095
+ self.model.embed_tokens = value
1096
+
1097
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
1098
+ def forward(
1099
+ self,
1100
+ input_ids: torch.LongTensor = None,
1101
+ attention_mask: Optional[torch.Tensor] = None,
1102
+ position_ids: Optional[torch.LongTensor] = None,
1103
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1104
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1105
+ labels: Optional[torch.LongTensor] = None,
1106
+ use_cache: Optional[bool] = None,
1107
+ output_attentions: Optional[bool] = None,
1108
+ output_hidden_states: Optional[bool] = None,
1109
+ return_dict: Optional[bool] = None,
1110
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1111
+ r"""
1112
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1113
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1114
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1115
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1116
+ """
1117
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1118
+
1119
+ transformer_outputs = self.model(
1120
+ input_ids,
1121
+ attention_mask=attention_mask,
1122
+ position_ids=position_ids,
1123
+ past_key_values=past_key_values,
1124
+ inputs_embeds=inputs_embeds,
1125
+ use_cache=use_cache,
1126
+ output_attentions=output_attentions,
1127
+ output_hidden_states=output_hidden_states,
1128
+ return_dict=return_dict,
1129
+ )
1130
+ hidden_states = transformer_outputs[0]
1131
+ logits = self.score(hidden_states)
1132
+
1133
+ if input_ids is not None:
1134
+ batch_size = input_ids.shape[0]
1135
+ else:
1136
+ batch_size = inputs_embeds.shape[0]
1137
+
1138
+ if self.config.pad_token_id is None and batch_size != 1:
1139
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
1140
+ if self.config.pad_token_id is None:
1141
+ sequence_lengths = -1
1142
+ else:
1143
+ if input_ids is not None:
1144
+ sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
1145
+ else:
1146
+ sequence_lengths = -1
1147
+
1148
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1149
+
1150
+ loss = None
1151
+ if labels is not None:
1152
+ labels = labels.to(logits.device)
1153
+ if self.config.problem_type is None:
1154
+ if self.num_labels == 1:
1155
+ self.config.problem_type = "regression"
1156
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1157
+ self.config.problem_type = "single_label_classification"
1158
+ else:
1159
+ self.config.problem_type = "multi_label_classification"
1160
+
1161
+ if self.config.problem_type == "regression":
1162
+ loss_fct = MSELoss()
1163
+ if self.num_labels == 1:
1164
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1165
+ else:
1166
+ loss = loss_fct(pooled_logits, labels)
1167
+ elif self.config.problem_type == "single_label_classification":
1168
+ loss_fct = CrossEntropyLoss()
1169
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1170
+ elif self.config.problem_type == "multi_label_classification":
1171
+ loss_fct = BCEWithLogitsLoss()
1172
+ loss = loss_fct(pooled_logits, labels)
1173
+ if not return_dict:
1174
+ output = (pooled_logits,) + transformer_outputs[1:]
1175
+ return ((loss,) + output) if loss is not None else output
1176
+
1177
+ return SequenceClassifierOutputWithPast(
1178
+ loss=loss,
1179
+ logits=pooled_logits,
1180
+ past_key_values=transformer_outputs.past_key_values,
1181
+ hidden_states=transformer_outputs.hidden_states,
1182
+ attentions=transformer_outputs.attentions,
1183
+ )