File size: 21,423 Bytes
30320c9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
from typing import List, Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import CrossEntropyLoss, CTCLoss
import transformers

from transformers import AutoConfig, AutoModelForCausalLM, \
                         LlamaConfig, LlamaModel, LlamaForCausalLM
from transformers.trainer_pt_utils import LabelSmoother
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers import (
    WhisperProcessor,
    WhisperModel,
)

IGNORE_TOKEN_ID = LabelSmoother.ignore_index


def padding_tensor(tensor, length, dim=0, pad=False):

    if length == 0:
        return tensor
    
    assert length > 0, f"Wrong padding length: {length}"

    shape = list(tensor.shape)
    assert dim < len(shape), f"dim {dim} out of shape {shape}"
    shape[dim] = length
    padding_tensor = torch.cat(
        (
            tensor,
            torch.full(tuple(shape), pad, dtype=tensor.dtype, device=tensor.device)
        ), 
        dim=dim
    )
    return padding_tensor


class T2ULlamaConfig(LlamaConfig):
    model_type = "T2ULlama"
    
class T2ULlamaForCausalLM(LlamaForCausalLM):
    config_class = T2ULlamaConfig

    def __init__(self, config, embedding_weight=None):
        
        self.current_step = 0
        self.log = {}

        super(LlamaForCausalLM, self).__init__(config)
        self.config = config
        self.training_stage = config.unit_output
        self.pad_token_id = 128009

        llama_config = T2ULlamaConfig(**config.to_dict(), 
                                        batch_first=True, 
                                        norm_first=True
                                    )
        llama_config.architectures = ["T2ULlamaForCausalLM"]
        llama_config.pad_token_id = self.pad_token_id
        llama_config.vocab_size += llama_config.unit_vocab_size
        #######################################################
        llama_config.unit_model = "medium"
        llama_config.max_position_embeddings = 2048     # 1024 1536 2048       # origin 1024 reduced 512
        #######################################################
        if hasattr(llama_config, "unit_model"):
            if llama_config.unit_model == "large":
                llama_config.num_hidden_layers = 2
                # llama_config.hidden_size = 4096
                # llama_config.num_attention_heads = 32
                # llama_config.intermediate_size = 14336
                # llama_config.head_dim = llama_config.hidden_size // llama_config.num_attention_heads

            elif llama_config.unit_model == "tiny":
                llama_config.num_hidden_layers = 4
                llama_config.hidden_size = 512
                llama_config.num_attention_heads = 8
                llama_config.intermediate_size = 2048
                llama_config.head_dim = llama_config.hidden_size // llama_config.num_attention_heads
            else:
                llama_config.num_hidden_layers = 8
                llama_config.hidden_size = 768
                llama_config.num_attention_heads = 12
                llama_config.num_key_value_heads = 12
                llama_config.intermediate_size = 2048
                llama_config.head_dim = llama_config.hidden_size // llama_config.num_attention_heads
        else:
            llama_config.num_hidden_layers = 6
            llama_config.hidden_size = 512
            llama_config.num_attention_heads = 8
            llama_config.intermediate_size = 2048
            llama_config.head_dim = llama_config.hidden_size // llama_config.num_attention_heads
        # print(llama_config)
        
        self.model = LlamaModel(llama_config)
        # share embedding 0501 by kkq
        self.model.embed_tokens = nn.Embedding(num_embeddings=config.vocab_size, embedding_dim=config.hidden_size, padding_idx=self.pad_token_id)   # redefine
        self.unit_embedding = nn.Linear(config.hidden_size, llama_config.unit_vocab_size, bias=False) 
        self.adapter = nn.Linear(config.hidden_size, llama_config.hidden_size, bias = True) 
        self.lm_head = nn.Linear(llama_config.hidden_size, llama_config.vocab_size, bias=False)

        if self.training_stage == "pretrain":
            pass
        elif self.training_stage == "finetune" or self.training_stage == "finetune_kd" or self.training_stage == "finetune_kd_online":
            self.aligner_MLP = nn.Sequential(
                nn.Linear(config.hidden_size, config.intermediate_size),
                nn.ReLU(),
                nn.Dropout(0.1),
                nn.Linear(config.intermediate_size, config.hidden_size),
            )
            torch.nn.init.ones_(self.aligner_MLP[0].weight)
            torch.nn.init.zeros_(self.aligner_MLP[0].bias)
            torch.nn.init.ones_(self.aligner_MLP[3].weight)
            torch.nn.init.zeros_(self.aligner_MLP[3].bias)

        # Initialize weights and apply final processing
        self.post_init()

    def get_model(self):
        return self.model

    def insert_text_embedding(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        text_labels: Optional[torch.LongTensor] = None,
        shift_text_labels: Optional[torch.LongTensor] = None,
        shift_text_hidden_states: Optional[torch.FloatTensor] = None,
        unit_targets: Optional[torch.LongTensor] = None,
        sub_lengths: Optional[torch.LongTensor] = None,
        text_start_index: Optional[torch.LongTensor] = None,
        do_task: str = None,
        **kwargs: dict,
    ):  

        if inputs_embeds == None:
            # share embedding 0501 by kkq
            embed_tokens_weight = torch.cat(
                [
                    self.model.embed_tokens.weight.detach(), self.unit_embedding.weight
                ],
                dim = 0,
            )
            # print(embed_tokens_weight, embed_tokens_weight.shape)
            inputs_embeds = F.embedding(input_ids, embed_tokens_weight, padding_idx=self.pad_token_id)
        
        emb_loss = None
        if do_task == "pretrain":
            if self.training:
                if hasattr(self, "embedding_dropout"):
                    emb_origin_mask = text_labels != -100
                    origin_padding_length = labels.shape[-1] - emb_origin_mask.shape[-1]
                    extend_emb_origin_mask = padding_tensor(emb_origin_mask, origin_padding_length, 1, False)
                    extend_emb_origin_mask = ~extend_emb_origin_mask.unsqueeze(-1).expand_as(inputs_embeds)

                    # Π-Model + noise
                    log_var = self.perturb(inputs_embeds)
                    perturbed_inputs_embeds_2 = inputs_embeds + torch.randn_like(inputs_embeds) * (torch.exp(0.5 * log_var) + 1e-6)
                    # Π-Model + dropout
                    perturbed_inputs_embeds_1 = self.embedding_dropout(inputs_embeds)
                    perturbed_inputs_embeds_2 = self.embedding_dropout(perturbed_inputs_embeds_2)
                    perturbed_inputs_embeds_1 = torch.where(extend_emb_origin_mask, inputs_embeds, perturbed_inputs_embeds_1)
                    perturbed_inputs_embeds_2 = torch.where(extend_emb_origin_mask, inputs_embeds, perturbed_inputs_embeds_2)

                    inputs_embeds = torch.cat(
                        (perturbed_inputs_embeds_1, perturbed_inputs_embeds_2),
                        dim=0,
                    )

                    kl_loss = -0.5 * (1 + log_var - log_var.exp()).mean(dim=-1).sum(dim=-1).mean()
                    contrastive_loss = (1 - F.cosine_similarity(perturbed_inputs_embeds_1, perturbed_inputs_embeds_2, dim=-1)).sum(dim=-1).mean()
                    emb_loss = kl_loss + contrastive_loss

                    if kl_loss.device == torch.device("cuda:0"):
                        self.log["kl_loss"] = kl_loss.item()
                        self.log["std"] = torch.exp(0.5 * log_var).mean().item()
                        self.log["contrastive_loss"] = contrastive_loss.item()

            pass
        elif do_task == "finetune":
            inputs_embeds = inputs_embeds.detach()
            inputs_embeds_refer = inputs_embeds.clone().detach()
            shift_text_hidden_states = self.aligner_MLP(shift_text_hidden_states)
            emb_origin_mask = text_labels != -100    # get output text pos
            emb_shift_mask = shift_text_labels != -100

            origin_padding_length = labels.shape[-1] - emb_origin_mask.shape[-1]
            shift_padding_length = labels.shape[-1] - emb_shift_mask.shape[-1]
            
            extend_emb_origin_mask = padding_tensor(emb_origin_mask, origin_padding_length, 1, False)
            extend_emb_shift_mask = padding_tensor(emb_shift_mask, shift_padding_length, 1, False)
            extend_shift_text_hidden_states = padding_tensor(shift_text_hidden_states, shift_padding_length, 1, 1e-9)
            # check
            extend_text_labels = padding_tensor(text_labels, origin_padding_length, 1, -100)
            extend_shift_text_labels = padding_tensor(shift_text_labels, shift_padding_length, 1, -100)
            
            assert torch.equal(
                extend_text_labels[extend_emb_origin_mask], 
                extend_shift_text_labels[extend_emb_shift_mask]
            ), "{}\n{}\n{}\n{}".format(labels, extend_emb_origin_mask, extend_shift_text_labels, extend_emb_shift_mask)
            
            inputs_embeds[extend_emb_origin_mask.unsqueeze(-1).expand_as(inputs_embeds)] = \
                extend_shift_text_hidden_states[extend_emb_shift_mask.unsqueeze(-1).expand_as(extend_shift_text_hidden_states)].to(dtype=inputs_embeds.dtype)
            
            if self.training:
                contrastive_loss = (1 - F.cosine_similarity(inputs_embeds, inputs_embeds_refer, dim=-1)).sum(-1).mean()
                emb_loss = contrastive_loss
                if emb_loss.device == torch.device("cuda:0"):
                    self.log["contrastive_loss"] = contrastive_loss.item()
                pass
        elif do_task == "finetune_kd" :
            #inputs_embeds = inputs_embeds.detach()
            #inputs_embeds_refer = inputs_embeds.clone().detach()
            #print(text_labels)
            #print(sub_lengths.sum())
            emb_origin_mask = text_labels != -100

            fetch_lables_list = [] 
            for batch in range(emb_origin_mask.shape[0]):
                fetch_lables_list.append(text_labels[batch][emb_origin_mask[batch]])
            shift_text_hidden_states = self.aligner_MLP(shift_text_hidden_states)

            #split the shift_text_hidden_states
            #[128006, 128000, 78191, 128007, 128000, 198, 128000]
            maxn_length = sub_lengths.max() + 8
            pad_ids = torch.full(size=(sub_lengths.shape[0], sub_lengths.shape[1], maxn_length), fill_value=self.pad_token_id, dtype=torch.long).to(shift_text_hidden_states.device)

            pad_text_ids = torch.full(size=(sub_lengths.shape[0], sub_lengths.shape[1], maxn_length), fill_value=self.pad_token_id, dtype=torch.long).to(shift_text_hidden_states.device)

            atten_mask = pad_ids.ne(self.pad_token_id)
            #target_mask_part1 = pad_ids.ne(self.pad_token_id)
            shift_text_hidden_states_slice = F.embedding(pad_ids, embed_tokens_weight, padding_idx=self.pad_token_id)
            
            #print(shift_text_hidden_states_slice.shape,shift_text_hidden_states.shape)
            for batch in range(sub_lengths.shape[0]):
                cot=0
                start_index = text_start_index[batch]
                for index, sub_length in enumerate(sub_lengths[batch]):
                    if sub_length==-1:
                        break
                    #print(shift_text_hidden_states_slice[batch][index][:sub_length].shape, shift_text_hidden_states[batch][cot:cot+sub_length].shape)
                    eos_id = torch.IntTensor([128009]).to(inputs_embeds.device)
                    eos = self.model.embed_tokens(eos_id)
                    if index == 0:
                        text_prefix_ids = torch.IntTensor([128006, 128000, 65576, 128007, 128000, 198]).to(inputs_embeds.device)
                        preifx_embed = self.model.embed_tokens(text_prefix_ids)
                        pad_text_ids[batch][index][:sub_length+7] = torch.cat([text_prefix_ids, fetch_lables_list[batch][cot:cot+sub_length], eos_id],dim=0)
                        atten_mask[batch][index][:sub_length+7]=True
                    else:
                        text_prefix_ids = torch.IntTensor([128006, 128000, 65576, 128007, 128000, 198, 12800]).to(inputs_embeds.device)
                        preifx_embed = self.model.embed_tokens(text_prefix_ids)
                        pad_text_ids[batch][index][:sub_length+8] = torch.cat([text_prefix_ids, fetch_lables_list[batch][cot:cot+sub_length], eos_id], dim=0)
                        atten_mask[batch][index][:sub_length+8]=True

                    new_shift_text_hidden_states = torch.cat([preifx_embed, shift_text_hidden_states[batch][start_index+cot:start_index+cot+sub_length], eos], dim = 0)
                    shift_text_hidden_states_slice[batch][index][:new_shift_text_hidden_states.shape[0]] = new_shift_text_hidden_states

                    cot+=sub_length
            shift_text_hidden_states_slice = shift_text_hidden_states_slice.reshape(shift_text_hidden_states_slice.shape[0]*shift_text_hidden_states_slice.shape[1],shift_text_hidden_states_slice.shape[2],shift_text_hidden_states_slice.shape[3])    


            padding_unit_targets = unit_targets.clone()
            padding_unit_targets = torch.where(padding_unit_targets == IGNORE_TOKEN_ID, self.pad_token_id, padding_unit_targets)
            target_mask_part = padding_unit_targets.ne(self.pad_token_id)
            atten_mask = torch.cat([atten_mask, target_mask_part], dim = -1)
            atten_mask = atten_mask.reshape(atten_mask.shape[0]*atten_mask.shape[1],atten_mask.shape[2])

            pad_text_ids = pad_text_ids.reshape(pad_text_ids.shape[0]*pad_text_ids.shape[1],pad_text_ids.shape[2])
            shift_text_embeddings = F.embedding(pad_text_ids, embed_tokens_weight, padding_idx=self.pad_token_id)

            unit_target_slice = F.embedding(padding_unit_targets, embed_tokens_weight, padding_idx=self.pad_token_id)
            # unit_target_slice = F.embedding(unit_targets, embed_tokens_weight, padding_idx=self.pad_token_id)             
            unit_target_slice = unit_target_slice.reshape(unit_target_slice.shape[0]*unit_target_slice.shape[1],unit_target_slice.shape[2],unit_target_slice.shape[3])

            inputs_embeds = torch.cat([shift_text_hidden_states_slice, unit_target_slice], dim = 1)

            ignore_ids = torch.full(size=(sub_lengths.shape[0], sub_lengths.shape[1], maxn_length), fill_value=IGNORE_TOKEN_ID, dtype=torch.long).to(shift_text_hidden_states.device)
            unit_targets = torch.cat([ignore_ids,unit_targets],dim=-1) 
            unit_targets = unit_targets.reshape(unit_targets.shape[0]*unit_targets.shape[1],unit_targets.shape[2])

            if self.training:
                #print(shift_text_hidden_states_slice.shape, shift_text_embeddings.shape)
                contrastive_loss = (1 - F.cosine_similarity(shift_text_hidden_states_slice, shift_text_embeddings, dim=-1)).sum(-1).mean()
                emb_loss = contrastive_loss
                if emb_loss.device == torch.device("cuda:0"):
                    self.log["contrastive_loss"] = contrastive_loss.item()

        elif do_task == "finetune_kd_online":
            shift_text_hidden_states = self.aligner_MLP(shift_text_hidden_states)
            gold_inputs_embeds = inputs_embeds.clone()
            for batch in range (inputs_embeds.shape[0]):
                start_index = text_start_index[batch]
                for slice_index in range (inputs_embeds.shape[1]):
                    sub_length= sub_lengths[batch][slice_index]
                    inputs_embeds[batch][slice_index][7:7+sub_length] = shift_text_hidden_states[batch][start_index+1:start_index+1+sub_length]
                    start_index += sub_length
            if self.training:
                #print(shift_text_hidden_states_slice.shape, shift_text_embeddings.shape)
                contrastive_loss = ((1 - F.cosine_similarity(inputs_embeds, gold_inputs_embeds, dim=-1)) * attention_mask).sum(-1).mean()
                emb_loss = contrastive_loss
                if emb_loss.device == torch.device("cuda:0"):
                    self.log["contrastive_loss"] = contrastive_loss.item()
            unit_embeds = F.embedding(unit_targets, embed_tokens_weight, padding_idx=self.pad_token_id)

            inputs_embeds = torch.cat([inputs_embeds,unit_embeds], dim=2)
        else:
            inputs_embeds = self.aligner_MLP(inputs_embeds)
            #[start_header_id] + _speaker + [end_header_id] + nl_tokens only for batch one!
            units_ids = torch.IntTensor([[128009, 128006, 128000, 65576, 128007, 128000, 198]]).to(inputs_embeds.device)
            units_prefix = self.model.embed_tokens(units_ids)
            text_ids = torch.IntTensor([[128006, 128000, 65576, 128007, 128000, 198, 12800]]).to(inputs_embeds.device)
            text_prefix = self.model.embed_tokens(text_ids)
            inputs_embeds = torch.cat([text_prefix, inputs_embeds, units_prefix], dim = 1)

        inputs_embeds = self.adapter(inputs_embeds)
        if do_task == "finetune_kd":
            return (emb_loss, inputs_embeds, unit_targets, atten_mask,)
        else:
            return (emb_loss, inputs_embeds)
    
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if inputs_embeds == None:
            # inputs_embeds = self.model.embed_tokens(input_ids)
            # share embedding 0501 by kkq
            embed_tokens_weight = torch.cat(
                [
                    self.model.embed_tokens.weight.detach(), self.unit_embedding.weight
                ],
                dim = 0,
            )
            # print(embed_tokens_weight, embed_tokens_weight.shape)
            inputs_embeds = F.embedding(input_ids, embed_tokens_weight, padding_idx=self.pad_token_id)
            inputs_embeds = self.adapter(inputs_embeds)
        
        outputs = self.model(
            input_ids=None,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states)

        loss = None
        cr_loss = None
        if labels != None:
            shift_labels = labels

            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = shift_labels[..., 1:].contiguous()

            loss_fct = CrossEntropyLoss()
        
            shift_logits = shift_logits.view(-1, (self.config.vocab_size + self.config.unit_vocab_size))
            shift_labels = shift_labels.view(-1)
            shift_labels = shift_labels.to(shift_logits.device)
            
            loss = loss_fct(shift_logits, shift_labels)
                
            if loss.device == torch.device("cuda:0"):
                self.log["unit_loss"] = loss.item()

            if cr_loss != None:
                target_scale = loss.item() * 0.2
                cr_loss_weight = target_scale / cr_loss.item() if cr_loss > target_scale else 1.0
                loss = loss + cr_loss_weight * cr_loss

            if loss.device == torch.device("cuda:0") and (self.current_step - 10) % 100 == 0:
                print(self.log, loss.device)
        
        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

AutoConfig.register("T2ULlama", T2ULlamaConfig)
AutoModelForCausalLM.register(T2ULlamaConfig, T2ULlamaForCausalLM)