File size: 8,102 Bytes
eacd98b
 
 
 
 
 
 
 
8bac0a3
 
 
 
ffe6ef0
8bac0a3
 
 
 
247118f
8bac0a3
 
 
 
 
 
 
54bfe84
8bac0a3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5af661b
 
8196b65
 
5af661b
7874fb0
d00c49d
3da56b6
 
ea97f7f
 
3da56b6
8bac0a3
ea97f7f
504e404
d00c49d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3da56b6
7874fb0
d00c49d
 
 
 
 
 
 
 
504e404
 
 
 
d00c49d
 
ea97f7f
 
ffe6ef0
 
d00c49d
 
 
 
1c368e2
ea97f7f
14a93fc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea97f7f
14a93fc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1c368e2
14a93fc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dd18e79
7bc62e5
 
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
import torch
from transformers.generation.utils import (
    GenerationMixin,
    validate_stopping_criteria,
    StoppingCriteriaList,
)
from transformers import TextStreamer

def custom_generate(
    self,
    input_ids,
    attention_mask=None,
    max_new_tokens=None,
    min_length=None,
    do_sample=None,
    early_stopping=None,
    num_beams=None,
    temperature=None,
    top_k=None,
    top_p=None,
    repetition_penalty=None,
    bad_words_ids=None,
    bos_token_id=None,
    pad_token_id=None,
    eos_token_id=None,
    streamer=None,
    length_penalty=None,
    no_repeat_ngram_size=None,
    num_return_sequences=None,
    decoder_start_token_id=None,
    use_cache=None,
    num_beam_groups=None,
    diversity_penalty=None,
    prefix_allowed_tokens_fn=None,
    output_attentions=None,
    output_hidden_states=None,
    output_scores=None,
    return_dict_in_generate=None,
    forced_bos_token_id=None,
    forced_eos_token_id=None,
    remove_invalid_values=None,
    synced_gpus=None,
    **kwargs,
):
    if input_ids is None or input_ids.nelement() == 0:
        # If input_ids is None or an empty tensor, create a default input tensor
        input_ids = torch.LongTensor([[self.tokenizer.bos_token_id]]).to(self.device)
        attention_mask = torch.ones_like(input_ids).to(self.device)

    device = input_ids.device
    with torch.no_grad():
        batch_size = input_ids.shape[0]
        finished_generating = torch.zeros(batch_size, dtype=torch.bool, device=device)

        max_length = input_ids.shape[1] + max_new_tokens
        generated_token_ids = torch.full((batch_size, max_new_tokens), self.tokenizer.pad_token_id, dtype=torch.long, device=device)

        cur_token_idx = 0
        while cur_token_idx < max_new_tokens:
            # Sample the next token
            new_ids = self(
                input_ids[~finished_generating],
                attention_mask=attention_mask[~finished_generating] if attention_mask is not None else None,
                **kwargs
            )['logits']

            # Mask out the start and end thought tokens so we don't accidentally sample them
            new_ids[:, :, self.tokenizer.vocab_size:] = -float("inf")

            for list_idx, answer_idx in enumerate((~finished_generating).nonzero(as_tuple=True)[0]):
                # Find the index of the last token that is not padding
                base_answer_ids = input_ids[answer_idx]
                new_answer_ids = new_ids[list_idx]
                last_token_idx = (base_answer_ids != self.tokenizer.pad_token_id).nonzero(as_tuple=True)[0].max()

                new_ids_sampled = torch.multinomial(
                    torch.nn.functional.softmax(new_answer_ids[last_token_idx] / temperature, dim=-1), 1)

                # Assign the new id to the last token
                if last_token_idx + 1 >= len(base_answer_ids):
                    # Add padding everywhere
                    new_padding = torch.full((batch_size, 1), self.tokenizer.pad_token_id, dtype=torch.long,
                                            device=device)
                    input_ids = torch.cat([input_ids, new_padding], dim=-1)
                    if attention_mask is not None:
                        attention_mask = torch.cat([attention_mask, torch.zeros_like(new_padding)], dim=-1)

                if attention_mask is not None:
                    attention_mask[answer_idx, last_token_idx + 1] = 1
                input_ids[answer_idx, last_token_idx + 1] = new_ids_sampled

                if cur_token_idx < max_new_tokens:
                    generated_token_ids[answer_idx, cur_token_idx] = new_ids_sampled

                if new_ids_sampled == self.tokenizer.eos_token_id or cur_token_idx + 1 == max_new_tokens:
                    finished_generating[answer_idx] = 1

            cur_token_idx += 1

            if finished_generating.all():
                break

            if streamer is not None:
                streamer.put(new_ids_sampled)

        return generated_token_ids


def generate(
    self,
    input_ids,
    attention_mask=None,
    max_new_tokens=None,
    min_length=None,
    do_sample=None,
    early_stopping=None,
    num_beams=None,
    temperature=1.1,
    streamer=None,
    top_k=None,
    top_p=None,
    repetition_penalty=None,
    bad_words_ids=None,
    bos_token_id=None,
    pad_token_id=None,
    eos_token_id=None,
    length_penalty=None,
    no_repeat_ngram_size=None,
    num_return_sequences=None,
    decoder_start_token_id=None,
    use_cache=None,
    num_beam_groups=None,
    diversity_penalty=None,
    prefix_allowed_tokens_fn=None,
    output_attentions=None,
    output_hidden_states=None,
    output_scores=None,
    return_dict_in_generate=None,
    forced_bos_token_id=None,
    forced_eos_token_id=None,
    remove_invalid_values=None,
    synced_gpus=None,
    n_ahead=4,
    n_ahead_talk=4,
    merged_talk_heads=True,
    merged_lm_and_talk_heads=False,
    merged_lm_and_think_heads=True,
    use_concat_talk_head=True,
    use_shallow_think=True,
    use_shallow_talk=False,
    use_complex_think_head=False,
    use_complex_talk_head=True,
    use_weighted_talk_head=True,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    dynamic_temperature=None,
    **model_kwargs,
):

    if max_new_tokens is None:
        max_new_tokens = 128 
    
    # Set model attributes
    self.max_thoughts = n_ahead + n_ahead_talk + 1
    self.merged_talk_heads = merged_talk_heads
    self.merged_lm_and_talk_heads = merged_lm_and_talk_heads
    self.merged_lm_and_think_heads = merged_lm_and_think_heads
    self.use_concat_talk_head = use_concat_talk_head
    self.use_shallow_think = use_shallow_think
    self.use_shallow_talk = use_shallow_talk
    self.use_complex_think_head = use_complex_think_head
    self.use_complex_talk_head = use_complex_talk_head
    self.use_weighted_talk_head = use_weighted_talk_head

    # Set model properties
    self.use_end_thought_token = True
    self.use_start_thought_token = True
    self.n_ahead = n_ahead
    self.n_passes = 1
    self.eval_mode = True
    self.first_run = False
    self.rm_initialized = True
    self.original_mode = False

    # Check if the input is a string (for compatibility with text-generation-webui)
    if isinstance(input_ids, str):
        input_ids = self.tokenizer.encode(input_ids, return_tensors='pt')

    # Move input_ids and attention_mask to the same device as the model
    input_ids = input_ids.to(self.device)
    if attention_mask is not None:
        attention_mask = attention_mask.to(self.device)

    generated_token_ids = custom_generate(
        self,
        input_ids=input_ids, 
        attention_mask=attention_mask,
        max_new_tokens=max_new_tokens,
        min_length=min_length,
        do_sample=do_sample,
        early_stopping=early_stopping,
        num_beams=num_beams,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        bad_words_ids=bad_words_ids,
        bos_token_id=bos_token_id,
        pad_token_id=pad_token_id,
        eos_token_id=eos_token_id,
        length_penalty=length_penalty,
        no_repeat_ngram_size=no_repeat_ngram_size,
        num_return_sequences=num_return_sequences,
        decoder_start_token_id=decoder_start_token_id,
        use_cache=use_cache,
        num_beam_groups=num_beam_groups,
        diversity_penalty=diversity_penalty,
        prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
        output_attentions=output_attentions,
        output_hidden_states=output_hidden_states,
        output_scores=output_scores,
        return_dict_in_generate=return_dict_in_generate,
        forced_bos_token_id=forced_bos_token_id,
        forced_eos_token_id=forced_eos_token_id,
        remove_invalid_values=remove_invalid_values,
        synced_gpus=synced_gpus,
        streamer=streamer,
        **model_kwargs,
    )


    generated_text = self.tokenizer.decode(generated_token_ids[0], skip_special_tokens=False)
    return generated_token_ids, generated_text