bjoernp commited on
Commit
1ef978a
1 Parent(s): 3563a93

Create modeling_flash_llama.py

Browse files
Files changed (1) hide show
  1. modeling_flash_llama.py +1010 -0
modeling_flash_llama.py ADDED
@@ -0,0 +1,1010 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # From https://huggingface.co/togethercomputer/LLaMA-2-7B-32K/blob/main/modeling_flash_llama.py
3
+ # With fix from Alex Birch: https://huggingface.co/togethercomputer/LLaMA-2-7B-32K/discussions/17
4
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
5
+ #
6
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
7
+ # and OPT implementations in this library. It has been modified from its
8
+ # original forms to accommodate minor architectural differences compared
9
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
10
+ #
11
+ # Licensed under the Apache License, Version 2.0 (the "License");
12
+ # you may not use this file except in compliance with the License.
13
+ # You may obtain a copy of the License at
14
+ #
15
+ # http://www.apache.org/licenses/LICENSE-2.0
16
+ #
17
+ # Unless required by applicable law or agreed to in writing, software
18
+ # distributed under the License is distributed on an "AS IS" BASIS,
19
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
20
+ # See the License for the specific language governing permissions and
21
+ # limitations under the License.
22
+ """ PyTorch LLaMA model."""
23
+ from typing import List, Optional, Tuple, Union
24
+
25
+ import torch
26
+ import torch.nn.functional as F
27
+ import torch.utils.checkpoint
28
+ from torch import nn
29
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
30
+
31
+ from transformers.activations import ACT2FN
32
+ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
33
+ from transformers.modeling_utils import PreTrainedModel
34
+ from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
35
+ from transformers.models.llama.configuration_llama import LlamaConfig
36
+
37
+
38
+ try:
39
+ from flash_attn.flash_attn_interface import (
40
+ flash_attn_kvpacked_func,
41
+ flash_attn_varlen_kvpacked_func,
42
+ )
43
+ from flash_attn.bert_padding import unpad_input, pad_input
44
+ flash_attn_v2_installed = True
45
+ print('>>>> Flash Attention installed')
46
+ except ImportError:
47
+ flash_attn_v2_installed = False
48
+ raise ImportError('Please install Flash Attention: `pip install flash-attn --no-build-isolation`')
49
+
50
+ try:
51
+ from flash_attn.layers.rotary import apply_rotary_emb_func
52
+ flash_rope_installed = True
53
+ print('>>>> Flash RoPE installed')
54
+ except ImportError:
55
+ flash_rope_installed = False
56
+ raise ImportError('Please install RoPE kernels: `pip install git+https://github.com/HazyResearch/flash-attention.git#subdirectory=csrc/rotary`')
57
+
58
+
59
+ logger = logging.get_logger(__name__)
60
+
61
+ _CONFIG_FOR_DOC = "LlamaConfig"
62
+
63
+
64
+ # @torch.jit.script
65
+ def rmsnorm_func(hidden_states, weight, variance_epsilon):
66
+ input_dtype = hidden_states.dtype
67
+ hidden_states = hidden_states.to(torch.float32)
68
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
69
+ hidden_states = hidden_states * torch.rsqrt(variance + variance_epsilon)
70
+ return (weight * hidden_states).to(input_dtype)
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.register_buffer(
81
+ "variance_epsilon",
82
+ torch.tensor(eps),
83
+ persistent=False,
84
+ )
85
+
86
+ def forward(self, hidden_states):
87
+ return rmsnorm_func(hidden_states, self.weight, self.variance_epsilon)
88
+
89
+
90
+ class FlashRotaryEmbedding(torch.nn.Module):
91
+ """
92
+ The rotary position embeddings from RoFormer_ (Su et. al).
93
+ A crucial insight from the method is that the query and keys are
94
+ transformed by rotation matrices which depend on the relative positions.
95
+
96
+ Other implementations are available in the Rotary Transformer repo_ and in
97
+ GPT-NeoX_, GPT-NeoX was an inspiration
98
+
99
+ .. _RoFormer: https://arxiv.org/abs/2104.09864
100
+ .. _repo: https://github.com/ZhuiyiTechnology/roformer
101
+ .. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
102
+
103
+ If scale_base is not None, this implements XPos (Sun et al., https://arxiv.org/abs/2212.10554).
104
+ A recommended value for scale_base is 512: https://github.com/HazyResearch/flash-attention/issues/96
105
+ Reference: https://github.com/sunyt32/torchscale/blob/main/torchscale/component/xpos_relative_position.py
106
+ """
107
+
108
+ def __init__(self, dim: int, base=10000.0, interleaved=False, scale_base=None,
109
+ scaling_factor=1.0, pos_idx_in_fp32=True, device=None):
110
+ """
111
+ interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
112
+ of 1st half and 2nd half (GPT-NeoX style).
113
+ pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
114
+ otherwise they might be in lower precision.
115
+ This option was added because previously (before 2023-07-02), when we construct
116
+ the position indices, we use the dtype of self.inv_freq. In most cases this would
117
+ be fp32, but if the model is trained in pure bf16 (not mixed precision), then
118
+ self.inv_freq would be bf16, and the position indices are also in bf16.
119
+ Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
120
+ embeddings for some positions will coincide.
121
+ To maintain compatibility with models previously trained in pure bf16,
122
+ we add this option.
123
+ scaling_factor: RotaryEmbedding extended with linear scaling.
124
+ """
125
+ super().__init__()
126
+ self.dim = dim
127
+ self.base = float(base)
128
+ self.pos_idx_in_fp32 = pos_idx_in_fp32
129
+ # Generate and save the inverse frequency buffer (non trainable)
130
+ inv_freq = self._compute_inv_freq(device)
131
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
132
+ self.interleaved = interleaved
133
+ self.scale_base = scale_base
134
+ self.scaling_factor = scaling_factor
135
+ scale = ((torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim)
136
+ / (1.4 * dim) if scale_base is not None else None)
137
+ self.register_buffer("scale", scale)
138
+
139
+ self._seq_len_cached = 0
140
+ self._cos_cached = None
141
+ self._sin_cached = None
142
+ self._cos_k_cached = None
143
+ self._sin_k_cached = None
144
+
145
+ def _compute_inv_freq(self, device=None):
146
+ return 1 / (self.base ** (torch.arange(0, self.dim, 2, device=device,
147
+ dtype=torch.float32) / self.dim))
148
+
149
+
150
+ def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
151
+ # Reset the tables if the sequence length has changed,
152
+ # if we're on a new device (possibly due to tracing for instance),
153
+ # or if we're switching from inference mode to training
154
+ if (seqlen > self._seq_len_cached or self._cos_cached.device != device
155
+ or self._cos_cached.dtype != dtype
156
+ or (self.training and self._cos_cached.is_inference())):
157
+ self._seq_len_cached = seqlen
158
+ # We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
159
+ # And the output of arange can be quite large, so bf16 would lose a lot of precision.
160
+ # However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
161
+ if self.pos_idx_in_fp32:
162
+ t = torch.arange(seqlen, device=device, dtype=torch.float32)
163
+ t /= self.scaling_factor
164
+ # We want fp32 here as well since inv_freq will be multiplied with t, and the output
165
+ # will be large. Having it in bf16 will lose a lot of precision and cause the
166
+ # cos & sin output to change significantly.
167
+ # We want to recompute self.inv_freq if it was not loaded in fp32
168
+ if self.inv_freq.dtype != torch.float32:
169
+ inv_freq = self.inv_freq.to(torch.float32)
170
+ else:
171
+ inv_freq = self.inv_freq
172
+ else:
173
+ t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
174
+ t /= self.scaling_factor
175
+ inv_freq = self.inv_freq
176
+ # Don't do einsum, it converts fp32 to fp16 under AMP
177
+ # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
178
+ freqs = torch.outer(t, inv_freq)
179
+ if self.scale is None:
180
+ self._cos_cached = torch.cos(freqs).to(dtype)
181
+ self._sin_cached = torch.sin(freqs).to(dtype)
182
+ else:
183
+ power = ((torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device)
184
+ - seqlen // 2) / self.scale_base)
185
+ scale = self.scale.to(device=power.device) ** power.unsqueeze(-1)
186
+ # We want the multiplication by scale to happen in fp32
187
+ self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
188
+ self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
189
+ self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
190
+ self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
191
+
192
+ def forward(self, q: torch.Tensor, k: torch.Tensor, seqlen_offset: int = 0) -> Tuple[torch.Tensor, torch.Tensor]:
193
+ """
194
+ q: (batch, seqlen, nheads, headdim)
195
+ k: (batch, seqlen, nheads, headdim)
196
+ seqlen_offset: can be used in generation where the qkv being passed in is only the last
197
+ token in the batch.
198
+ """
199
+ self._update_cos_sin_cache(q.shape[1] + seqlen_offset, device=q.device, dtype=q.dtype)
200
+ if self.scale is None:
201
+ return apply_rotary_emb_func(
202
+ q, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:],
203
+ self.interleaved, True # inplace=True
204
+ ), apply_rotary_emb_func(
205
+ k, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:],
206
+ self.interleaved, True # inplace=True
207
+ )
208
+ else:
209
+ assert False
210
+
211
+ class LlamaMLP(nn.Module):
212
+ def __init__(self, config):
213
+ super().__init__()
214
+ self.config = config
215
+ self.hidden_size = config.hidden_size
216
+ self.intermediate_size = config.intermediate_size
217
+ self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
218
+ self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
219
+ self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
220
+ self.act_fn = ACT2FN[config.hidden_act]
221
+
222
+ def forward(self, x):
223
+ if self.config.pretraining_tp > 1:
224
+ slice = self.intermediate_size // self.config.pretraining_tp
225
+ gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
226
+ up_proj_slices = self.up_proj.weight.split(slice, dim=0)
227
+ down_proj_slices = self.down_proj.weight.split(slice, dim=1)
228
+
229
+ gate_proj = torch.cat(
230
+ [F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
231
+ )
232
+ up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
233
+
234
+ intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
235
+ down_proj = [
236
+ F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
237
+ ]
238
+ down_proj = sum(down_proj)
239
+ else:
240
+ down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
241
+
242
+ return down_proj
243
+
244
+ @torch.jit.script
245
+ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
246
+ """
247
+ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
248
+ num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
249
+ """
250
+ batch, slen, _, num_key_value_heads, head_dim = hidden_states.shape
251
+ if n_rep == 1:
252
+ return hidden_states
253
+ hidden_states = hidden_states[:, :, :, :, None, :].expand(batch, slen, 2, num_key_value_heads, n_rep, head_dim)
254
+ return hidden_states.reshape(batch, slen, 2, num_key_value_heads * n_rep, head_dim)
255
+
256
+
257
+ class LlamaAttention(nn.Module):
258
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
259
+
260
+ def __init__(self, config: LlamaConfig):
261
+ super().__init__()
262
+ self.config = config
263
+ self.hidden_size = config.hidden_size
264
+ self.num_heads = config.num_attention_heads
265
+ self.head_dim = self.hidden_size // self.num_heads
266
+ self.num_key_value_heads = config.num_key_value_heads
267
+ self.num_key_value_groups = self.num_heads // self.num_key_value_heads
268
+ self.max_position_embeddings = config.max_position_embeddings
269
+
270
+ if (self.head_dim * self.num_heads) != self.hidden_size:
271
+ raise ValueError(
272
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
273
+ f" and `num_heads`: {self.num_heads})."
274
+ )
275
+ self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
276
+ self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
277
+ self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
278
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
279
+
280
+ self.register_buffer(
281
+ "norm_factor",
282
+ torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype()),
283
+ persistent=False,
284
+ )
285
+
286
+ if self.config.rope_scaling is None:
287
+ scaling_factor = 1
288
+ else:
289
+ scaling_type = self.config.rope_scaling["type"]
290
+ scaling_factor = self.config.rope_scaling["factor"]
291
+ assert scaling_type == 'linear'
292
+
293
+ self.rotary_emb = FlashRotaryEmbedding(
294
+ self.head_dim, base=10000, interleaved=False, scaling_factor=scaling_factor,
295
+ )
296
+
297
+ def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
298
+ return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
299
+
300
+ def forward(
301
+ self,
302
+ hidden_states: torch.Tensor,
303
+ attention_mask: Optional[torch.Tensor] = None,
304
+ position_ids: Optional[torch.LongTensor] = None,
305
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
306
+ output_attentions: bool = False,
307
+ use_cache: bool = False,
308
+ is_padded_inputs: Optional[bool] = False,
309
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
310
+ bsz, q_len, h_size = hidden_states.size()
311
+
312
+ has_layer_past = past_key_value is not None
313
+
314
+ if has_layer_past:
315
+ past_kv = past_key_value[0]
316
+ past_len = past_key_value[1]
317
+ else:
318
+ past_len = 0
319
+
320
+ if self.config.pretraining_tp > 1:
321
+ key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
322
+ query_slices = self.q_proj.weight.split(
323
+ (self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
324
+ )
325
+ key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
326
+ value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
327
+
328
+ q = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
329
+ q = torch.cat(q, dim=-1)
330
+
331
+ k = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
332
+ k = torch.cat(k, dim=-1)
333
+
334
+ v = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
335
+ v = torch.cat(v, dim=-1)
336
+
337
+ else:
338
+ q = self.q_proj(hidden_states)
339
+ k = self.k_proj(hidden_states)
340
+ v = self.v_proj(hidden_states)
341
+
342
+ q = q.view(bsz, q_len, self.num_heads, self.head_dim)
343
+ k = k.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
344
+ v = v.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
345
+
346
+ q, k = self.rotary_emb(q, k, past_len)
347
+
348
+ kv = torch.stack([k, v], 2)
349
+ kv = repeat_kv(kv, self.num_key_value_groups)
350
+
351
+ # Cache QKV values
352
+ if has_layer_past:
353
+ new_len = past_len+q.size(1)
354
+ if new_len > past_kv.size(1):
355
+ past_kv = torch.cat([past_kv, torch.empty(bsz, 256, 2, kv.size(3), kv.size(4), dtype=kv.dtype, device=kv.device)], 1)
356
+ past_kv[:, past_len:new_len] = kv
357
+ kv = past_kv[:, :new_len]
358
+ else:
359
+ past_kv = kv
360
+
361
+ past_key_value = (past_kv, past_len+q.size(1)) if use_cache else None
362
+
363
+ if is_padded_inputs:
364
+
365
+ # varlen, ignore padding tokens, efficient for large batch with many paddings
366
+
367
+ assert attention_mask is not None
368
+
369
+ unpadded_kv, indices_k, cu_seqlens_k, max_seqlen_k = unpad_input(kv, attention_mask)
370
+ unpadded_q, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, attention_mask[:, -q.size(1):])
371
+ attn_outputs = flash_attn_varlen_kvpacked_func(
372
+ unpadded_q, unpadded_kv, cu_seqlens_q, cu_seqlens_k,
373
+ max_seqlen_q, max_seqlen_k,
374
+ dropout_p=0.0, softmax_scale=1.0/self.norm_factor,
375
+ causal=(not has_layer_past), return_attn_probs=output_attentions
376
+ )
377
+
378
+ attn_output = attn_outputs[0] if output_attentions else attn_outputs
379
+ attn_output = pad_input(
380
+ attn_output, indices_q, bsz, q_len
381
+ ).reshape(bsz, q_len, h_size)
382
+ attn_weights = attn_outputs[2] if output_attentions else None
383
+
384
+ else:
385
+
386
+ # no padding tokens, more efficient
387
+
388
+ attn_outputs = flash_attn_kvpacked_func(
389
+ q, kv, dropout_p=0.0, softmax_scale=1.0/self.norm_factor, causal=(not has_layer_past), return_attn_probs=output_attentions)
390
+
391
+ attn_output = attn_outputs[0] if output_attentions else attn_outputs
392
+ attn_output = attn_output.reshape(bsz, q_len, h_size)
393
+ attn_weights = attn_outputs[2] if output_attentions else None
394
+
395
+ if self.config.pretraining_tp > 1:
396
+ attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
397
+ o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
398
+ attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
399
+ else:
400
+ attn_output = self.o_proj(attn_output)
401
+
402
+ if not output_attentions:
403
+ attn_weights = None
404
+
405
+ return attn_output, attn_weights, past_key_value
406
+
407
+
408
+ class LlamaDecoderLayer(nn.Module):
409
+ def __init__(self, config: LlamaConfig):
410
+ super().__init__()
411
+ self.hidden_size = config.hidden_size
412
+ self.self_attn = LlamaAttention(config=config)
413
+ self.mlp = LlamaMLP(config)
414
+ self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
415
+ self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
416
+
417
+ def forward(
418
+ self,
419
+ hidden_states: torch.Tensor,
420
+ attention_mask: Optional[torch.Tensor] = None,
421
+ position_ids: Optional[torch.LongTensor] = None,
422
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
423
+ is_padded_inputs: Optional[bool] = False,
424
+ output_attentions: Optional[bool] = False,
425
+ use_cache: Optional[bool] = False,
426
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
427
+ """
428
+ Args:
429
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
430
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
431
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
432
+ output_attentions (`bool`, *optional*):
433
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
434
+ returned tensors for more detail.
435
+ use_cache (`bool`, *optional*):
436
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
437
+ (see `past_key_values`).
438
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
439
+ """
440
+
441
+ residual = hidden_states
442
+
443
+ hidden_states = self.input_layernorm(hidden_states)
444
+
445
+ # Self Attention
446
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
447
+ hidden_states=hidden_states,
448
+ attention_mask=attention_mask,
449
+ position_ids=position_ids,
450
+ past_key_value=past_key_value,
451
+ output_attentions=output_attentions,
452
+ use_cache=use_cache,
453
+ is_padded_inputs=is_padded_inputs,
454
+ )
455
+ hidden_states = residual + hidden_states
456
+
457
+ # Fully Connected
458
+ residual = hidden_states
459
+ hidden_states = self.post_attention_layernorm(hidden_states)
460
+ hidden_states = self.mlp(hidden_states)
461
+ hidden_states = residual + hidden_states
462
+
463
+ outputs = (hidden_states,)
464
+
465
+ if output_attentions:
466
+ outputs += (self_attn_weights,)
467
+
468
+ if use_cache:
469
+ outputs += (present_key_value,)
470
+
471
+ return outputs
472
+
473
+
474
+ LLAMA_START_DOCSTRING = r"""
475
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
476
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
477
+ etc.)
478
+
479
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
480
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
481
+ and behavior.
482
+
483
+ Parameters:
484
+ config ([`LlamaConfig`]):
485
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
486
+ load the weights associated with the model, only the configuration. Check out the
487
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
488
+ """
489
+
490
+
491
+ @add_start_docstrings(
492
+ "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
493
+ LLAMA_START_DOCSTRING,
494
+ )
495
+ class LlamaPreTrainedModel(PreTrainedModel):
496
+ config_class = LlamaConfig
497
+ base_model_prefix = "model"
498
+ supports_gradient_checkpointing = True
499
+ _no_split_modules = ["LlamaDecoderLayer"]
500
+ _skip_keys_device_placement = "past_key_values"
501
+
502
+ def _init_weights(self, module):
503
+ std = self.config.initializer_range
504
+ if isinstance(module, nn.Linear):
505
+ module.weight.data.normal_(mean=0.0, std=std)
506
+ if module.bias is not None:
507
+ module.bias.data.zero_()
508
+ elif isinstance(module, nn.Embedding):
509
+ module.weight.data.normal_(mean=0.0, std=std)
510
+ if module.padding_idx is not None:
511
+ module.weight.data[module.padding_idx].zero_()
512
+
513
+ def _set_gradient_checkpointing(self, module, value=False):
514
+ if isinstance(module, LlamaModel):
515
+ module.gradient_checkpointing = value
516
+
517
+
518
+ LLAMA_INPUTS_DOCSTRING = r"""
519
+ Args:
520
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
521
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
522
+ it.
523
+
524
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
525
+ [`PreTrainedTokenizer.__call__`] for details.
526
+
527
+ [What are input IDs?](../glossary#input-ids)
528
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
529
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
530
+
531
+ - 1 for tokens that are **not masked**,
532
+ - 0 for tokens that are **masked**.
533
+
534
+ [What are attention masks?](../glossary#attention-mask)
535
+
536
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
537
+ [`PreTrainedTokenizer.__call__`] for details.
538
+
539
+ If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
540
+ `past_key_values`).
541
+
542
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
543
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
544
+ information on the default strategy.
545
+
546
+ - 1 indicates the head is **not masked**,
547
+ - 0 indicates the head is **masked**.
548
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
549
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
550
+ config.n_positions - 1]`.
551
+
552
+ [What are position IDs?](../glossary#position-ids)
553
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
554
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
555
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
556
+ `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
557
+
558
+ Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
559
+ blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
560
+
561
+ If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
562
+ don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
563
+ `decoder_input_ids` of shape `(batch_size, sequence_length)`.
564
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
565
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
566
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
567
+ model's internal embedding lookup matrix.
568
+ use_cache (`bool`, *optional*):
569
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
570
+ `past_key_values`).
571
+ output_attentions (`bool`, *optional*):
572
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
573
+ tensors for more detail.
574
+ output_hidden_states (`bool`, *optional*):
575
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
576
+ more detail.
577
+ return_dict (`bool`, *optional*):
578
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
579
+ """
580
+
581
+
582
+ @add_start_docstrings(
583
+ "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
584
+ LLAMA_START_DOCSTRING,
585
+ )
586
+ class LlamaModel(LlamaPreTrainedModel):
587
+ """
588
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
589
+
590
+ Args:
591
+ config: LlamaConfig
592
+ """
593
+
594
+ def __init__(self, config: LlamaConfig):
595
+ super().__init__(config)
596
+ self.padding_idx = config.pad_token_id
597
+ self.vocab_size = config.vocab_size
598
+
599
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
600
+ self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
601
+ self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
602
+
603
+ self.gradient_checkpointing = False
604
+ # Initialize weights and apply final processing
605
+ self.post_init()
606
+
607
+ def get_input_embeddings(self):
608
+ return self.embed_tokens
609
+
610
+ def set_input_embeddings(self, value):
611
+ self.embed_tokens = value
612
+
613
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
614
+ def forward(
615
+ self,
616
+ input_ids: torch.LongTensor = None,
617
+ attention_mask: Optional[torch.Tensor] = None,
618
+ position_ids: Optional[torch.LongTensor] = None,
619
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
620
+ inputs_embeds: Optional[torch.FloatTensor] = None,
621
+ use_cache: Optional[bool] = None,
622
+ output_attentions: Optional[bool] = None,
623
+ output_hidden_states: Optional[bool] = None,
624
+ return_dict: Optional[bool] = None,
625
+ is_padded_inputs: Optional[bool] = False,
626
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
627
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
628
+ output_hidden_states = (
629
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
630
+ )
631
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
632
+
633
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
634
+
635
+ # retrieve input_ids and inputs_embeds
636
+ if input_ids is not None and inputs_embeds is not None:
637
+ raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
638
+ elif input_ids is not None:
639
+ batch_size, seq_length = input_ids.shape
640
+ elif inputs_embeds is not None:
641
+ batch_size, seq_length, _ = inputs_embeds.shape
642
+ else:
643
+ raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
644
+
645
+ seq_length_with_past = seq_length
646
+ past_key_values_length = 0
647
+
648
+ if past_key_values is not None:
649
+ past_key_values_length = past_key_values[0][0].shape[2]
650
+ seq_length_with_past = seq_length_with_past + past_key_values_length
651
+
652
+ position_ids = None
653
+
654
+ if inputs_embeds is None:
655
+ inputs_embeds = self.embed_tokens(input_ids)
656
+
657
+ hidden_states = inputs_embeds
658
+
659
+ if self.gradient_checkpointing and self.training:
660
+ if use_cache:
661
+ logger.warning_once(
662
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
663
+ )
664
+ use_cache = False
665
+
666
+ # decoder layers
667
+ all_hidden_states = () if output_hidden_states else None
668
+ all_self_attns = () if output_attentions else None
669
+ next_decoder_cache = () if use_cache else None
670
+
671
+ for idx, decoder_layer in enumerate(self.layers):
672
+ if output_hidden_states:
673
+ all_hidden_states += (hidden_states,)
674
+
675
+ past_key_value = past_key_values[idx] if past_key_values is not None else None
676
+
677
+ if self.gradient_checkpointing and self.training:
678
+
679
+ def create_custom_forward(module):
680
+ def custom_forward(*inputs):
681
+ # None for past_key_value
682
+ return module(*inputs, output_attentions, None)
683
+
684
+ return custom_forward
685
+
686
+ layer_outputs = torch.utils.checkpoint.checkpoint(
687
+ create_custom_forward(decoder_layer),
688
+ hidden_states,
689
+ attention_mask,
690
+ position_ids,
691
+ None,
692
+ is_padded_inputs
693
+ )
694
+ else:
695
+ layer_outputs = decoder_layer(
696
+ hidden_states,
697
+ attention_mask=attention_mask,
698
+ position_ids=position_ids,
699
+ past_key_value=past_key_value,
700
+ output_attentions=output_attentions,
701
+ use_cache=use_cache,
702
+ is_padded_inputs=is_padded_inputs,
703
+ )
704
+
705
+ hidden_states = layer_outputs[0]
706
+
707
+ if use_cache:
708
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
709
+
710
+ if output_attentions:
711
+ all_self_attns += (layer_outputs[1],)
712
+
713
+ hidden_states = self.norm(hidden_states)
714
+
715
+ # add hidden states from the last decoder layer
716
+ if output_hidden_states:
717
+ all_hidden_states += (hidden_states,)
718
+
719
+ next_cache = next_decoder_cache if use_cache else None
720
+ if not return_dict:
721
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
722
+ return BaseModelOutputWithPast(
723
+ last_hidden_state=hidden_states,
724
+ past_key_values=next_cache,
725
+ hidden_states=all_hidden_states,
726
+ attentions=all_self_attns,
727
+ )
728
+
729
+
730
+ class LlamaForCausalLM(LlamaPreTrainedModel):
731
+ _tied_weights_keys = ["lm_head.weight"]
732
+
733
+ def __init__(self, config):
734
+ super().__init__(config)
735
+ self.model = LlamaModel(config)
736
+ self.vocab_size = config.vocab_size
737
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
738
+
739
+ # Initialize weights and apply final processing
740
+ self.post_init()
741
+
742
+ def get_input_embeddings(self):
743
+ return self.model.embed_tokens
744
+
745
+ def set_input_embeddings(self, value):
746
+ self.model.embed_tokens = value
747
+
748
+ def get_output_embeddings(self):
749
+ return self.lm_head
750
+
751
+ def set_output_embeddings(self, new_embeddings):
752
+ self.lm_head = new_embeddings
753
+
754
+ def set_decoder(self, decoder):
755
+ self.model = decoder
756
+
757
+ def get_decoder(self):
758
+ return self.model
759
+
760
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
761
+ @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
762
+ def forward(
763
+ self,
764
+ input_ids: torch.LongTensor = None,
765
+ attention_mask: Optional[torch.Tensor] = None,
766
+ position_ids: Optional[torch.LongTensor] = None,
767
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
768
+ inputs_embeds: Optional[torch.FloatTensor] = None,
769
+ labels: Optional[torch.LongTensor] = None,
770
+ use_cache: Optional[bool] = None,
771
+ output_attentions: Optional[bool] = None,
772
+ output_hidden_states: Optional[bool] = None,
773
+ return_dict: Optional[bool] = None,
774
+ is_padded_inputs: Optional[bool] = None,
775
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
776
+ r"""
777
+ Args:
778
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
779
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
780
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
781
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
782
+
783
+ Returns:
784
+
785
+ Example:
786
+
787
+ ```python
788
+ >>> from transformers import AutoTokenizer, LlamaForCausalLM
789
+
790
+ >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
791
+ >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
792
+
793
+ >>> prompt = "Hey, are you conscious? Can you talk to me?"
794
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
795
+
796
+ >>> # Generate
797
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
798
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
799
+ "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
800
+ ```"""
801
+
802
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
803
+ output_hidden_states = (
804
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
805
+ )
806
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
807
+
808
+ is_padded_inputs = ((attention_mask is not None) and (not attention_mask.all().item()))
809
+
810
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
811
+ outputs = self.model(
812
+ input_ids=input_ids,
813
+ attention_mask=attention_mask,
814
+ position_ids=position_ids,
815
+ past_key_values=past_key_values,
816
+ inputs_embeds=inputs_embeds,
817
+ use_cache=use_cache,
818
+ output_attentions=output_attentions,
819
+ output_hidden_states=output_hidden_states,
820
+ return_dict=return_dict,
821
+ is_padded_inputs=is_padded_inputs,
822
+ )
823
+
824
+ hidden_states = outputs[0]
825
+ if self.config.pretraining_tp > 1:
826
+ lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.config.pretraining_tp, dim=0)
827
+ logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.config.pretraining_tp)]
828
+ logits = torch.cat(logits, dim=-1)
829
+ else:
830
+ logits = self.lm_head(hidden_states)
831
+ logits = logits.float()
832
+
833
+ loss = None
834
+ if labels is not None:
835
+ # Shift so that tokens < n predict n
836
+ shift_logits = logits[..., :-1, :].contiguous()
837
+ shift_labels = labels[..., 1:].contiguous()
838
+ # Flatten the tokens
839
+ loss_fct = CrossEntropyLoss()
840
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
841
+ shift_labels = shift_labels.view(-1)
842
+ # Enable model parallelism
843
+ shift_labels = shift_labels.to(shift_logits.device)
844
+ loss = loss_fct(shift_logits, shift_labels)
845
+
846
+ if not return_dict:
847
+ output = (logits,) + outputs[1:]
848
+ return (loss,) + output if loss is not None else output
849
+
850
+ return CausalLMOutputWithPast(
851
+ loss=loss,
852
+ logits=logits,
853
+ past_key_values=outputs.past_key_values,
854
+ hidden_states=outputs.hidden_states,
855
+ attentions=outputs.attentions,
856
+ )
857
+
858
+ def prepare_inputs_for_generation(
859
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
860
+ ):
861
+ if past_key_values:
862
+ input_ids = input_ids[:, -1:]
863
+
864
+ position_ids = kwargs.get("position_ids", None)
865
+
866
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
867
+ if inputs_embeds is not None and past_key_values is None:
868
+ model_inputs = {"inputs_embeds": inputs_embeds}
869
+ else:
870
+ model_inputs = {"input_ids": input_ids}
871
+
872
+ model_inputs.update(
873
+ {
874
+ "position_ids": position_ids,
875
+ "past_key_values": past_key_values,
876
+ "use_cache": kwargs.get("use_cache"),
877
+ "attention_mask": attention_mask,
878
+ "is_padded_inputs": ((attention_mask is not None) and (not attention_mask.all().item()))
879
+ }
880
+ )
881
+ return model_inputs
882
+
883
+ @staticmethod
884
+ def _reorder_cache(past_key_values, beam_idx):
885
+ reordered_past = ()
886
+ for layer_past in past_key_values:
887
+ reordered_past += (
888
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
889
+ )
890
+ return reordered_past
891
+
892
+
893
+ @add_start_docstrings(
894
+ """
895
+ The LLaMa Model transformer with a sequence classification head on top (linear layer).
896
+
897
+ [`LlamaForSequenceClassification`] uses the last token in order to do the classification, as other causal models
898
+ (e.g. GPT-2) do.
899
+
900
+ Since it does classification on the last token, it requires to know the position of the last token. If a
901
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
902
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
903
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
904
+ each row of the batch).
905
+ """,
906
+ LLAMA_START_DOCSTRING,
907
+ )
908
+ class LlamaForSequenceClassification(LlamaPreTrainedModel):
909
+ def __init__(self, config):
910
+ super().__init__(config)
911
+ self.num_labels = config.num_labels
912
+ self.model = LlamaModel(config)
913
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
914
+
915
+ # Initialize weights and apply final processing
916
+ self.post_init()
917
+
918
+ def get_input_embeddings(self):
919
+ return self.model.embed_tokens
920
+
921
+ def set_input_embeddings(self, value):
922
+ self.model.embed_tokens = value
923
+
924
+ @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
925
+ def forward(
926
+ self,
927
+ input_ids: torch.LongTensor = None,
928
+ attention_mask: Optional[torch.Tensor] = None,
929
+ position_ids: Optional[torch.LongTensor] = None,
930
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
931
+ inputs_embeds: Optional[torch.FloatTensor] = None,
932
+ labels: Optional[torch.LongTensor] = None,
933
+ use_cache: Optional[bool] = None,
934
+ output_attentions: Optional[bool] = None,
935
+ output_hidden_states: Optional[bool] = None,
936
+ return_dict: Optional[bool] = None,
937
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
938
+ r"""
939
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
940
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
941
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
942
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
943
+ """
944
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
945
+
946
+ transformer_outputs = self.model(
947
+ input_ids,
948
+ attention_mask=attention_mask,
949
+ position_ids=position_ids,
950
+ past_key_values=past_key_values,
951
+ inputs_embeds=inputs_embeds,
952
+ use_cache=use_cache,
953
+ output_attentions=output_attentions,
954
+ output_hidden_states=output_hidden_states,
955
+ return_dict=return_dict,
956
+ )
957
+ hidden_states = transformer_outputs[0]
958
+ logits = self.score(hidden_states)
959
+
960
+ if input_ids is not None:
961
+ batch_size = input_ids.shape[0]
962
+ else:
963
+ batch_size = inputs_embeds.shape[0]
964
+
965
+ if self.config.pad_token_id is None and batch_size != 1:
966
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
967
+ if self.config.pad_token_id is None:
968
+ sequence_lengths = -1
969
+ else:
970
+ if input_ids is not None:
971
+ sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
972
+ else:
973
+ sequence_lengths = -1
974
+
975
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
976
+
977
+ loss = None
978
+ if labels is not None:
979
+ labels = labels.to(logits.device)
980
+ if self.config.problem_type is None:
981
+ if self.num_labels == 1:
982
+ self.config.problem_type = "regression"
983
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
984
+ self.config.problem_type = "single_label_classification"
985
+ else:
986
+ self.config.problem_type = "multi_label_classification"
987
+
988
+ if self.config.problem_type == "regression":
989
+ loss_fct = MSELoss()
990
+ if self.num_labels == 1:
991
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
992
+ else:
993
+ loss = loss_fct(pooled_logits, labels)
994
+ elif self.config.problem_type == "single_label_classification":
995
+ loss_fct = CrossEntropyLoss()
996
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
997
+ elif self.config.problem_type == "multi_label_classification":
998
+ loss_fct = BCEWithLogitsLoss()
999
+ loss = loss_fct(pooled_logits, labels)
1000
+ if not return_dict:
1001
+ output = (pooled_logits,) + transformer_outputs[1:]
1002
+ return ((loss,) + output) if loss is not None else output
1003
+
1004
+ return SequenceClassifierOutputWithPast(
1005
+ loss=loss,
1006
+ logits=pooled_logits,
1007
+ past_key_values=transformer_outputs.past_key_values,
1008
+ hidden_states=transformer_outputs.hidden_states,
1009
+ attentions=transformer_outputs.attentions,
1010
+ )