Text Generation
Transformers
PyTorch
English
gpt_neox
gpt
llm
large language model
open-source
text-generation-inference
File size: 32,720 Bytes
d01fb16
 
 
d5d0b9d
d01fb16
 
 
 
980dc16
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d01fb16
980dc16
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d01fb16
 
980dc16
 
 
 
 
 
 
d01fb16
 
 
 
 
 
 
980dc16
 
 
 
 
 
 
 
 
d01fb16
d5d0b9d
 
980dc16
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
from transformers import TextGenerationPipeline
from transformers.pipelines.text_generation import ReturnType





class H2OTextGenerationPipeline(TextGenerationPipeline):
    def __init__(self, *args, debug=False, chat=False, stream_output=False,
                 sanitize_bot_response=True,
                 use_prompter=True, prompter=None, prompt_type=None,
                 max_input_tokens=2048 - 256, **kwargs):
        """
        HF-like pipeline, but handle instruction prompting and stopping (for some models)
        :param args:
        :param debug:
        :param chat:
        :param stream_output:
        :param sanitize_bot_response:
        :param use_prompter: Whether to use prompter.  If pass prompt_type, will make prompter
        :param prompter: prompter, can pass if have already
        :param prompt_type: prompt_type, e.g. human_bot.  See prompt_type to model mapping in from prompter.py.
                            If use_prompter, then will make prompter and use it.
        :param max_input_tokens:
        :param kwargs:
        """
        super().__init__(*args, **kwargs)
        self.prompt_text = None
        self.use_prompter = use_prompter
        self.prompt_type = prompt_type
        self.prompter = prompter
        if self.use_prompter:
            if self.prompter is not None:
                assert self.prompter.prompt_type is not None
            else:
                self.prompter = Prompter(self.prompt_type, debug=debug, chat=chat, stream_output=stream_output)
            self.human = self.prompter.humanstr
            self.bot = self.prompter.botstr
            self.can_stop = True
        else:
            self.prompter = None
            self.human = None
            self.bot = None
            self.can_stop = False
        self.sanitize_bot_response = sanitize_bot_response
        self.max_input_tokens = max_input_tokens  # not for generate, so ok that not kwargs

    def preprocess(self, prompt_text, prefix="", handle_long_generation=None, **generate_kwargs):
        data_point = dict(context='', instruction=prompt_text, input='')
        if self.prompter is not None:
            prompt_text = self.prompter.generate_prompt(data_point)
        self.prompt_text = prompt_text
        if handle_long_generation is None:
            # forces truncation of inputs to avoid critical failure
            handle_long_generation = 'hole'
        return super().preprocess(prompt_text, prefix=prefix, handle_long_generation=handle_long_generation,
                                  **generate_kwargs)

    def postprocess(self, model_outputs, return_type=ReturnType.FULL_TEXT, clean_up_tokenization_spaces=True):
        records = super().postprocess(model_outputs, return_type=return_type,
                                      clean_up_tokenization_spaces=clean_up_tokenization_spaces)
        for rec in records:
            if self.use_prompter:
                outputs = rec['generated_text']
                outputs = self.prompter.get_response(outputs, prompt=self.prompt_text,
                                                     sanitize_bot_response=self.sanitize_bot_response)
            elif self.bot and self.human:
                outputs = rec['generated_text'].split(self.bot)[1].strip().split(self.human)[0].strip()
            else:
                outputs = rec['generated_text']
            rec['generated_text'] = outputs
        return records

    def _forward(self, model_inputs, **generate_kwargs):
        if self.can_stop:
            stopping_criteria = get_stopping(self.prompt_type, self.tokenizer, self.device, human=self.human,
                                             bot=self.bot)
            generate_kwargs['stopping_criteria'] = stopping_criteria
        # return super()._forward(model_inputs, **generate_kwargs)
        return self.__forward(model_inputs, **generate_kwargs)

    # FIXME: Copy-paste of original _forward, but removed copy.deepcopy()
    # FIXME: https://github.com/h2oai/h2ogpt/issues/172
    def __forward(self, model_inputs, **generate_kwargs):
        input_ids = model_inputs["input_ids"]
        attention_mask = model_inputs.get("attention_mask", None)
        # Allow empty prompts
        if input_ids.shape[1] == 0:
            input_ids = None
            attention_mask = None
            in_b = 1
        else:
            in_b = input_ids.shape[0]
        prompt_text = model_inputs.pop("prompt_text")

        ## If there is a prefix, we may need to adjust the generation length. Do so without permanently modifying
        ## generate_kwargs, as some of the parameterization may come from the initialization of the pipeline.
        # generate_kwargs = copy.deepcopy(generate_kwargs)
        prefix_length = generate_kwargs.pop("prefix_length", 0)
        if prefix_length > 0:
            has_max_new_tokens = "max_new_tokens" in generate_kwargs or (
                    "generation_config" in generate_kwargs
                    and generate_kwargs["generation_config"].max_new_tokens is not None
            )
            if not has_max_new_tokens:
                generate_kwargs["max_length"] = generate_kwargs.get("max_length") or self.model.config.max_length
                generate_kwargs["max_length"] += prefix_length
            has_min_new_tokens = "min_new_tokens" in generate_kwargs or (
                    "generation_config" in generate_kwargs
                    and generate_kwargs["generation_config"].min_new_tokens is not None
            )
            if not has_min_new_tokens and "min_length" in generate_kwargs:
                generate_kwargs["min_length"] += prefix_length

        # BS x SL
        generated_sequence = self.model.generate(input_ids=input_ids, attention_mask=attention_mask, **generate_kwargs)
        out_b = generated_sequence.shape[0]
        if self.framework == "pt":
            generated_sequence = generated_sequence.reshape(in_b, out_b // in_b, *generated_sequence.shape[1:])
        elif self.framework == "tf":
            from transformers import is_tf_available
            if is_tf_available():
                import tensorflow as tf
                generated_sequence = tf.reshape(generated_sequence,
                                                (in_b, out_b // in_b, *generated_sequence.shape[1:]))
            else:
                raise ValueError("TF not avaialble.")
        return {"generated_sequence": generated_sequence, "input_ids": input_ids, "prompt_text": prompt_text}
import torch
from transformers import StoppingCriteria, StoppingCriteriaList



class StoppingCriteriaSub(StoppingCriteria):

    def __init__(self, stops=[], encounters=[], device="cuda"):
        super().__init__()
        assert len(stops) % len(encounters) == 0, "Number of stops and encounters must match"
        self.encounters = encounters
        self.stops = [stop.to(device) for stop in stops]
        self.num_stops = [0] * len(stops)

    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        for stopi, stop in enumerate(self.stops):
            if torch.all((stop == input_ids[0][-len(stop):])).item():
                self.num_stops[stopi] += 1
                if self.num_stops[stopi] >= self.encounters[stopi % len(self.encounters)]:
                    # print("Stopped", flush=True)
                    return True
        # print("Tokens: %s" % input_ids[0].cpu().numpy(), flush=True)
        # print("Stop Tokens: %s" % [x.cpu().numpy() for x in self.stops], flush=True)
        return False


def get_stopping(prompt_type, tokenizer, device, human='<human>:', bot="<bot>:"):
    if prompt_type in [PromptType.human_bot.name, PromptType.instruct_vicuna.name, PromptType.instruct_with_end.name]:
        if prompt_type == PromptType.human_bot.name:
            # encounters = [prompt.count(human) + 1, prompt.count(bot) + 1]
            # stopping only starts once output is beyond prompt
            # 1 human is enough to trigger, but need 2 bots, because very first view back will be bot we added
            stop_words = [human, bot, '\n' + human, '\n' + bot]
            encounters = [1, 2]
        elif prompt_type == PromptType.instruct_vicuna.name:
            # even below is not enough, generic strings and many ways to encode
            stop_words = [
                '### Human:',
                """
### Human:""",
                """
### Human:
""",
                '### Assistant:',
                """
### Assistant:""",
                """
### Assistant:
""",
            ]
            encounters = [1, 2]
        else:
            # some instruct prompts have this as end, doesn't hurt to stop on it since not common otherwise
            stop_words = ['### End']
            encounters = [1]
        stop_words_ids = [
            tokenizer(stop_word, return_tensors='pt')['input_ids'].squeeze() for stop_word in stop_words]
        # handle single token case
        stop_words_ids = [x if len(x.shape) > 0 else torch.tensor([x]) for x in stop_words_ids]
        stop_words_ids = [x for x in stop_words_ids if x.shape[0] > 0]
        # avoid padding in front of tokens
        if tokenizer._pad_token:  # use hidden variable to avoid annoying properly logger bug
            stop_words_ids = [x[1:] if x[0] == tokenizer.pad_token_id and len(x) > 1 else x for x in stop_words_ids]
        # handle fake \n added
        stop_words_ids = [x[1:] if y[0] == '\n' else x for x, y in zip(stop_words_ids, stop_words)]
        # build stopper
        stopping_criteria = StoppingCriteriaList(
            [StoppingCriteriaSub(stops=stop_words_ids, encounters=encounters, device=device)])
    else:
        stopping_criteria = StoppingCriteriaList()
    return stopping_criteria
import time
from enum import Enum

non_hf_types = ['gpt4all_llama', 'llama', 'gptj']


class PromptType(Enum):
    plain = 0
    instruct = 1
    quality = 2
    human_bot = 3
    dai_faq = 4
    summarize = 5
    simple_instruct = 6
    instruct_vicuna = 7
    instruct_with_end = 8
    human_bot_orig = 9
    prompt_answer = 10
    open_assistant = 11
    wizard_lm = 12
    wizard_mega = 13
    instruct_vicuna2 = 14
    instruct_vicuna3 = 15
    wizard2 = 16
    wizard3 = 17


prompt_type_to_model_name = {
    'plain': [
        'EleutherAI/gpt-j-6B',
        'EleutherAI/pythia-6.9b',
        'EleutherAI/pythia-12b',
        'EleutherAI/pythia-12b-deduped',
        'EleutherAI/gpt-neox-20b',
        'openlm-research/open_llama_7b_700bt_preview',
        'decapoda-research/llama-7b-hf',
        'decapoda-research/llama-13b-hf',
        'decapoda-research/llama-30b-hf',
        'decapoda-research/llama-65b-hf',
        'facebook/mbart-large-50-many-to-many-mmt',
        'philschmid/bart-large-cnn-samsum',
        'philschmid/flan-t5-base-samsum',
        'gpt2',
        'distilgpt2',
        'mosaicml/mpt-7b-storywriter',
        'mosaicml/mpt-7b-instruct',  # internal code handles instruct
        'mosaicml/mpt-7b-chat',  # NC, internal code handles instruct
        'gptj',  # internally handles prompting
        'llama',  # plain, or need to choose prompt_type for given TheBloke model
        'gpt4all_llama',  # internally handles prompting
    ],
    'prompt_answer': [
        'h2oai/h2ogpt-gm-oasst1-en-1024-20b',
        'h2oai/h2ogpt-gm-oasst1-en-1024-12b',
        'h2oai/h2ogpt-gm-oasst1-multilang-1024-20b',
        'h2oai/h2ogpt-gm-oasst1-en-2048-open-llama-7b-preview-300bt',
        'h2oai/h2ogpt-gm-oasst1-en-2048-open-llama-7b-preview-300bt-v2',
        'h2oai/h2ogpt-gm-oasst1-en-2048-open-llama-7b-preview-700bt',
    ],
    'instruct': [],
    'instruct_with_end': ['databricks/dolly-v2-12b'],
    'quality': [],
    'human_bot': [
        'h2oai/h2ogpt-oasst1-512-12b',
        'h2oai/h2ogpt-oasst1-512-20b',
        'h2oai/h2ogpt-oig-oasst1-256-6_9b',
        'h2oai/h2ogpt-oig-oasst1-512-6_9b',
        'h2oai/h2ogpt-oig-oasst1-256-6.9b',  # legacy
        'h2oai/h2ogpt-oig-oasst1-512-6.9b',  # legacy
        'h2oai/h2ogpt-research-oasst1-512-30b',
        'h2oai/h2ogpt-oasst1-falcon-40b',
    ],
    'dai_faq': [],
    'summarize': [],
    'simple_instruct': ['t5-small', 't5-large', 'google/flan-t5', 'google/flan-t5-xxl', 'google/flan-ul2'],
    'instruct_vicuna': ['AlekseyKorshuk/vicuna-7b', 'TheBloke/stable-vicuna-13B-HF', 'junelee/wizard-vicuna-13b'],
    'human_bot_orig': ['togethercomputer/GPT-NeoXT-Chat-Base-20B'],
    "open_assistant": ['OpenAssistant/oasst-sft-7-llama-30b-xor', 'oasst-sft-7-llama-30b'],
    "wizard_lm": ['ehartford/WizardLM-7B-Uncensored', 'ehartford/WizardLM-13B-Uncensored'],
    "wizard_mega": ['openaccess-ai-collective/wizard-mega-13b'],
}

inv_prompt_type_to_model_name = {v.strip(): k for k, l in prompt_type_to_model_name.items() for v in l}
inv_prompt_type_to_model_lower = {v.strip().lower(): k for k, l in prompt_type_to_model_name.items() for v in l}

prompt_types_strings = []
for p in PromptType:
    prompt_types_strings.extend([p.name])

prompt_types = []
for p in PromptType:
    prompt_types.extend([p.name, p.value, str(p.value)])


def get_prompt(prompt_type, chat, context, reduced):
    if prompt_type in [PromptType.plain.value, str(PromptType.plain.value),
                       PromptType.plain.name]:
        promptA = promptB = PreInstruct = PreInput = PreResponse = ''
        terminate_response = []
        chat_sep = ''
        humanstr = ''
        botstr = ''
    elif prompt_type == 'simple_instruct':
        promptA = promptB = PreInstruct = PreInput = PreResponse = None
        terminate_response = []
        chat_sep = '\n'
        humanstr = ''
        botstr = ''
    elif prompt_type in [PromptType.instruct.value, str(PromptType.instruct.value),
                         PromptType.instruct.name] + [PromptType.instruct_with_end.value,
                                                      str(PromptType.instruct_with_end.value),
                                                      PromptType.instruct_with_end.name]:
        promptA = 'Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n' if not (
                chat and reduced) else ''
        promptB = 'Below is an instruction that describes a task. Write a response that appropriately completes the request.\n' if not (
                chat and reduced) else ''

        PreInstruct = """
### Instruction:
"""

        PreInput = """
### Input:
"""

        PreResponse = """
### Response:
"""
        if prompt_type in [PromptType.instruct_with_end.value, str(PromptType.instruct_with_end.value),
                           PromptType.instruct_with_end.name]:
            terminate_response = ['### End']
        else:
            terminate_response = None
        chat_sep = '\n'
        humanstr = PreInstruct
        botstr = PreResponse
    elif prompt_type in [PromptType.quality.value, str(PromptType.quality.value),
                         PromptType.quality.name]:
        promptA = 'Write a detailed high-quality, accurate, fair, Response with about 100 words by following the Instruction as applied on the Input.\n' if not (
                chat and reduced) else ''
        promptB = 'Write a detailed high-quality, accurate, fair, Response with about 100 words by following the Instruction.\n' if not (
                chat and reduced) else ''

        PreInstruct = """
### Instruction:
"""

        PreInput = """
### Input:
"""

        PreResponse = """
### Response:
"""
        terminate_response = None
        chat_sep = '\n'
        humanstr = PreInstruct  # first thing human says
        botstr = PreResponse  # first thing bot says
    elif prompt_type in [PromptType.human_bot.value, str(PromptType.human_bot.value),
                         PromptType.human_bot.name] + [PromptType.human_bot_orig.value,
                                                       str(PromptType.human_bot_orig.value),
                                                       PromptType.human_bot_orig.name]:
        human = '<human>:'
        bot = "<bot>:"
        if reduced or context or prompt_type in [PromptType.human_bot.value, str(PromptType.human_bot.value),
                                                 PromptType.human_bot.name]:
            preprompt = ''
        else:
            cur_date = time.strftime('%Y-%m-%d')
            cur_time = time.strftime('%H:%M:%S %p %Z')

            PRE_PROMPT = """\
Current Date: {}
Current Time: {}

"""
            preprompt = PRE_PROMPT.format(cur_date, cur_time)
        start = human
        promptB = promptA = '%s%s ' % (preprompt, start)

        PreInstruct = ""

        PreInput = None

        if reduced:
            # when making context, want it to appear as-if LLM generated, which starts with space after :
            PreResponse = bot + ' '
        else:
            # normally LLM adds space after this, because was how trained.
            # if add space here, non-unique tokenization will often make LLM produce wrong output
            PreResponse = bot

        terminate_response = [start, PreResponse]
        chat_sep = '\n'
        humanstr = human  # tag before human talks
        botstr = bot  # tag before bot talks
    elif prompt_type in [PromptType.dai_faq.value, str(PromptType.dai_faq.value),
                         PromptType.dai_faq.name]:
        promptA = ''
        promptB = 'Answer the following Driverless AI question.\n'

        PreInstruct = """
### Driverless AI frequently asked question:
"""

        PreInput = None

        PreResponse = """
### Driverless AI documentation answer:
"""
        terminate_response = ['\n\n']
        chat_sep = terminate_response
        humanstr = PreInstruct
        botstr = PreResponse
    elif prompt_type in [PromptType.summarize.value, str(PromptType.summarize.value),
                         PromptType.summarize.name]:
        promptA = promptB = PreInput = ''
        PreInstruct = '## Main Text\n\n'
        PreResponse = '\n\n## Summary\n\n'
        terminate_response = None
        chat_sep = '\n'
        humanstr = PreInstruct
        botstr = PreResponse
    elif prompt_type in [PromptType.instruct_vicuna.value, str(PromptType.instruct_vicuna.value),
                         PromptType.instruct_vicuna.name]:
        promptA = promptB = "A chat between a curious human and an artificial intelligence assistant. " \
                            "The assistant gives helpful, detailed, and polite answers to the human's questions." if not (
                chat and reduced) else ''

        PreInstruct = """
### Human:
"""

        PreInput = None

        PreResponse = """
### Assistant:
"""
        terminate_response = [
            '### Human:']  # but only allow terminate after prompt is found correctly, else can't terminate
        chat_sep = '\n'
        humanstr = PreInstruct
        botstr = PreResponse
    elif prompt_type in [PromptType.prompt_answer.value, str(PromptType.prompt_answer.value),
                         PromptType.prompt_answer.name]:
        preprompt = ''
        prompt_tokens = "<|prompt|>"
        answer_tokens = "<|answer|>"
        start = prompt_tokens
        promptB = promptA = '%s%s' % (preprompt, start)
        PreInstruct = ""
        PreInput = None
        PreResponse = answer_tokens
        eos = '<|endoftext|>'  # neox eos
        terminate_response = [start, PreResponse, eos]
        chat_sep = eos
        humanstr = prompt_tokens
        botstr = answer_tokens
    elif prompt_type in [PromptType.open_assistant.value, str(PromptType.open_assistant.value),
                         PromptType.open_assistant.name]:
        # From added_tokens.json
        preprompt = ''
        prompt_tokens = "<|prompter|>"
        answer_tokens = "<|assistant|>"
        start = prompt_tokens
        promptB = promptA = '%s%s' % (preprompt, start)
        PreInstruct = ""
        PreInput = None
        PreResponse = answer_tokens
        pend = "<|prefix_end|>"
        eos = "</s>"
        terminate_response = [start, PreResponse, pend, eos]
        chat_sep = eos
        humanstr = prompt_tokens
        botstr = answer_tokens
    elif prompt_type in [PromptType.wizard_lm.value, str(PromptType.wizard_lm.value),
                         PromptType.wizard_lm.name]:
        # https://github.com/ehartford/WizardLM/blob/main/src/train_freeform.py
        preprompt = ''
        start = ''
        promptB = promptA = '%s%s' % (preprompt, start)
        PreInstruct = ""
        PreInput = None
        PreResponse = "\n\n### Response\n"
        eos = "</s>"
        terminate_response = [PreResponse, eos]
        chat_sep = eos
        humanstr = promptA
        botstr = PreResponse
    elif prompt_type in [PromptType.wizard_mega.value, str(PromptType.wizard_mega.value),
                         PromptType.wizard_mega.name]:
        preprompt = ''
        start = ''
        promptB = promptA = '%s%s' % (preprompt, start)
        PreInstruct = """
### Instruction:
"""
        PreInput = None
        PreResponse = """
### Assistant:
"""
        terminate_response = [PreResponse]
        chat_sep = '\n'
        humanstr = PreInstruct
        botstr = PreResponse
    elif prompt_type in [PromptType.instruct_vicuna2.value, str(PromptType.instruct_vicuna2.value),
                         PromptType.instruct_vicuna2.name]:
        promptA = promptB = "" if not (
                chat and reduced) else ''

        PreInstruct = """
HUMAN:
"""

        PreInput = None

        PreResponse = """
ASSISTANT:
"""
        terminate_response = [
            'HUMAN:']  # but only allow terminate after prompt is found correctly, else can't terminate
        chat_sep = '\n'
        humanstr = PreInstruct
        botstr = PreResponse
    elif prompt_type in [PromptType.instruct_vicuna3.value, str(PromptType.instruct_vicuna3.value),
                         PromptType.instruct_vicuna3.name]:
        promptA = promptB = "" if not (
                chat and reduced) else ''

        PreInstruct = """
### User:
"""

        PreInput = None

        PreResponse = """
### Assistant:
"""
        terminate_response = [
            '### User:']  # but only allow terminate after prompt is found correctly, else can't terminate
        chat_sep = '\n'
        humanstr = PreInstruct
        botstr = PreResponse
    elif prompt_type in [PromptType.wizard2.value, str(PromptType.wizard2.value),
                         PromptType.wizard2.name]:
        # https://huggingface.co/TheBloke/WizardLM-7B-uncensored-GGML
        preprompt = """Below is an instruction that describes a task. Write a response that appropriately completes the request."""
        start = ''
        promptB = promptA = '%s%s' % (preprompt, start)
        PreInstruct = """
### Instruction:
"""
        PreInput = None
        PreResponse = """
### Response:
"""
        terminate_response = [PreResponse]
        chat_sep = '\n'
        humanstr = PreInstruct
        botstr = PreResponse
    elif prompt_type in [PromptType.wizard3.value, str(PromptType.wizard3.value),
                         PromptType.wizard3.name]:
        # https://huggingface.co/TheBloke/wizardLM-13B-1.0-GGML
        preprompt = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions."""
        start = ''
        promptB = promptA = '%s%s' % (preprompt, start)
        PreInstruct = """USER: """
        PreInput = None
        PreResponse = """ASSISTANT: """
        terminate_response = [PreResponse]
        chat_sep = '\n'
        humanstr = PreInstruct
        botstr = PreResponse

    else:
        raise RuntimeError("No such prompt_type=%s" % prompt_type)

    return promptA, promptB, PreInstruct, PreInput, PreResponse, terminate_response, chat_sep, humanstr, botstr


def generate_prompt(data_point, prompt_type, chat, reduced):
    context = data_point.get('context')
    if context is None:
        context = ''
    instruction = data_point.get('instruction')
    input = data_point.get('input')
    output = data_point.get('output')
    prompt_type = data_point.get('prompt_type', prompt_type)
    assert prompt_type in prompt_types, "Bad prompt type: %s" % prompt_type
    promptA, promptB, PreInstruct, PreInput, PreResponse, \
        terminate_response, chat_sep, humanstr, botstr = get_prompt(prompt_type, chat, context, reduced)

    prompt = context if not reduced else ''

    if input and promptA:
        prompt += f"""{promptA}"""
    elif promptB:
        prompt += f"""{promptB}"""

    if instruction and PreInstruct is not None and input and PreInput is not None:
        prompt += f"""{PreInstruct}{instruction}{PreInput}{input}"""
        prompt = inject_newline(prompt_type, prompt)
    elif instruction and input and PreInstruct is None and PreInput is not None:
        prompt += f"""{PreInput}{instruction}
{input}"""
        prompt = inject_newline(prompt_type, prompt)
    elif input and instruction and PreInput is None and PreInstruct is not None:
        prompt += f"""{PreInstruct}{instruction}
{input}"""
        prompt = inject_newline(prompt_type, prompt)
    elif instruction and PreInstruct is not None:
        prompt += f"""{PreInstruct}{instruction}"""
        prompt = inject_newline(prompt_type, prompt)
    elif input and PreInput is not None:
        prompt += f"""{PreInput}{input}"""
        prompt = inject_newline(prompt_type, prompt)
    elif input and instruction and PreInput is not None:
        prompt += f"""{PreInput}{instruction}{input}"""
        prompt = inject_newline(prompt_type, prompt)
    elif input and instruction and PreInstruct is not None:
        prompt += f"""{PreInstruct}{instruction}{input}"""
        prompt = inject_newline(prompt_type, prompt)
    elif input and instruction:
        # i.e. for simple_instruct
        prompt += f"""{instruction}: {input}"""
        prompt = inject_newline(prompt_type, prompt)
    elif input:
        prompt += f"""{input}"""
        prompt = inject_newline(prompt_type, prompt)
    elif instruction:
        prompt += f"""{instruction}"""
        prompt = inject_newline(prompt_type, prompt)

    if PreResponse is not None:
        prompt += f"""{PreResponse}"""
        pre_response = PreResponse  # Don't use strip
    else:
        pre_response = ''

    if output:
        prompt += f"""{output}"""

    return prompt, pre_response, terminate_response, chat_sep


def inject_newline(prompt_type, prompt):
    if prompt_type not in [-1, '-1', 'plain', 'simple_instruct']:
        # only add new line if structured prompt, while 'plain' is just generation of next tokens from input
        prompt += '\n'
    return prompt


class Prompter(object):
    def __init__(self, prompt_type, debug=False, chat=False, stream_output=False, repeat_penalty=True,
                 allowed_repeat_line_length=10):
        self.prompt_type = prompt_type
        data_point = dict(instruction='', input='', output='')
        _, self.pre_response, self.terminate_response, self.chat_sep = \
            generate_prompt(data_point, prompt_type, chat, False)
        self.debug = debug
        self.chat = chat
        self.stream_output = stream_output
        self.repeat_penalty = repeat_penalty
        self.allowed_repeat_line_length = allowed_repeat_line_length
        self.prompt = None
        context = ""  # not for chat context
        reduced = False  # not for chat context
        self.promptA, self.promptB, self.PreInstruct, self.PreInput, self.PreResponse, \
            self.terminate_response, self.chat_sep, self.humanstr, self.botstr = \
            get_prompt(prompt_type, chat, context, reduced)

    def generate_prompt(self, data_point):
        reduced = False
        prompt, _, _, _ = generate_prompt(data_point, self.prompt_type, self.chat, reduced)
        if self.debug:
            print("prompt: ", prompt, flush=True)
        self.prompt = prompt
        return prompt

    def get_response(self, outputs, prompt=None, sanitize_bot_response=True):
        if isinstance(outputs, str):
            outputs = [outputs]
        if self.debug:
            print("output:\n", '\n\n'.join(outputs), flush=True)
        if prompt is not None:
            self.prompt = prompt

        def clean_response(response):
            meaningless_words = ['<pad>', '</s>', '<|endoftext|>']
            for word in meaningless_words:
                response = response.replace(word, "")
            if sanitize_bot_response:
                from better_profanity import profanity
                response = profanity.censor(response)
            response = response.strip("\n")
            return response

        def clean_repeats(response):
            lines = response.split('\n')
            new_lines = []
            [new_lines.append(line) for line in lines if
             line not in new_lines or len(line) < self.allowed_repeat_line_length]
            if self.debug and len(lines) != len(new_lines):
                print("cleaned repeats: %s %s" % (len(lines), len(new_lines)), flush=True)
            response = '\n'.join(new_lines)
            return response

        multi_output = len(outputs) > 1

        for oi, output in enumerate(outputs):
            if self.prompt_type in [PromptType.plain.value, str(PromptType.plain.value), PromptType.plain.name]:
                output = clean_response(output)
            elif prompt is None:
                # then use most basic parsing like pipeline
                if self.botstr in output:
                    if self.humanstr:
                        output = clean_response(output.split(self.botstr)[1].strip().split(self.humanstr)[0].strip())
                    else:
                        # i.e. use after bot but only up to next bot
                        output = clean_response(output.split(self.botstr)[1].strip().split(self.botstr)[0].strip())
                else:
                    # output = clean_response(output.strip())
                    # assume just not printed yet
                    output = ""
            else:
                # find first instance of prereponse
                # prompt sometimes has odd characters, that mutate length,
                # so can't go by length alone
                if self.pre_response:
                    outputi = output.find(prompt)
                    if outputi >= 0:
                        output = output[outputi + len(prompt):]
                        allow_terminate = True
                    else:
                        # subtraction is risky due to space offsets sometimes, so only do if necessary
                        output = output[len(prompt) - len(self.pre_response):]
                        # [1] to avoid repeated pre_response, just take first (after prompt - pre_response for chat)
                        if self.pre_response in output:
                            output = output.split(self.pre_response)[1]
                            allow_terminate = True
                        else:
                            if output:
                                print("Failure of parsing or not enough output yet: %s" % output, flush=True)
                            allow_terminate = False
                else:
                    allow_terminate = True
                    output = output[len(prompt):]
                # clean after subtract prompt out, so correct removal of pre_response
                output = clean_response(output).strip()
                if self.repeat_penalty:
                    output = clean_repeats(output).strip()
                if self.terminate_response and allow_terminate:
                    finds = []
                    for term in self.terminate_response:
                        finds.append(output.find(term))
                    finds = [x for x in finds if x >= 0]
                    if len(finds) > 0:
                        termi = finds[0]
                        output = output[:termi].strip()
                    else:
                        output = output.strip()
                else:
                    output = output.strip()
            if multi_output:
                # prefix with output counter
                output = "\n=========== Output %d\n\n" % (1 + oi) + output
                if oi > 0:
                    # post fix outputs with seperator
                    output += '\n'
            outputs[oi] = output
        # join all outputs, only one extra new line between outputs
        output = '\n'.join(outputs)
        if self.debug:
            print("outputclean:\n", '\n\n'.join(outputs), flush=True)
        return output