File size: 30,890 Bytes
140387c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
import os
import time
import uuid
from enum import Enum
from threading import Thread
from typing import Any, Iterator, Union, List
from llama2_wrapper.types import (
    Completion,
    CompletionChunk,
    ChatCompletion,
    ChatCompletionChunk,
    # ChatCompletionMessage,
    Message,
    B_INST,
    E_INST,
    B_SYS,
    E_SYS,
)


class LLAMA2_WRAPPER:
    def __init__(
        self,
        model_path: str = "",
        tokenizer_path: str = "",
        backend_type: str = "llama.cpp",
        max_tokens: int = 4000,
        load_in_8bit: bool = True,
        verbose: bool = False,
    ):
        """Load a llama2 model from `model_path`.

        Args:
            model_path: Path to the model.
            backend_type: Backend for llama2, options: llama.cpp, gptq, transformers
            max_tokens: Maximum context size.
            load_in_8bit: Use bitsandbytes to run model in 8 bit mode (only for transformers models).
            verbose: Print verbose output to stderr.

        Raises:
            ValueError: If the model path does not exist.

        Returns:
            A LLAMA2_WRAPPER instance.
        """
        self.model_path = model_path
        self.tokenizer_path = tokenizer_path
        self.backend_type = BackendType.get_type(backend_type)
        self.max_tokens = max_tokens
        self.load_in_8bit = load_in_8bit

        self.model = None
        self.tokenizer = None

        self.verbose = verbose

        if self.backend_type is BackendType.LLAMA_CPP:
            print("Running on backend llama.cpp.")
        elif self.backend_type is BackendType.LLAMA2_CU:
            print("Running on backend llama2.cu.")
        else:
            import torch

            if torch.cuda.is_available():
                print("Running on GPU with backend torch transformers.")
            else:
                print("GPU CUDA not found.")

        self.default_llamacpp_path = "./models/llama-2-7b-chat.Q4_0.gguf"
        self.default_gptq_path = "./models/Llama-2-7b-Chat-GPTQ"
        self.default_llama2cu_path = "./models/llama2_7b.bin"
        # Download default ggml/gptq model
        if self.model_path == "":
            print("Model path is empty.")
            if self.backend_type is BackendType.LLAMA_CPP:
                print("Use default llama.cpp model path: " + self.default_llamacpp_path)
                if not os.path.exists(self.default_llamacpp_path):
                    print("Start downloading model to: " + self.default_llamacpp_path)
                    from huggingface_hub import hf_hub_download

                    hf_hub_download(
                        repo_id="TheBloke/Llama-2-7b-Chat-GGUF",
                        filename="llama-2-7b-chat.Q4_0.gguf",
                        local_dir="./models/",
                    )
                else:
                    print("Model exists in ./models/llama-2-7b-chat.Q4_0.gguf.")
                self.model_path = self.default_llamacpp_path
            elif self.backend_type is BackendType.LLAMA2_CU:
                if not os.path.exists(self.default_llama2cu_path):
                    print("Default model not found in " + self.default_llama2cu_path)
                    exit(1)
                else:
                    print("Model exists in " + self.default_llama2cu_path)
                self.model_path = self.default_llama2cu_path
            elif self.backend_type is BackendType.GPTQ:
                print("Use default gptq model path: " + self.default_gptq_path)
                if not os.path.exists(self.default_gptq_path):
                    print("Start downloading model to: " + self.default_gptq_path)
                    from huggingface_hub import snapshot_download

                    snapshot_download(
                        "TheBloke/Llama-2-7b-Chat-GPTQ",
                        local_dir=self.default_gptq_path,
                    )
                else:
                    print("Model exists in " + self.default_gptq_path)
                self.model_path = self.default_gptq_path

        self.init_tokenizer()
        self.init_model()

    def init_model(self):
        if self.model is None:
            self.model = LLAMA2_WRAPPER.create_llama2_model(
                self.model_path,
                self.backend_type,
                self.max_tokens,
                self.load_in_8bit,
                self.verbose,
                self.tokenizer_path,
            )
        if self.backend_type not in [BackendType.LLAMA_CPP, BackendType.LLAMA2_CU]:
            self.model.eval()

    def init_tokenizer(self):
        if self.backend_type not in [BackendType.LLAMA_CPP, BackendType.LLAMA2_CU]:
            if self.tokenizer is None:
                self.tokenizer = LLAMA2_WRAPPER.create_llama2_tokenizer(self.model_path)
        elif self.backend_type is BackendType.LLAMA2_CU:
            self.default_llama2cu_tokenizer = "./models/tokenizer.bin"
            if not os.path.exists(self.default_llama2cu_tokenizer):
                print("Default tokenizer not found in " + self.default_llama2cu_tokenizer)
                exit(1)
            else:
                print("Tokenizer exists in " + self.default_llama2cu_tokenizer)
            self.tokenizer_path = self.default_llama2cu_tokenizer

    @classmethod
    def create_llama2_model(
        cls, model_path, backend_type, max_tokens, load_in_8bit, verbose, tokenizer_path
    ):
        if backend_type is BackendType.LLAMA_CPP:
            from llama_cpp import Llama

            model = Llama(
                model_path=model_path,
                n_ctx=max_tokens,
                n_batch=max_tokens,
                verbose=verbose,
            )
        elif backend_type is BackendType.LLAMA2_CU:
            from llama2_cu_python import Llama2

            model = Llama2(model_path=model_path, tokenizer_path=tokenizer_path, n_ctx=max_tokens, n_batch=max_tokens)
        elif backend_type is BackendType.GPTQ:
            from auto_gptq import AutoGPTQForCausalLM

            model = AutoGPTQForCausalLM.from_quantized(
                model_path,
                use_safetensors=True,
                trust_remote_code=True,
                device="cuda:0",
                use_triton=False,
                quantize_config=None,
            )
        elif backend_type is BackendType.TRANSFORMERS:
            import torch
            from transformers import AutoModelForCausalLM

            model = AutoModelForCausalLM.from_pretrained(
                model_path,
                device_map="auto",
                torch_dtype=torch.float16,
                load_in_8bit=load_in_8bit,
            )
        else:
            print(backend_type + "not implemented.")
        return model

    @classmethod
    def create_llama2_tokenizer(cls, model_path):
        from transformers import AutoTokenizer

        tokenizer = AutoTokenizer.from_pretrained(model_path)
        return tokenizer

    def get_token_length(
        self,
        prompt: str,
    ) -> int:
        if self.backend_type is BackendType.LLAMA_CPP:
            input_ids = self.model.tokenize(bytes(prompt, "utf-8"))
            return len(input_ids)
        elif self.backend_type is BackendType.LLAMA2_CU:
            input_ids = self.model.tokenize(prompt)
            return len(input_ids)
        else:
            input_ids = self.tokenizer([prompt], return_tensors="np")["input_ids"]
            return input_ids.shape[-1]

    def get_input_token_length(
        self,
        message: str,
        chat_history: list[tuple[str, str]] = [],
        system_prompt: str = "",
    ) -> int:
        prompt = get_prompt(message, chat_history, system_prompt)

        return self.get_token_length(prompt)

    def generate(
        self,
        prompt: str,
        max_new_tokens: int = 1000,
        temperature: float = 0.9,
        top_p: float = 1.0,
        top_k: int = 40,
        repetition_penalty: float = 1.0,
        **kwargs: Any,
    ) -> Iterator[str]:
        """Create a generator of response from a prompt.

        Examples:
            >>> llama2_wrapper = LLAMA2_WRAPPER()
            >>> prompt = get_prompt("Hi do you know Pytorch?")
            >>> for response in llama2_wrapper.generate(prompt):
            ...     print(response)

        Args:
            prompt: The prompt to generate text from.
            max_new_tokens: The maximum number of tokens to generate.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for sampling.
            top_k: The top-k value to use for sampling.
            repetition_penalty: The penalty to apply to repeated tokens.
            kwargs: all other arguments.

        Yields:
            The generated text.
        """
        if self.backend_type is BackendType.LLAMA_CPP:
            result = self.model(
                prompt=prompt,
                stream=True,
                max_tokens=max_new_tokens,
                top_k=top_k,
                top_p=top_p,
                temperature=temperature,
                repeat_penalty=repetition_penalty,
                **kwargs,
            )
            outputs = []
            for part in result:
                text = part["choices"][0]["text"]
                outputs.append(text)
                yield "".join(outputs)
        elif self.backend_type is BackendType.LLAMA2_CU:
            result = self.model(
                prompt=prompt,
                stream=True,
                max_tokens=max_new_tokens,
                top_k=top_k,
                top_p=top_p,
                temperature=temperature,
                repeat_penalty=repetition_penalty,
                **kwargs,
            )
            outputs = []
            for part in result:
                outputs.append(part)
                yield "".join(outputs)
        else:
            from transformers import TextIteratorStreamer

            inputs = self.tokenizer([prompt], return_tensors="pt").to("cuda")

            streamer = TextIteratorStreamer(
                self.tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True
            )
            generate_kwargs = dict(
                inputs,
                streamer=streamer,
                max_new_tokens=max_new_tokens,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                repetition_penalty=repetition_penalty,
                # num_beams=1,
            )
            generate_kwargs = (
                generate_kwargs if kwargs is None else {**generate_kwargs, **kwargs}
            )
            t = Thread(target=self.model.generate, kwargs=generate_kwargs)
            t.start()

            outputs = []
            for text in streamer:
                outputs.append(text)
                yield "".join(outputs)

    def run(
        self,
        message: str,
        chat_history: list[tuple[str, str]] = [],
        system_prompt: str = "",
        max_new_tokens: int = 1000,
        temperature: float = 0.9,
        top_p: float = 1.0,
        top_k: int = 40,
        repetition_penalty: float = 1.0,
    ) -> Iterator[str]:
        """Create a generator of response from a chat message.
        Process message to llama2 prompt with chat history
        and system_prompt for chatbot.

        Args:
            message: The origianl chat message to generate text from.
            chat_history: Chat history list from chatbot.
            system_prompt: System prompt for chatbot.
            max_new_tokens: The maximum number of tokens to generate.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for sampling.
            top_k: The top-k value to use for sampling.
            repetition_penalty: The penalty to apply to repeated tokens.
            kwargs: all other arguments.

        Yields:
            The generated text.
        """
        prompt = get_prompt(message, chat_history, system_prompt)
        return self.generate(
            prompt, max_new_tokens, temperature, top_p, top_k, repetition_penalty
        )

    def __call__(
        self,
        prompt: str,
        stream: bool = False,
        max_new_tokens: int = 1000,
        temperature: float = 0.9,
        top_p: float = 1.0,
        top_k: int = 40,
        repetition_penalty: float = 1.0,
        **kwargs: Any,
    ) -> Union[str, Iterator[str]]:
        """Generate text from a prompt.

        Examples:
            >>> llama2_wrapper = LLAMA2_WRAPPER()
            >>> prompt = get_prompt("Hi do you know Pytorch?")
            >>> print(llama2_wrapper(prompt))

        Args:
            prompt: The prompt to generate text from.
            stream: Whether to stream the results.
            max_new_tokens: The maximum number of tokens to generate.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for sampling.
            top_k: The top-k value to use for sampling.
            repetition_penalty: The penalty to apply to repeated tokens.
            kwargs: all other arguments.

        Raises:
            ValueError: If the requested tokens exceed the context window.
            RuntimeError: If the prompt fails to tokenize or the model fails to evaluate the prompt.

        Returns:
            Generated text.
        """
        if self.backend_type is BackendType.LLAMA_CPP:
            completion_or_chunks = self.model.__call__(
                prompt,
                stream=stream,
                max_tokens=max_new_tokens,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                repeat_penalty=repetition_penalty,
                **kwargs,
            )
            if stream:

                def chunk_generator(chunks):
                    for part in chunks:
                        chunk = part["choices"][0]["text"]
                        yield chunk

                chunks: Iterator[str] = chunk_generator(completion_or_chunks)
                return chunks
            return completion_or_chunks["choices"][0]["text"]
        elif self.backend_type is BackendType.LLAMA2_CU:
            pass    # TODO
        else:
            inputs = self.tokenizer([prompt], return_tensors="pt").input_ids
            prompt_tokens_len = len(inputs[0])
            inputs = inputs.to("cuda")
            generate_kwargs = dict(
                inputs=inputs,
                max_new_tokens=max_new_tokens,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                repetition_penalty=repetition_penalty,
                # num_beams=1,
            )
            generate_kwargs = (
                generate_kwargs if kwargs is None else {**generate_kwargs, **kwargs}
            )
            if stream:
                from transformers import TextIteratorStreamer

                streamer = TextIteratorStreamer(
                    self.tokenizer,
                    timeout=10.0,
                    skip_prompt=True,
                    skip_special_tokens=True,
                )
                generate_kwargs["streamer"] = streamer

                t = Thread(target=self.model.generate, kwargs=generate_kwargs)
                t.start()
                return streamer
            else:
                output_ids = self.model.generate(
                    **generate_kwargs,
                )
                # skip prompt, skip special tokens
                output = self.tokenizer.decode(
                    output_ids[0][prompt_tokens_len:], skip_special_tokens=True
                )
                return output

    def completion(
        self,
        prompt: str,
        stream: bool = False,
        max_new_tokens: int = 1000,
        temperature: float = 0.9,
        top_p: float = 1.0,
        top_k: int = 40,
        repetition_penalty: float = 1.0,
        **kwargs: Any,
    ) -> Union[Completion, Iterator[CompletionChunk]]:
        """For OpenAI compatible API /v1/completions
        Generate text from a prompt.

        Examples:
            >>> llama2_wrapper = LLAMA2_WRAPPER()
            >>> prompt = get_prompt("Hi do you know Pytorch?")
            >>> print(llm.completion(prompt))

        Args:
            prompt: The prompt to generate text from.
            stream: Whether to stream the results.
            max_new_tokens: The maximum number of tokens to generate.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for sampling.
            top_k: The top-k value to use for sampling.
            repetition_penalty: The penalty to apply to repeated tokens.
            kwargs: all other arguments.

        Raises:
            ValueError: If the requested tokens exceed the context window.
            RuntimeError: If the prompt fails to tokenize or the model fails to evaluate the prompt.

        Returns:
            Response object containing the generated text.
        """
        completion_id: str = f"cmpl-{str(uuid.uuid4())}"
        created: int = int(time.time())
        model_name: str = (
            self.backend_type + " default model"
            if self.model_path == ""
            else self.model_path
        )
        if self.backend_type is BackendType.LLAMA_CPP:
            completion_or_chunks = self.model.__call__(
                prompt,
                stream=stream,
                max_tokens=max_new_tokens,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                repeat_penalty=repetition_penalty,
                **kwargs,
            )
            if stream:
                chunks: Iterator[CompletionChunk] = completion_or_chunks
                return chunks
            return completion_or_chunks
        elif self.backend_type is BackendType.LLAMA2_CU:
            pass # TODO
        else:
            inputs = self.tokenizer([prompt], return_tensors="pt").input_ids
            prompt_tokens_len = len(inputs[0])
            inputs = inputs.to("cuda")
            generate_kwargs = dict(
                inputs=inputs,
                max_new_tokens=max_new_tokens,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                repetition_penalty=repetition_penalty,
                # num_beams=1,
            )
            generate_kwargs = (
                generate_kwargs if kwargs is None else {**generate_kwargs, **kwargs}
            )
            if stream:
                from transformers import TextIteratorStreamer

                streamer = TextIteratorStreamer(
                    self.tokenizer,
                    timeout=10.0,
                    skip_prompt=True,
                    skip_special_tokens=True,
                )
                generate_kwargs["streamer"] = streamer

                t = Thread(target=self.model.generate, kwargs=generate_kwargs)
                t.start()

                def chunk_generator(chunks):
                    for part in chunks:
                        yield {
                            "id": completion_id,
                            "object": "text_completion",
                            "created": created,
                            "model": model_name,
                            "choices": [
                                {
                                    "text": part,
                                    "index": 0,
                                    "logprobs": None,
                                    "finish_reason": None,
                                }
                            ],
                        }

                chunks: Iterator[CompletionChunk] = chunk_generator(streamer)
                return chunks

            else:
                output_ids = self.model.generate(
                    **generate_kwargs,
                )
                total_tokens_len = len(output_ids[0])
                output = self.tokenizer.decode(
                    output_ids[0][prompt_tokens_len:], skip_special_tokens=True
                )
                completion: Completion = {
                    "id": completion_id,
                    "object": "text_completion",
                    "created": created,
                    "model": model_name,
                    "choices": [
                        {
                            "text": output,
                            "index": 0,
                            "logprobs": None,
                            "finish_reason": None,
                        }
                    ],
                    "usage": {
                        "prompt_tokens": prompt_tokens_len,
                        "completion_tokens": total_tokens_len - prompt_tokens_len,
                        "total_tokens": total_tokens_len,
                    },
                }
                return completion

    def chat_completion(
        self,
        messages: List[Message],
        stream: bool = False,
        max_new_tokens: int = 1000,
        temperature: float = 0.9,
        top_p: float = 1.0,
        top_k: int = 40,
        repetition_penalty: float = 1.0,
        **kwargs: Any,
    ) -> Union[ChatCompletion, Iterator[ChatCompletionChunk]]:
        """For OpenAI compatible API /v1/chat/completions
        Generate text from a dialog (chat history).

        Examples:
            >>> llama2_wrapper = LLAMA2_WRAPPER()
            >>> dialog = [
                    {
                        "role":"system",
                        "content":"You are a helpful, respectful and honest assistant. "
                    },{
                        "role":"user",
                        "content":"Hi do you know Pytorch?",
                    },
                ]
            >>> print(llm.chat_completion(dialog))

        Args:
            dialog: The dialog (chat history) to generate text from.
            stream: Whether to stream the results.
            max_new_tokens: The maximum number of tokens to generate.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for sampling.
            top_k: The top-k value to use for sampling.
            repetition_penalty: The penalty to apply to repeated tokens.
            kwargs: all other arguments.

        Raises:
            ValueError: If the requested tokens exceed the context window.
            RuntimeError: If the prompt fails to tokenize or the model fails to evaluate the prompt.

        Returns:
            Response object containing the generated text.
        """
        completion_id: str = f"cmpl-{str(uuid.uuid4())}"
        created: int = int(time.time())
        model_name: str = (
            self.backend_type + " default model"
            if self.model_path == ""
            else self.model_path
        )
        if self.backend_type is BackendType.LLAMA_CPP:
            completion_or_chunks = self.model.create_chat_completion(
                messages,
                stream=stream,
                max_tokens=max_new_tokens,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                repeat_penalty=repetition_penalty,
                **kwargs,
            )
            if stream:
                chunks: Iterator[ChatCompletionChunk] = completion_or_chunks
                return chunks
            return completion_or_chunks
        elif self.backend_type is BackendType.LLAMA2_CU:
            pass # TODO
        else:
            prompt = get_prompt_for_dialog(messages)
            inputs = self.tokenizer([prompt], return_tensors="pt").input_ids
            prompt_tokens_len = len(inputs[0])
            inputs = inputs.to("cuda")
            generate_kwargs = dict(
                inputs=inputs,
                max_new_tokens=max_new_tokens,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                repetition_penalty=repetition_penalty,
                # num_beams=1,
            )
            generate_kwargs = (
                generate_kwargs if kwargs is None else {**generate_kwargs, **kwargs}
            )
            if stream:
                from transformers import TextIteratorStreamer

                streamer = TextIteratorStreamer(
                    self.tokenizer,
                    timeout=10.0,
                    skip_prompt=True,
                    skip_special_tokens=True,
                )
                generate_kwargs["streamer"] = streamer
                t = Thread(target=self.model.generate, kwargs=generate_kwargs)
                t.start()

                def chunk_generator(chunks):
                    yield {
                        "id": "chat" + completion_id,
                        "model": model_name,
                        "created": created,
                        "object": "chat.completion.chunk",
                        "choices": [
                            {
                                "index": 0,
                                "delta": {
                                    "role": "assistant",
                                },
                                "finish_reason": None,
                            }
                        ],
                    }
                    for part in enumerate(chunks):
                        yield {
                            "id": "chat" + completion_id,
                            "model": model_name,
                            "created": created,
                            "object": "chat.completion.chunk",
                            "choices": [
                                {
                                    "index": 0,
                                    "delta": {
                                        "content": part,
                                    },
                                    "finish_reason": None,
                                }
                            ],
                        }

                chunks: Iterator[ChatCompletionChunk] = chunk_generator(streamer)
                return chunks

            else:
                output_ids = self.model.generate(
                    **generate_kwargs,
                )
                total_tokens_len = len(output_ids[0])
                output = self.tokenizer.decode(
                    output_ids[0][prompt_tokens_len:], skip_special_tokens=True
                )
                chatcompletion: ChatCompletion = {
                    "id": "chat" + completion_id,
                    "object": "chat.completion",
                    "created": created,
                    "model": model_name,
                    "choices": [
                        {
                            "index": 0,
                            "message": {
                                "role": "assistant",
                                "content": output,
                            },
                            "finish_reason": None,
                        }
                    ],
                    "usage": {
                        "prompt_tokens": prompt_tokens_len,
                        "completion_tokens": total_tokens_len - prompt_tokens_len,
                        "total_tokens": total_tokens_len,
                    },
                }
                return chatcompletion


def get_prompt_for_dialog(dialog: List[Message]) -> str:
    """Process dialog (chat history) to llama2 prompt for
    OpenAI compatible API /v1/chat/completions.

    Examples:
        >>> dialog = [
                {
                    "role":"system",
                    "content":"You are a helpful, respectful and honest assistant. "
                },{
                    "role":"user",
                    "content":"Hi do you know Pytorch?",
                },
            ]
        >>> prompt = get_prompt_for_dialog("Hi do you know Pytorch?")

    Args:
        dialog: The dialog (chat history) to generate text from.

    Yields:
        prompt string.
    """
    # add "<<SYS>>\n{system_prompt}\n<</SYS>>\n\n" in first dialog
    if dialog[0]["role"] == "system":
        dialog = [
            {
                "role": dialog[1]["role"],
                "content": B_SYS + dialog[0]["content"] + E_SYS + dialog[1]["content"],
            }
        ] + dialog[2:]
    # check roles
    assert all([msg["role"] == "user" for msg in dialog[::2]]) and all(
        [msg["role"] == "assistant" for msg in dialog[1::2]]
    ), (
        "model only supports 'system', 'user' and 'assistant' roles, "
        "starting with 'system', then 'user' and alternating (u/a/u/a/u...)"
    )
    # add chat history
    texts = []
    for prompt, answer in zip(
        dialog[::2],
        dialog[1::2],
    ):
        texts.append(
            f"{B_INST} {(prompt['content']).strip()} {E_INST} {(answer['content']).strip()} "
        )
    # check last message if role is user, then add it to prompt text
    assert (
        dialog[-1]["role"] == "user"
    ), f"Last message must be from user, got {dialog[-1]['role']}"
    texts.append(f"{B_INST} {(dialog[-1]['content']).strip()} {E_INST}")
    return "".join(texts)


def get_prompt(
    message: str, chat_history: list[tuple[str, str]] = [], system_prompt: str = ""
) -> str:
    """Process message to llama2 prompt with chat history
    and system_prompt for chatbot.

    Examples:
        >>> prompt = get_prompt("Hi do you know Pytorch?")

    Args:
        message: The origianl chat message to generate text from.
        chat_history: Chat history list from chatbot.
        system_prompt: System prompt for chatbot.

    Yields:
        prompt string.
    """
    texts = [f"[INST] <<SYS>>\n{system_prompt}\n<</SYS>>\n\n"]
    for user_input, response in chat_history:
        texts.append(f"{user_input.strip()} [/INST] {response.strip()} </s><s> [INST] ")
    texts.append(f"{message.strip()} [/INST]")
    return "".join(texts)


class BackendType(Enum):
    UNKNOWN = 0
    TRANSFORMERS = 1
    GPTQ = 2
    LLAMA_CPP = 3
    LLAMA2_CU = 4

    @classmethod
    def get_type(cls, backend_name: str):
        backend_type = None
        backend_name_lower = backend_name.lower()
        if "transformers" in backend_name_lower:
            backend_type = BackendType.TRANSFORMERS
        elif "gptq" in backend_name_lower:
            backend_type = BackendType.GPTQ
        elif "cpp" in backend_name_lower:
            backend_type = BackendType.LLAMA_CPP
        elif "cu" in backend_name_lower:
            backend_type = BackendType.LLAMA2_CU
        else:
            raise Exception("Unknown backend: " + backend_name)
            # backend_type = BackendType.UNKNOWN
        return backend_type