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  99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364

classLlama:"""High-level Python wrapper for a llama.cpp model."""__backend_initialized=Falsedef__init__(self,model_path:str,*,# Model Paramsn_gpu_layers:int=0,split_mode:int=llama_cpp.LLAMA_SPLIT_MODE_LAYER,main_gpu:int=0,tensor_split:Optional[List[float]]=None,vocab_only:bool=False,use_mmap:bool=True,use_mlock:bool=False,kv_overrides:Optional[Dict[str,Union[bool,int,float,str]]]=None,# Context Paramsseed:int=llama_cpp.LLAMA_DEFAULT_SEED,n_ctx:int=512,n_batch:int=512,n_ubatch:int=512,n_threads:Optional[int]=None,n_threads_batch:Optional[int]=None,rope_scaling_type:Optional[int]=llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED,pooling_type:int=llama_cpp.LLAMA_POOLING_TYPE_UNSPECIFIED,rope_freq_base:float=0.0,rope_freq_scale:float=0.0,yarn_ext_factor:float=-1.0,yarn_attn_factor:float=1.0,yarn_beta_fast:float=32.0,yarn_beta_slow:float=1.0,yarn_orig_ctx:int=0,logits_all:bool=False,embedding:bool=False,offload_kqv:bool=True,flash_attn:bool=False,op_offloat:Optional[bool]=None,swa_full:Optional[bool]=None,# Sampling Paramsno_perf:bool=False,last_n_tokens_size:int=64,# LoRA Paramslora_base:Optional[str]=None,lora_scale:float=1.0,lora_path:Optional[str]=None,# Backend Paramsnuma:Union[bool,int]=False,# Chat Format Paramschat_format:Optional[str]=None,chat_handler:Optional[llama_chat_format.LlamaChatCompletionHandler]=None,# Speculative Decodingdraft_model:Optional[LlamaDraftModel]=None,# Tokenizer Overridetokenizer:Optional[BaseLlamaTokenizer]=None,# KV cache quantizationtype_k:Optional[int]=None,type_v:Optional[int]=None,# Miscspm_infill:bool=False,verbose:bool=True,# Extra Params**kwargs,# type: ignore):"""Load a llama.cpp model from `model_path`.        Examples:            Basic usage            >>> import llama_cpp            >>> model = llama_cpp.Llama(            ...     model_path="path/to/model",            ... )            >>> print(model("The quick brown fox jumps ", stop=["."])["choices"][0]["text"])            the lazy dog            Loading a chat model            >>> import llama_cpp            >>> model = llama_cpp.Llama(            ...     model_path="path/to/model",            ...     chat_format="llama-2",            ... )            >>> print(model.create_chat_completion(            ...     messages=[{            ...         "role": "user",            ...         "content": "what is the meaning of life?"            ...     }]            ... ))        Args:            model_path: Path to the model.            n_gpu_layers: Number of layers to offload to GPU (-ngl). If -1, all layers are offloaded.            split_mode: How to split the model across GPUs. See llama_cpp.LLAMA_SPLIT_* for options.            main_gpu: main_gpu interpretation depends on split_mode: LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model. LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results. LLAMA_SPLIT_MODE_LAYER: ignored            tensor_split: How split tensors should be distributed across GPUs. If None, the model is not split.            vocab_only: Only load the vocabulary no weights.            use_mmap: Use mmap if possible.            use_mlock: Force the system to keep the model in RAM.            kv_overrides: Key-value overrides for the model.            seed: RNG seed, -1 for random            n_ctx: Text context, 0 = from model            n_batch: Prompt processing maximum batch size            n_ubatch: Physical batch size            n_threads: Number of threads to use for generation            n_threads_batch: Number of threads to use for batch processing            rope_scaling_type: RoPE scaling type, from `enum llama_rope_scaling_type`. ref: https://github.com/ggerganov/llama.cpp/pull/2054            pooling_type: Pooling type, from `enum llama_pooling_type`.            rope_freq_base: RoPE base frequency, 0 = from model            rope_freq_scale: RoPE frequency scaling factor, 0 = from model            yarn_ext_factor: YaRN extrapolation mix factor, negative = from model            yarn_attn_factor: YaRN magnitude scaling factor            yarn_beta_fast: YaRN low correction dim            yarn_beta_slow: YaRN high correction dim            yarn_orig_ctx: YaRN original context size            logits_all: Return logits for all tokens, not just the last token. Must be True for completion to return logprobs.            embedding: Embedding mode only.            offload_kqv: Offload K, Q, V to GPU.            flash_attn: Use flash attention.            op_offloat: offload host tensor operations to device            swa_full: use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)            no_perf: Measure performance timings.            last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque.            lora_base: Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model.            lora_path: Path to a LoRA file to apply to the model.            numa: numa policy            chat_format: String specifying the chat format to use when calling create_chat_completion.            chat_handler: Optional chat handler to use when calling create_chat_completion.            draft_model: Optional draft model to use for speculative decoding.            tokenizer: Optional tokenizer to override the default tokenizer from llama.cpp.            verbose: Print verbose output to stderr.            type_k: KV cache data type for K (default: f16)            type_v: KV cache data type for V (default: f16)            spm_infill: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this.        Raises:            ValueError: If the model path does not exist.        Returns:            A Llama instance.        """self.verbose=verboseself._stack=contextlib.ExitStack()set_verbose(verbose)ifnotLlama.__backend_initialized:withsuppress_stdout_stderr(disable=verbose):llama_cpp.llama_backend_init()Llama.__backend_initialized=Trueifisinstance(numa,bool):self.numa=(llama_cpp.GGML_NUMA_STRATEGY_DISTRIBUTEifnumaelsellama_cpp.GGML_NUMA_STRATEGY_DISABLED)else:self.numa=numaifself.numa!=llama_cpp.GGML_NUMA_STRATEGY_DISABLED:withsuppress_stdout_stderr(disable=verbose):llama_cpp.llama_numa_init(self.numa)self.model_path=model_path# Model Paramsself.model_params=llama_cpp.llama_model_default_params()self.model_params.n_gpu_layers=(0x7FFFFFFFifn_gpu_layers==-1elsen_gpu_layers)# 0x7FFFFFFF is INT32 max, will be auto set to all layersself.model_params.split_mode=split_modeself.model_params.main_gpu=main_gpuself.tensor_split=tensor_splitself._c_tensor_split=Noneifself.tensor_splitisnotNone:iflen(self.tensor_split)>llama_cpp.LLAMA_MAX_DEVICES:raiseValueError(f"Attempt to split tensors that exceed maximum supported devices. Current LLAMA_MAX_DEVICES={llama_cpp.LLAMA_MAX_DEVICES}")# Type conversion and expand the list to the length of LLAMA_MAX_DEVICESFloatArray=ctypes.c_float*llama_cpp.LLAMA_MAX_DEVICESself._c_tensor_split=FloatArray(*tensor_split# type: ignore)# keep a reference to the array so it is not gc'dself.model_params.tensor_split=self._c_tensor_splitself.model_params.vocab_only=vocab_onlyself.model_params.use_mmap=use_mmapiflora_pathisNoneelseFalseself.model_params.use_mlock=use_mlock# kv_overrides is the original python dictself.kv_overrides=kv_overridesifkv_overridesisnotNone:# _kv_overrides_array is a ctypes.Array of llama_model_kv_override Structskvo_array_len=len(kv_overrides)+1# for sentinel elementself._kv_overrides_array=(llama_cpp.llama_model_kv_override*kvo_array_len)()fori,(k,v)inenumerate(kv_overrides.items()):self._kv_overrides_array[i].key=k.encode("utf-8")ifisinstance(v,bool):self._kv_overrides_array[i].tag=llama_cpp.LLAMA_KV_OVERRIDE_TYPE_BOOLself._kv_overrides_array[i].value.val_bool=velifisinstance(v,int):self._kv_overrides_array[i].tag=llama_cpp.LLAMA_KV_OVERRIDE_TYPE_INTself._kv_overrides_array[i].value.val_i64=velifisinstance(v,float):self._kv_overrides_array[i].tag=llama_cpp.LLAMA_KV_OVERRIDE_TYPE_FLOATself._kv_overrides_array[i].value.val_f64=velifisinstance(v,str):# type: ignorev_bytes=v.encode("utf-8")iflen(v_bytes)>128:# TODO: Make this a constantraiseValueError(f"Value for{k} is too long:{v}")v_bytes=v_bytes.ljust(128,b"\0")self._kv_overrides_array[i].tag=llama_cpp.LLAMA_KV_OVERRIDE_TYPE_STR# copy min(v_bytes, 128) to str_valueaddress=typing.cast(int,ctypes.addressof(self._kv_overrides_array[i].value)+llama_cpp.llama_model_kv_override_value.val_str.offset,)buffer_start=ctypes.cast(address,ctypes.POINTER(ctypes.c_char))ctypes.memmove(buffer_start,v_bytes,128,)else:raiseValueError(f"Unknown value type for{k}:{v}")self._kv_overrides_array[-1].key=b"\0"# ensure sentinel element is zeroedself.model_params.kv_overrides=self._kv_overrides_arrayself.n_batch=min(n_ctx,n_batch)# ???self.n_threads=n_threadsormax(multiprocessing.cpu_count()//2,1)self.n_threads_batch=n_threads_batchormultiprocessing.cpu_count()# Used by the samplerself._seed=seedorllama_cpp.LLAMA_DEFAULT_SEED# Context Paramsself.context_params=llama_cpp.llama_context_default_params()self.context_params.n_ctx=n_ctxself.context_params.n_batch=self.n_batchself.context_params.n_ubatch=min(self.n_batch,n_ubatch)self.context_params.n_threads=self.n_threadsself.context_params.n_threads_batch=self.n_threads_batchself.context_params.rope_scaling_type=(rope_scaling_typeifrope_scaling_typeisnotNoneelsellama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED)self.context_params.pooling_type=pooling_typeself.context_params.rope_freq_base=(rope_freq_baseifrope_freq_base!=0.0else0)self.context_params.rope_freq_scale=(rope_freq_scaleifrope_freq_scale!=0.0else0)self.context_params.yarn_ext_factor=(yarn_ext_factorifyarn_ext_factor!=0.0else0)self.context_params.yarn_attn_factor=(yarn_attn_factorifyarn_attn_factor!=0.0else0)self.context_params.yarn_beta_fast=(yarn_beta_fastifyarn_beta_fast!=0.0else0)self.context_params.yarn_beta_slow=(yarn_beta_slowifyarn_beta_slow!=0.0else0)self.context_params.yarn_orig_ctx=yarn_orig_ctxifyarn_orig_ctx!=0else0self._logits_all=logits_allifdraft_modelisNoneelseTrueself.context_params.embeddings=embedding# TODO: Rename to embeddingsself.context_params.offload_kqv=offload_kqvself.context_params.flash_attn=flash_attnifop_offloatisnotNone:self.context_params.op_offloat=op_offloatifswa_fullisnotNone:self.context_params.swa_full=swa_full#  KV cache quantizationiftype_kisnotNone:self.context_params.type_k=type_kiftype_visnotNone:self.context_params.type_v=type_v# Sampling Paramsself.context_params.no_perf=no_perfself.last_n_tokens_size=last_n_tokens_sizeself.cache:Optional[BaseLlamaCache]=Noneself.lora_base=lora_baseself.lora_scale=lora_scaleself.lora_path=lora_pathself.spm_infill=spm_infillifnotos.path.exists(model_path):raiseValueError(f"Model path does not exist:{model_path}")self._model=self._stack.enter_context(contextlib.closing(internals.LlamaModel(path_model=self.model_path,params=self.model_params,verbose=self.verbose,)))# Override tokenizerself.tokenizer_=tokenizerorLlamaTokenizer(self)# Set the default value for the context and correct the batchifn_ctx==0:n_ctx=self._model.n_ctx_train()self.n_batch=min(n_ctx,n_batch)self.context_params.n_ctx=self._model.n_ctx_train()self.context_params.n_batch=self.n_batchself.context_params.n_ubatch=min(self.n_batch,n_ubatch)self._ctx=self._stack.enter_context(contextlib.closing(internals.LlamaContext(model=self._model,params=self.context_params,verbose=self.verbose,)))self._batch=self._stack.enter_context(contextlib.closing(internals.LlamaBatch(n_tokens=self.n_batch,embd=0,n_seq_max=self.context_params.n_ctx,verbose=self.verbose,)))self._lora_adapter:Optional[llama_cpp.llama_adapter_lora_p]=Noneifself.lora_path:self._lora_adapter=llama_cpp.llama_adapter_lora_init(self._model.model,self.lora_path.encode("utf-8"),)ifself._lora_adapterisNone:raiseRuntimeError(f"Failed to initialize LoRA adapter from lora path:{self.lora_path}")deffree_lora_adapter():ifself._lora_adapterisNone:returnllama_cpp.llama_adapter_lora_free(self._lora_adapter)self._lora_adapter=Noneself._stack.callback(free_lora_adapter)ifllama_cpp.llama_set_adapter_lora(self._ctx.ctx,self._lora_adapter,self.lora_scale):raiseRuntimeError(f"Failed to set LoRA adapter from lora path:{self.lora_path}")ifself.verbose:print(llama_cpp.llama_print_system_info().decode("utf-8"),file=sys.stderr)self.chat_format=chat_formatself.chat_handler=chat_handlerself._chat_handlers:Dict[str,llama_chat_format.LlamaChatCompletionHandler]={}self.draft_model=draft_modelself._n_vocab=self.n_vocab()self._n_ctx=self.n_ctx()self._token_nl=self.token_nl()self._token_eos=self.token_eos()self._candidates=internals.LlamaTokenDataArray(n_vocab=self._n_vocab)self.n_tokens=0self.input_ids:npt.NDArray[np.intc]=np.ndarray((n_ctx,),dtype=np.intc)self.scores:npt.NDArray[np.single]=np.ndarray((n_ctxiflogits_all==Trueelsen_batch,self._n_vocab),dtype=np.single)self._mirostat_mu=ctypes.c_float(2.0*5.0)# TODO: Move this to sampling contexttry:self.metadata=self._model.metadata()exceptExceptionase:self.metadata={}ifself.verbose:print(f"Failed to load metadata:{e}",file=sys.stderr)ifself.verbose:print(f"Model metadata:{self.metadata}",file=sys.stderr)eos_token_id=self.token_eos()bos_token_id=self.token_bos()eos_token=(self._model.token_get_text(eos_token_id)ifeos_token_id!=-1else"")bos_token=(self._model.token_get_text(bos_token_id)ifbos_token_id!=-1else"")# Unfortunately the llama.cpp API does not return metadata arrays, so we can't get template names from tokenizer.chat_templatestemplate_choices=dict((name[10:],template)forname,templateinself.metadata.items()ifname.startswith("tokenizer.chat_template."))if"tokenizer.chat_template"inself.metadata:template_choices["chat_template.default"]=self.metadata["tokenizer.chat_template"]ifself.verboseandtemplate_choices:print(f"Available chat formats from metadata:{', '.join(template_choices.keys())}",file=sys.stderr,)forname,templateintemplate_choices.items():self._chat_handlers[name]=llama_chat_format.Jinja2ChatFormatter(template=template,eos_token=eos_token,bos_token=bos_token,stop_token_ids=[eos_token_id],).to_chat_handler()if(self.chat_formatisNoneandself.chat_handlerisNoneand"chat_template.default"intemplate_choices):chat_format=llama_chat_format.guess_chat_format_from_gguf_metadata(self.metadata)ifchat_formatisnotNone:self.chat_format=chat_formatifself.verbose:print(f"Guessed chat format:{chat_format}",file=sys.stderr)else:ifself.verbose:print(f"Using gguf chat template:{template_choices['chat_template.default']}",file=sys.stderr,)print(f"Using chat eos_token:{eos_token}",file=sys.stderr)print(f"Using chat bos_token:{bos_token}",file=sys.stderr)self.chat_format="chat_template.default"ifself.chat_formatisNoneandself.chat_handlerisNone:self.chat_format="llama-2"ifself.verbose:print(f"Using fallback chat format:{self.chat_format}",file=sys.stderr)self._sampler=None@propertydefctx(self)->llama_cpp.llama_context_p:returnself._ctx.ctx@propertydefmodel(self)->llama_cpp.llama_model_p:returnself._model.model@propertydef_input_ids(self)->npt.NDArray[np.intc]:returnself.input_ids[:self.n_tokens]@propertydef_scores(self)->npt.NDArray[np.single]:returnself.scores[:self.n_tokens,:]@propertydefeval_tokens(self)->Deque[int]:returndeque(self.input_ids[:self.n_tokens].tolist(),maxlen=self._n_ctx)@propertydefeval_logits(self)->Deque[List[float]]:returndeque(self.scores[:self.n_tokens,:].tolist(),maxlen=self._n_ctxifself._logits_allelse1,)deftokenize(self,text:bytes,add_bos:bool=True,special:bool=False)->List[int]:"""Tokenize a string.        Args:            text: The utf-8 encoded string to tokenize.            add_bos: Whether to add a beginning of sequence token.            special: Whether to tokenize special tokens.        Raises:            RuntimeError: If the tokenization failed.        Returns:            A list of tokens.        """returnself.tokenizer_.tokenize(text,add_bos,special)defdetokenize(self,tokens:List[int],prev_tokens:Optional[List[int]]=None,special:bool=False,)->bytes:"""Detokenize a list of tokens.        Args:            tokens: The list of tokens to detokenize.            prev_tokens: The list of previous tokens. Offset mapping will be performed if provided.            special: Whether to detokenize special tokens.        Returns:            The detokenized string.        """returnself.tokenizer_.detokenize(tokens,prev_tokens=prev_tokens,special=special)defset_cache(self,cache:Optional[BaseLlamaCache]):"""Set the cache.        Args:            cache: The cache to set.        """self.cache=cachedefset_seed(self,seed:int):"""Set the random seed.        Args:            seed: The random seed.        """self._seed=seeddefreset(self):"""Reset the model state."""self.n_tokens=0defeval(self,tokens:Sequence[int]):"""Evaluate a list of tokens.        Args:            tokens: The list of tokens to evaluate.        """self._ctx.kv_cache_seq_rm(-1,self.n_tokens,-1)foriinrange(0,len(tokens),self.n_batch):batch=tokens[i:min(len(tokens),i+self.n_batch)]n_past=self.n_tokensn_tokens=len(batch)self._batch.set_batch(batch=batch,n_past=n_past,logits_all=self._logits_all)self._ctx.decode(self._batch)# Save tokensself.input_ids[n_past:n_past+n_tokens]=batch# Save logitsifself._logits_all:rows=n_tokenscols=self._n_vocablogits=np.ctypeslib.as_array(self._ctx.get_logits(),shape=(rows*cols,))self.scores[n_past:n_past+n_tokens,:].reshape(-1)[::]=logitselse:# rows = 1# cols = self._n_vocab# logits = np.ctypeslib.as_array(#     self._ctx.get_logits(), shape=(rows * cols,)# )# self.scores[n_past + n_tokens - 1, :].reshape(-1)[::] = logits# NOTE: Now that sampling is done inside the sampler, logits are only needed for logprobs which requires logits_allpass# Update n_tokensself.n_tokens+=n_tokensdef_init_sampler(self,top_k:int=40,top_p:float=0.95,min_p:float=0.05,typical_p:float=1.0,temp:float=0.80,repeat_penalty:float=1.0,frequency_penalty:float=0.0,presence_penalty:float=0.0,tfs_z:float=1.0,mirostat_mode:int=0,mirostat_eta:float=0.1,mirostat_tau:float=5.0,penalize_nl:bool=True,logits_processor:Optional[LogitsProcessorList]=None,grammar:Optional[LlamaGrammar]=None,):sampler=internals.LlamaSampler()iflogits_processorisnotNone:# Create and add a custom samplerdefapply_func(token_data_array:llama_cpp.llama_token_data_array_p):size=token_data_array.contents.sizedata_soa=token_data_array.contents.datadata_soa_address=ctypes.addressof(data_soa.contents)# NOTE: This is probably brokenrecarray=np.recarray(shape=(size,),dtype=np.dtype([("id",np.intc),("logit",np.single),("p",np.single)],align=True,),buf=(llama_cpp.llama_token_data*size).from_address(data_soa_address),)forlogit_processorinlogits_processor:recarray.logit[:]=logit_processor(self._input_ids,recarray.logit)sampler.add_custom(apply_func)sampler.add_penalties(# n_vocab=self._n_vocab,# special_eos_id=self._token_eos,# linefeed_id=self._token_nl,penalty_last_n=self.last_n_tokens_size,penalty_repeat=repeat_penalty,penalty_freq=frequency_penalty,penalty_present=presence_penalty,# penalize_nl=penalize_nl,# ignore_eos=False,)ifgrammarisnotNone:sampler.add_grammar(self._model,grammar)iftemp<0.0:sampler.add_softmax()sampler.add_dist(self._seed)eliftemp==0.0:sampler.add_greedy()else:ifmirostat_mode==1:mirostat_m=100sampler.add_mirostat(self._n_vocab,self._seed,mirostat_tau,mirostat_eta,mirostat_m,)elifmirostat_mode==2:sampler.add_mirostat_v2(self._seed,mirostat_tau,mirostat_eta,)else:n_probs=0min_keep=max(1,n_probs)sampler.add_top_k(top_k)sampler.add_typical(typical_p,min_keep)sampler.add_top_p(top_p,min_keep)sampler.add_min_p(min_p,min_keep)sampler.add_temp(temp)sampler.add_dist(self._seed)returnsamplerdefsample(self,top_k:int=40,top_p:float=0.95,min_p:float=0.05,typical_p:float=1.0,temp:float=0.80,repeat_penalty:float=1.0,frequency_penalty:float=0.0,presence_penalty:float=0.0,tfs_z:float=1.0,mirostat_mode:int=0,mirostat_eta:float=0.1,mirostat_tau:float=5.0,penalize_nl:bool=True,logits_processor:Optional[LogitsProcessorList]=None,grammar:Optional[LlamaGrammar]=None,idx:Optional[int]=None,):"""Sample a token from the model.        Args:            top_k: The top-k sampling parameter.            top_p: The top-p sampling parameter.            temp: The temperature parameter.            repeat_penalty: The repeat penalty parameter.        Returns:            The sampled token.        """assertself.n_tokens>0tmp_sampler=Falseifself._samplerisNone:tmp_sampler=Trueself._sampler=self._init_sampler(top_k=top_k,top_p=top_p,min_p=min_p,typical_p=typical_p,temp=temp,repeat_penalty=repeat_penalty,frequency_penalty=frequency_penalty,presence_penalty=presence_penalty,tfs_z=tfs_z,mirostat_mode=mirostat_mode,mirostat_tau=mirostat_tau,mirostat_eta=mirostat_eta,penalize_nl=penalize_nl,logits_processor=logits_processor,grammar=grammar,)ridx=idx-self.n_tokensifidxisnotNoneelse-1assertself.ctxisnotNonetoken=self._sampler.sample(self._ctx,ridx)iftmp_sampler:self._sampler=Nonereturntokendefgenerate(self,tokens:Sequence[int],top_k:int=40,top_p:float=0.95,min_p:float=0.05,typical_p:float=1.0,temp:float=0.80,repeat_penalty:float=1.0,reset:bool=True,frequency_penalty:float=0.0,presence_penalty:float=0.0,tfs_z:float=1.0,mirostat_mode:int=0,mirostat_tau:float=5.0,mirostat_eta:float=0.1,penalize_nl:bool=True,logits_processor:Optional[LogitsProcessorList]=None,stopping_criteria:Optional[StoppingCriteriaList]=None,grammar:Optional[LlamaGrammar]=None,)->Generator[int,Optional[Sequence[int]],None]:"""Create a generator of tokens from a prompt.        Examples:            >>> llama = Llama("models/ggml-7b.bin")            >>> tokens = llama.tokenize(b"Hello, world!")            >>> for token in llama.generate(tokens, top_k=40, top_p=0.95, temp=1.0, repeat_penalty=1.0):            ...     print(llama.detokenize([token]))        Args:            tokens: The prompt tokens.            top_k: The top-k sampling parameter.            top_p: The top-p sampling parameter.            temp: The temperature parameter.            repeat_penalty: The repeat penalty parameter.            reset: Whether to reset the model state.        Yields:            The generated tokens.        """# Reset mirostat samplingself._mirostat_mu=ctypes.c_float(2.0*mirostat_tau)self._sampler=self._init_sampler(top_k=top_k,top_p=top_p,min_p=min_p,typical_p=typical_p,temp=temp,repeat_penalty=repeat_penalty,frequency_penalty=frequency_penalty,presence_penalty=presence_penalty,tfs_z=tfs_z,mirostat_mode=mirostat_mode,mirostat_tau=mirostat_tau,mirostat_eta=mirostat_eta,penalize_nl=penalize_nl,logits_processor=logits_processor,grammar=grammar,)# Check for kv cache prefix matchifresetandself.n_tokens>0:longest_prefix=0fora,binzip(self._input_ids,tokens[:-1]):ifa==b:longest_prefix+=1else:breakiflongest_prefix>0:reset=Falsetokens=tokens[longest_prefix:]self.n_tokens=longest_prefixifself.verbose:print(f"Llama.generate:{longest_prefix} prefix-match hit, "f"remaining{len(tokens)} prompt tokens to eval",file=sys.stderr,)# Reset the model stateifreset:self.reset()# # Reset the grammar# if grammar is not None:#     grammar.reset()sample_idx=self.n_tokens+len(tokens)-1tokens=list(tokens)# Eval and samplewhileTrue:self.eval(tokens)whilesample_idx<self.n_tokens:token=self.sample(top_k=top_k,top_p=top_p,min_p=min_p,typical_p=typical_p,temp=temp,repeat_penalty=repeat_penalty,frequency_penalty=frequency_penalty,presence_penalty=presence_penalty,tfs_z=tfs_z,mirostat_mode=mirostat_mode,mirostat_tau=mirostat_tau,mirostat_eta=mirostat_eta,logits_processor=logits_processor,grammar=grammar,penalize_nl=penalize_nl,idx=sample_idx,)sample_idx+=1ifstopping_criteriaisnotNoneandstopping_criteria(self._input_ids[:sample_idx],self._scores[sample_idx-self.n_tokens,:]):returntokens_or_none=yieldtokentokens.clear()tokens.append(token)iftokens_or_noneisnotNone:tokens.extend(tokens_or_none)ifsample_idx<self.n_tokensandtoken!=self._input_ids[sample_idx]:self.n_tokens=sample_idxself._ctx.kv_cache_seq_rm(-1,self.n_tokens,-1)breakifself.draft_modelisnotNone:self.input_ids[self.n_tokens:self.n_tokens+len(tokens)]=tokensdraft_tokens=self.draft_model(self.input_ids[:self.n_tokens+len(tokens)])tokens.extend(draft_tokens.astype(int)[:self._n_ctx-self.n_tokens-len(tokens)])defcreate_embedding(self,input:Union[str,List[str]],model:Optional[str]=None)->CreateEmbeddingResponse:"""Embed a string.        Args:            input: The utf-8 encoded string to embed.        Returns:            An embedding object.        """model_name:str=modelifmodelisnotNoneelseself.model_pathinput=inputifisinstance(input,list)else[input]# get numeric embeddingsembeds:Union[List[List[float]],List[List[List[float]]]]total_tokens:intembeds,total_tokens=self.embed(input,return_count=True)# type: ignore# convert to CreateEmbeddingResponsedata:List[Embedding]=[{"object":"embedding","embedding":emb,"index":idx,}foridx,embinenumerate(embeds)]return{"object":"list","data":data,"model":model_name,"usage":{"prompt_tokens":total_tokens,"total_tokens":total_tokens,},}defembed(self,input:Union[str,List[str]],normalize:bool=False,truncate:bool=True,return_count:bool=False,):"""Embed a string.        Args:            input: The utf-8 encoded string to embed.        Returns:            A list of embeddings        """n_embd=self.n_embd()n_batch=self.n_batch# get pooling informationpooling_type=self.pooling_type()logits_all=pooling_type==llama_cpp.LLAMA_POOLING_TYPE_NONEifself.context_params.embeddingsisFalse:raiseRuntimeError("Llama model must be created with embedding=True to call this method")ifself.verbose:llama_cpp.llama_perf_context_reset(self._ctx.ctx)ifisinstance(input,str):inputs=[input]else:inputs=input# reset batchself._batch.reset()# decode and fetch embeddingsdata:Union[List[List[float]],List[List[List[float]]]]=[]defdecode_batch(seq_sizes:List[int]):llama_cpp.llama_kv_self_clear(self._ctx.ctx)self._ctx.decode(self._batch)self._batch.reset()# store embeddingsifpooling_type==llama_cpp.LLAMA_POOLING_TYPE_NONE:pos:int=0fori,sizeinenumerate(seq_sizes):ptr=llama_cpp.llama_get_embeddings(self._ctx.ctx)embedding:List[List[float]]=[ptr[pos+j*n_embd:pos+(j+1)*n_embd]forjinrange(size)]ifnormalize:embedding=[internals.normalize_embedding(e)foreinembedding]data.append(embedding)pos+=sizeelse:foriinrange(len(seq_sizes)):ptr=llama_cpp.llama_get_embeddings_seq(self._ctx.ctx,i)embedding:List[float]=ptr[:n_embd]ifnormalize:embedding=internals.normalize_embedding(embedding)data.append(embedding)# init statetotal_tokens=0s_batch=[]t_batch=0p_batch=0# accumulate batches and encodefortextininputs:tokens=self.tokenize(text.encode("utf-8"))iftruncate:tokens=tokens[:n_batch]n_tokens=len(tokens)total_tokens+=n_tokens# check for overrunifn_tokens>n_batch:raiseValueError(f"Requested tokens ({n_tokens}) exceed batch size of{n_batch}")# time to eval batchift_batch+n_tokens>n_batch:decode_batch(s_batch)s_batch=[]t_batch=0p_batch=0# add to batchself._batch.add_sequence(tokens,p_batch,logits_all)# update batch statss_batch.append(n_tokens)t_batch+=n_tokensp_batch+=1# hanlde last batchdecode_batch(s_batch)ifself.verbose:llama_cpp.llama_perf_context_print(self._ctx.ctx)output=data[0]ifisinstance(input,str)elsedatallama_cpp.llama_kv_self_clear(self._ctx.ctx)self.reset()ifreturn_count:returnoutput,total_tokenselse:returnoutputdef_create_completion(self,prompt:Union[str,List[int]],suffix:Optional[str]=None,max_tokens:Optional[int]=16,temperature:float=0.8,top_p:float=0.95,min_p:float=0.05,typical_p:float=1.0,logprobs:Optional[int]=None,echo:bool=False,stop:Optional[Union[str,List[str]]]=[],frequency_penalty:float=0.0,presence_penalty:float=0.0,repeat_penalty:float=1.0,top_k:int=40,stream:bool=False,seed:Optional[int]=None,tfs_z:float=1.0,mirostat_mode:int=0,mirostat_tau:float=5.0,mirostat_eta:float=0.1,model:Optional[str]=None,stopping_criteria:Optional[StoppingCriteriaList]=None,logits_processor:Optional[LogitsProcessorList]=None,grammar:Optional[LlamaGrammar]=None,logit_bias:Optional[Dict[int,float]]=None,)->Union[Iterator[CreateCompletionResponse],Iterator[CreateCompletionStreamResponse]]:assertsuffixisNoneorsuffix.__class__isstrcompletion_id:str=f"cmpl-{str(uuid.uuid4())}"created:int=int(time.time())bos_token_id:int=self.token_bos()cls_token_id:int=self._model.token_cls()sep_token_id:int=self._model.token_sep()prefix_token_id:int=0# self._model.token_prefix() # TODO: Fixmiddle_token_id:int=0# self._model.token_middle() # TODO: Fixsuffix_token_id:int=0# self._model.token_suffix() # TODO: Fixadd_space_prefix:bool=(self.metadata.get("tokenizer.ggml.add_space_prefix","true")=="true")bos_tokens:List[int]=[cls_token_idifcls_token_id!=-1elsebos_token_id]eos_tokens:List[int]=[sep_token_idifsep_token_id!=-1elseself.token_eos()]if((isinstance(prompt,list)andsuffixisNone)ornotself._model.add_bos_token()orbos_tokens[:1]==[-1]):bos_tokens=[]if(isinstance(prompt,list)andsuffixisNone)or(notself._model.add_eos_token()andsep_token_id==-1):eos_tokens=[]suffix_space_prefix:int=0# Tokenizer hack to remove leading spaceifadd_space_prefixandsuffix_token_id>=0andsuffix:suffix="☺"+suffixsuffix_space_prefix=2# If prompt is empty, initialize completion with BOS token to avoid# detokenization including a space at the beginning of the completioncompletion_tokens:List[int]=[]iflen(prompt)>0else[bos_token_id]# Add blank space to start of prompt to match OG llama tokenizerprefix_tokens:List[int]=([prefix_token_id]ifprefix_token_id>=0andsuffixisnotNoneelse[])+((self.tokenize(prompt.encode("utf-8"),add_bos=False,special=(prefix_token_id<0orsuffixisNone),)ifprompt!=""else[])ifisinstance(prompt,str)elseprompt)suffix_tokens:List[int]=(([suffix_token_id]+(self.tokenize(suffix.encode("utf-8"),add_bos=False,special=False)[suffix_space_prefix:]ifsuffixelse[]))ifsuffix_token_id>=0andsuffixisnotNoneelse[])middle_tokens:List[int]=([middle_token_id]ifmiddle_token_id>=0andsuffixisnotNoneelse[])prompt_tokens:List[int]=(bos_tokens+((suffix_tokens+prefix_tokens+middle_tokens)ifself.spm_infillelse(prefix_tokens+suffix_tokens+middle_tokens))+eos_tokens)text:bytes=b""returned_tokens:int=0stop=(stopifisinstance(stop,list)else[stop]ifisinstance(stop,str)else[])model_name:str=modelifmodelisnotNoneelseself.model_pathifprompt_tokens[:2]==[self.token_bos()]*2:warnings.warn(f'Detected duplicate leading "{self._model.token_get_text(self.token_bos())}" in prompt, this will likely reduce response quality, consider removing it...',RuntimeWarning,)# NOTE: This likely doesn't work correctly for the first token in the prompt# because of the extra space added to the start of the prompt_tokensiflogit_biasisnotNone:logit_bias_map={int(k):float(v)fork,vinlogit_bias.items()}deflogit_bias_processor(input_ids:npt.NDArray[np.intc],scores:npt.NDArray[np.single],)->npt.NDArray[np.single]:new_scores=np.copy(scores)# Does it make sense to copy the whole array or can we just overwrite the original one?forinput_id,scoreinlogit_bias_map.items():new_scores[input_id]=score+scores[input_id]returnnew_scores_logit_bias_processor=LogitsProcessorList([logit_bias_processor])iflogits_processorisNone:logits_processor=_logit_bias_processorelse:logits_processor=logits_processor.extend(_logit_bias_processor)ifself.verbose:self._ctx.reset_timings()iflen(prompt_tokens)>=self._n_ctx:raiseValueError(f"Requested tokens ({len(prompt_tokens)}) exceed context window of{llama_cpp.llama_n_ctx(self.ctx)}")ifmax_tokensisNoneormax_tokens<=0:# Unlimited, depending on n_ctx.max_tokens=self._n_ctx-len(prompt_tokens)# Truncate max_tokens if requested tokens would exceed the context windowmax_tokens=(max_tokensifmax_tokens+len(prompt_tokens)<self._n_ctxelse(self._n_ctx-len(prompt_tokens)))ifstop!=[]:stop_sequences=[s.encode("utf-8")forsinstop]else:stop_sequences=[]iflogprobsisnotNoneandself._logits_allisFalse:raiseValueError("logprobs is not supported for models created with logits_all=False")ifself.cache:try:cache_item=self.cache[prompt_tokens]cache_prefix_len=Llama.longest_token_prefix(cache_item.input_ids.tolist(),prompt_tokens)eval_prefix_len=Llama.longest_token_prefix(self._input_ids.tolist(),prompt_tokens)ifcache_prefix_len>eval_prefix_len:self.load_state(cache_item)ifself.verbose:print("Llama._create_completion: cache hit",file=sys.stderr)exceptKeyError:ifself.verbose:print("Llama._create_completion: cache miss",file=sys.stderr)ifseedisnotNone:self.set_seed(seed)else:self.set_seed(random.Random(self._seed).randint(0,2**32))finish_reason="length"multibyte_fix=0fortokeninself.generate(prompt_tokens,top_k=top_k,top_p=top_p,min_p=min_p,typical_p=typical_p,temp=temperature,tfs_z=tfs_z,mirostat_mode=mirostat_mode,mirostat_tau=mirostat_tau,mirostat_eta=mirostat_eta,frequency_penalty=frequency_penalty,presence_penalty=presence_penalty,repeat_penalty=repeat_penalty,stopping_criteria=stopping_criteria,logits_processor=logits_processor,grammar=grammar,):ifllama_cpp.llama_token_is_eog(self._model.vocab,token):text=self.detokenize(completion_tokens,prev_tokens=prompt_tokens)finish_reason="stop"breakcompletion_tokens.append(token)all_text=self.detokenize(completion_tokens,prev_tokens=prompt_tokens)# Contains multi-byte UTF8fork,charinenumerate(all_text[-3:]):k=3-kfornum,patternin[(2,192),(3,224),(4,240)]:# Bitwise AND checkifnum>kandpattern&char==pattern:multibyte_fix=num-k# Stop incomplete bytes from passingifmultibyte_fix>0:multibyte_fix-=1continueany_stop=[sforsinstop_sequencesifsinall_text]iflen(any_stop)>0:first_stop=any_stop[0]text=all_text[:all_text.index(first_stop)]finish_reason="stop"breakifstream:remaining_tokens=completion_tokens[returned_tokens:]remaining_text=self.detokenize(remaining_tokens,prev_tokens=prompt_tokens+completion_tokens[:returned_tokens],)remaining_length=len(remaining_text)# We want to avoid yielding any characters from# the generated text if they are part of a stop# sequence.first_stop_position=0forsinstop_sequences:foriinrange(min(len(s),remaining_length),0,-1):ifremaining_text.endswith(s[:i]):ifi>first_stop_position:first_stop_position=ibreaktoken_end_position=0iflogprobsisnotNone:# not sure how to handle this branch when dealing# with CJK output, so keep it unchangedfortokeninremaining_tokens:iftoken==bos_token_id:continuetoken_end_position+=len(self.detokenize([token],prev_tokens=prompt_tokens+completion_tokens[:returned_tokens],))# Check if stop sequence is in the tokeniftoken_end_position>(remaining_length-first_stop_position):breaktoken_str=self.detokenize([token],prev_tokens=prompt_tokens+completion_tokens[:returned_tokens],).decode("utf-8",errors="ignore")text_offset=len(prompt)+len(self.detokenize(completion_tokens[:returned_tokens],prev_tokens=prompt_tokens+completion_tokens[:returned_tokens],).decode("utf-8",errors="ignore"))token_offset=len(prompt_tokens)+returned_tokenslogits=self._scores[token_offset-1,:]current_logprobs=Llama.logits_to_logprobs(logits).tolist()sorted_logprobs=list(sorted(zip(current_logprobs,range(len(current_logprobs))),reverse=True,))top_logprob={self.detokenize([i]).decode("utf-8",errors="ignore"):logprobforlogprob,iinsorted_logprobs[:logprobs]}top_logprob.update({token_str:current_logprobs[int(token)]})logprobs_or_none={"tokens":[self.detokenize([token],prev_tokens=prompt_tokens+completion_tokens[:returned_tokens],).decode("utf-8",errors="ignore")],"text_offset":[text_offset],"token_logprobs":[current_logprobs[int(token)]],"top_logprobs":[top_logprob],}returned_tokens+=1yield{"id":completion_id,"object":"text_completion","created":created,"model":model_name,"choices":[{"text":self.detokenize([token],prev_tokens=prompt_tokens+completion_tokens[:returned_tokens],).decode("utf-8",errors="ignore"),"index":0,"logprobs":logprobs_or_none,"finish_reason":None,}],}else:whilelen(remaining_tokens)>0:decode_success=Falseforiinrange(1,len(remaining_tokens)+1):try:bs=self.detokenize(remaining_tokens[:i],prev_tokens=prompt_tokens+completion_tokens[:returned_tokens],)ts=bs.decode("utf-8")decode_success=TruebreakexceptUnicodeError:passelse:breakifnotdecode_success:# all remaining tokens cannot be decoded to a UTF-8 characterbreaktoken_end_position+=len(bs)iftoken_end_position>(remaining_length-first_stop_position):breakremaining_tokens=remaining_tokens[i:]returned_tokens+=iyield{"id":completion_id,"object":"text_completion","created":created,"model":model_name,"choices":[{"text":ts,"index":0,"logprobs":None,"finish_reason":None,}],}iflen(completion_tokens)>=max_tokens:text=self.detokenize(completion_tokens,prev_tokens=prompt_tokens)finish_reason="length"breakifstopping_criteriaisnotNoneandstopping_criteria(self._input_ids,self._scores[-1,:]):text=self.detokenize(completion_tokens,prev_tokens=prompt_tokens)finish_reason="stop"ifself.verbose:self._ctx.print_timings()ifstream:remaining_tokens=completion_tokens[returned_tokens:]remaining_text=self.detokenize(remaining_tokens,prev_tokens=prompt_tokens+completion_tokens[:returned_tokens],)any_stop=[sforsinstop_sequencesifsinremaining_text]iflen(any_stop)>0:end=min(remaining_text.index(stop)forstopinany_stop)else:end=len(remaining_text)token_end_position=0fortokeninremaining_tokens:token_end_position+=len(self.detokenize([token],prev_tokens=prompt_tokens+completion_tokens[:returned_tokens],))logprobs_or_none:Optional[CompletionLogprobs]=NoneiflogprobsisnotNone:iftoken==bos_token_id:continuetoken_str=self.detokenize([token]).decode("utf-8",errors="ignore")text_offset=len(prompt)+len(self.detokenize(completion_tokens[:returned_tokens],prev_tokens=prompt_tokens+completion_tokens[:returned_tokens],))token_offset=len(prompt_tokens)+returned_tokens-1logits=self._scores[token_offset,:]current_logprobs=Llama.logits_to_logprobs(logits).tolist()sorted_logprobs=list(sorted(zip(current_logprobs,range(len(current_logprobs))),reverse=True,))top_logprob={self.detokenize([i]).decode("utf-8",errors="ignore"):logprobforlogprob,iinsorted_logprobs[:logprobs]}top_logprob.update({token_str:current_logprobs[int(token)]})logprobs_or_none={"tokens":[self.detokenize([token]).decode("utf-8",errors="ignore")],"text_offset":[text_offset],"token_logprobs":[current_logprobs[int(token)]],"top_logprobs":[top_logprob],}iftoken_end_position>=end:last_text=self.detokenize([token])iftoken_end_position==end-1:breakreturned_tokens+=1yield{"id":completion_id,"object":"text_completion","created":created,"model":model_name,"choices":[{"text":last_text[:len(last_text)-(token_end_position-end)].decode("utf-8",errors="ignore"),"index":0,"logprobs":logprobs_or_none,"finish_reason":None,}],}breakreturned_tokens+=1yield{"id":completion_id,"object":"text_completion","created":created,"model":model_name,"choices":[{"text":self.detokenize([token]).decode("utf-8",errors="ignore"),"index":0,"logprobs":logprobs_or_none,"finish_reason":None,}],}yield{"id":completion_id,"object":"text_completion","created":created,"model":model_name,"choices":[{"text":"","index":0,"logprobs":None,"finish_reason":finish_reason,}],}ifself.cache:ifself.verbose:print("Llama._create_completion: cache save",file=sys.stderr)self.cache[prompt_tokens+completion_tokens]=self.save_state()ifself.verbose:print("Llama._create_completion: cache saved",file=sys.stderr)returnifself.cache:ifself.verbose:print("Llama._create_completion: cache save",file=sys.stderr)self.cache[prompt_tokens+completion_tokens]=self.save_state()text_str=text.decode("utf-8",errors="ignore")ifecho:text_str=prompt+text_strifsuffix_token_id<0andsuffixisnotNone:text_str=text_str+suffixlogprobs_or_none:Optional[CompletionLogprobs]=NoneiflogprobsisnotNone:text_offset=0ifechoelselen(prompt)token_offset=0ifechoelselen(prompt_tokens[1:])text_offsets:List[int]=[]token_logprobs:List[Optional[float]]=[]tokens:List[str]=[]top_logprobs:List[Optional[Dict[str,float]]]=[]ifecho:# Remove leading BOS token if existsall_tokens=(prompt_tokens[1ifprompt_tokens[0]==self.token_bos()else0:]+completion_tokens)else:all_tokens=completion_tokensall_token_strs=[self.detokenize([token],prev_tokens=all_tokens[:i]).decode("utf-8",errors="ignore")fori,tokeninenumerate(all_tokens)]all_logprobs=Llama.logits_to_logprobs(self._scores)[token_offset:]# TODO: may be able to change this loop to use np.take_along_dimforidx,(token,token_str,logprobs_token)inenumerate(zip(all_tokens,all_token_strs,all_logprobs)):iftoken==bos_token_id:continuetext_offsets.append(text_offset+len(self.detokenize(all_tokens[:idx]).decode("utf-8",errors="ignore")))tokens.append(token_str)sorted_logprobs=list(sorted(zip(logprobs_token,range(len(logprobs_token))),reverse=True))token_logprobs.append(logprobs_token[int(token)])top_logprob:Optional[Dict[str,float]]={self.detokenize([i],prev_tokens=all_tokens[:idx]).decode("utf-8",errors="ignore"):logprobforlogprob,iinsorted_logprobs[:logprobs]}top_logprob.update({token_str:logprobs_token[int(token)]})top_logprobs.append(top_logprob)# Weird idosincracy of the OpenAI API where# token_logprobs and top_logprobs are null for# the first token.ifechoandlen(all_tokens)>0:token_logprobs[0]=Nonetop_logprobs[0]=Nonelogprobs_or_none={"tokens":tokens,"text_offset":text_offsets,"token_logprobs":token_logprobs,"top_logprobs":top_logprobs,}yield{"id":completion_id,"object":"text_completion","created":created,"model":model_name,"choices":[{"text":text_str,"index":0,"logprobs":logprobs_or_none,"finish_reason":finish_reason,}],"usage":{"prompt_tokens":len(prompt_tokens),"completion_tokens":len(completion_tokens),"total_tokens":len(prompt_tokens)+len(completion_tokens),},}defcreate_completion(self,prompt:Union[str,List[int]],suffix:Optional[str]=None,max_tokens:Optional[int]=16,temperature:float=0.8,top_p:float=0.95,min_p:float=0.05,typical_p:float=1.0,logprobs:Optional[int]=None,echo:bool=False,stop:Optional[Union[str,List[str]]]=[],frequency_penalty:float=0.0,presence_penalty:float=0.0,repeat_penalty:float=1.0,top_k:int=40,stream:bool=False,seed:Optional[int]=None,tfs_z:float=1.0,mirostat_mode:int=0,mirostat_tau:float=5.0,mirostat_eta:float=0.1,model:Optional[str]=None,stopping_criteria:Optional[StoppingCriteriaList]=None,logits_processor:Optional[LogitsProcessorList]=None,grammar:Optional[LlamaGrammar]=None,logit_bias:Optional[Dict[int,float]]=None,)->Union[CreateCompletionResponse,Iterator[CreateCompletionStreamResponse]]:"""Generate text from a prompt.        Args:            prompt: The prompt to generate text from.            suffix: A suffix to append to the generated text. If None, no suffix is appended.            max_tokens: The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.            temperature: The temperature to use for sampling.            top_p: The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751            min_p: The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841            typical_p: The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.            logprobs: The number of logprobs to return. If None, no logprobs are returned.            echo: Whether to echo the prompt.            stop: A list of strings to stop generation when encountered.            frequency_penalty: The penalty to apply to tokens based on their frequency in the prompt.            presence_penalty: The penalty to apply to tokens based on their presence in the prompt.            repeat_penalty: The penalty to apply to repeated tokens.            top_k: The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751            stream: Whether to stream the results.            seed: The seed to use for sampling.            tfs_z: The tail-free sampling parameter. Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.            mirostat_mode: The mirostat sampling mode.            mirostat_tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.            mirostat_eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.            model: The name to use for the model in the completion object.            stopping_criteria: A list of stopping criteria to use.            logits_processor: A list of logits processors to use.            grammar: A grammar to use for constrained sampling.            logit_bias: A logit bias to use.        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_or_chunks=self._create_completion(prompt=prompt,suffix=suffix,max_tokens=-1ifmax_tokensisNoneelsemax_tokens,temperature=temperature,top_p=top_p,min_p=min_p,typical_p=typical_p,logprobs=logprobs,echo=echo,stop=stop,frequency_penalty=frequency_penalty,presence_penalty=presence_penalty,repeat_penalty=repeat_penalty,top_k=top_k,stream=stream,seed=seed,tfs_z=tfs_z,mirostat_mode=mirostat_mode,mirostat_tau=mirostat_tau,mirostat_eta=mirostat_eta,model=model,stopping_criteria=stopping_criteria,logits_processor=logits_processor,grammar=grammar,logit_bias=logit_bias,)ifstream:chunks:Iterator[CreateCompletionStreamResponse]=completion_or_chunksreturnchunkscompletion:Completion=next(completion_or_chunks)# type: ignorereturncompletiondef__call__(self,prompt:str,suffix:Optional[str]=None,max_tokens:Optional[int]=16,temperature:float=0.8,top_p:float=0.95,min_p:float=0.05,typical_p:float=1.0,logprobs:Optional[int]=None,echo:bool=False,stop:Optional[Union[str,List[str]]]=[],frequency_penalty:float=0.0,presence_penalty:float=0.0,repeat_penalty:float=1.0,top_k:int=40,stream:bool=False,seed:Optional[int]=None,tfs_z:float=1.0,mirostat_mode:int=0,mirostat_tau:float=5.0,mirostat_eta:float=0.1,model:Optional[str]=None,stopping_criteria:Optional[StoppingCriteriaList]=None,logits_processor:Optional[LogitsProcessorList]=None,grammar:Optional[LlamaGrammar]=None,logit_bias:Optional[Dict[int,float]]=None,)->Union[CreateCompletionResponse,Iterator[CreateCompletionStreamResponse]]:"""Generate text from a prompt.        Args:            prompt: The prompt to generate text from.            suffix: A suffix to append to the generated text. If None, no suffix is appended.            max_tokens: The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.            temperature: The temperature to use for sampling.            top_p: The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751            min_p: The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841            typical_p: The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.            logprobs: The number of logprobs to return. If None, no logprobs are returned.            echo: Whether to echo the prompt.            stop: A list of strings to stop generation when encountered.            frequency_penalty: The penalty to apply to tokens based on their frequency in the prompt.            presence_penalty: The penalty to apply to tokens based on their presence in the prompt.            repeat_penalty: The penalty to apply to repeated tokens.            top_k: The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751            stream: Whether to stream the results.            seed: The seed to use for sampling.            tfs_z: The tail-free sampling parameter. Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.            mirostat_mode: The mirostat sampling mode.            mirostat_tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.            mirostat_eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.            model: The name to use for the model in the completion object.            stopping_criteria: A list of stopping criteria to use.            logits_processor: A list of logits processors to use.            grammar: A grammar to use for constrained sampling.            logit_bias: A logit bias to use.        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.        """returnself.create_completion(prompt=prompt,suffix=suffix,max_tokens=max_tokens,temperature=temperature,top_p=top_p,min_p=min_p,typical_p=typical_p,logprobs=logprobs,echo=echo,stop=stop,frequency_penalty=frequency_penalty,presence_penalty=presence_penalty,repeat_penalty=repeat_penalty,top_k=top_k,stream=stream,seed=seed,tfs_z=tfs_z,mirostat_mode=mirostat_mode,mirostat_tau=mirostat_tau,mirostat_eta=mirostat_eta,model=model,stopping_criteria=stopping_criteria,logits_processor=logits_processor,grammar=grammar,logit_bias=logit_bias,)defcreate_chat_completion(self,messages:List[ChatCompletionRequestMessage],functions:Optional[List[ChatCompletionFunction]]=None,function_call:Optional[ChatCompletionRequestFunctionCall]=None,tools:Optional[List[ChatCompletionTool]]=None,tool_choice:Optional[ChatCompletionToolChoiceOption]=None,temperature:float=0.2,top_p:float=0.95,top_k:int=40,min_p:float=0.05,typical_p:float=1.0,stream:bool=False,stop:Optional[Union[str,List[str]]]=[],seed:Optional[int]=None,response_format:Optional[ChatCompletionRequestResponseFormat]=None,max_tokens:Optional[int]=None,presence_penalty:float=0.0,frequency_penalty:float=0.0,repeat_penalty:float=1.0,tfs_z:float=1.0,mirostat_mode:int=0,mirostat_tau:float=5.0,mirostat_eta:float=0.1,model:Optional[str]=None,logits_processor:Optional[LogitsProcessorList]=None,grammar:Optional[LlamaGrammar]=None,logit_bias:Optional[Dict[int,float]]=None,logprobs:Optional[bool]=None,top_logprobs:Optional[int]=None,)->Union[CreateChatCompletionResponse,Iterator[CreateChatCompletionStreamResponse]]:"""Generate a chat completion from a list of messages.        Args:            messages: A list of messages to generate a response for.            functions: A list of functions to use for the chat completion.            function_call: A function call to use for the chat completion.            tools: A list of tools to use for the chat completion.            tool_choice: A tool choice to use for the chat completion.            temperature: The temperature to use for sampling.            top_p: The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751            top_k: The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751            min_p: The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841            typical_p: The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.            stream: Whether to stream the results.            stop: A list of strings to stop generation when encountered.            seed: The seed to use for sampling.            response_format: The response format to use for the chat completion. Use { "type": "json_object" } to contstrain output to only valid json.            max_tokens: The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.            presence_penalty: The penalty to apply to tokens based on their presence in the prompt.            frequency_penalty: The penalty to apply to tokens based on their frequency in the prompt.            repeat_penalty: The penalty to apply to repeated tokens.            tfs_z: The tail-free sampling parameter.            mirostat_mode: The mirostat sampling mode.            mirostat_tau: The mirostat sampling tau parameter.            mirostat_eta: The mirostat sampling eta parameter.            model: The name to use for the model in the completion object.            logits_processor: A list of logits processors to use.            grammar: A grammar to use.            logit_bias: A logit bias to use.        Returns:            Generated chat completion or a stream of chat completion chunks.        """handler=(self.chat_handlerorself._chat_handlers.get(self.chat_format)orllama_chat_format.get_chat_completion_handler(self.chat_format))returnhandler(llama=self,messages=messages,functions=functions,function_call=function_call,tools=tools,tool_choice=tool_choice,temperature=temperature,top_p=top_p,top_k=top_k,min_p=min_p,typical_p=typical_p,logprobs=logprobs,top_logprobs=top_logprobs,stream=stream,stop=stop,seed=seed,response_format=response_format,max_tokens=max_tokens,presence_penalty=presence_penalty,frequency_penalty=frequency_penalty,repeat_penalty=repeat_penalty,tfs_z=tfs_z,mirostat_mode=mirostat_mode,mirostat_tau=mirostat_tau,mirostat_eta=mirostat_eta,model=model,logits_processor=logits_processor,grammar=grammar,logit_bias=logit_bias,)defcreate_chat_completion_openai_v1(self,*args:Any,**kwargs:Any,):"""Generate a chat completion with return type based on the the OpenAI v1 API.        OpenAI python package is required to use this method.        You can install it with `pip install openai`.        Args:            *args: Positional arguments to pass to create_chat_completion.            **kwargs: Keyword arguments to pass to create_chat_completion.        Returns:            Generated chat completion or a stream of chat completion chunks.        """try:fromopenai.types.chatimportChatCompletion,ChatCompletionChunkstream=kwargs.get("stream",False)# type: ignoreassertisinstance(stream,bool)ifstream:return(ChatCompletionChunk(**chunk)forchunkinself.create_chat_completion(*args,**kwargs))# type: ignoreelse:returnChatCompletion(**self.create_chat_completion(*args,**kwargs))# type: ignoreexceptImportError:raiseImportError("To use create_chat_completion_openai_v1, you must install the openai package.""You can install it with `pip install openai`.")def__getstate__(self):returndict(model_path=self.model_path,# Model Paramsn_gpu_layers=self.model_params.n_gpu_layers,split_mode=self.model_params.split_mode,main_gpu=self.model_params.main_gpu,tensor_split=self.tensor_split,vocab_only=self.model_params.vocab_only,use_mmap=self.model_params.use_mmap,use_mlock=self.model_params.use_mlock,kv_overrides=self.kv_overrides,# Context Paramsseed=self._seed,n_ctx=self.context_params.n_ctx,n_batch=self.n_batch,n_ubatch=self.context_params.n_ubatch,n_threads=self.context_params.n_threads,n_threads_batch=self.context_params.n_threads_batch,rope_scaling_type=self.context_params.rope_scaling_type,pooling_type=self.context_params.pooling_type,rope_freq_base=self.context_params.rope_freq_base,rope_freq_scale=self.context_params.rope_freq_scale,yarn_ext_factor=self.context_params.yarn_ext_factor,yarn_attn_factor=self.context_params.yarn_attn_factor,yarn_beta_fast=self.context_params.yarn_beta_fast,yarn_beta_slow=self.context_params.yarn_beta_slow,yarn_orig_ctx=self.context_params.yarn_orig_ctx,logits_all=self._logits_all,embedding=self.context_params.embeddings,offload_kqv=self.context_params.offload_kqv,flash_attn=self.context_params.flash_attn,op_offloat=self.context_params.op_offloat,swa_full=self.context_params.swa_full,# Sampling Paramsno_perf=self.context_params.no_perf,last_n_tokens_size=self.last_n_tokens_size,# LoRA Paramslora_base=self.lora_base,lora_scale=self.lora_scale,lora_path=self.lora_path,# Backend Paramsnuma=self.numa,# Chat Format Paramschat_format=self.chat_format,chat_handler=self.chat_handler,# Speculative Decidngdraft_model=self.draft_model,# KV cache quantizationtype_k=self.context_params.type_k,type_v=self.context_params.type_v,# Miscspm_infill=self.spm_infill,verbose=self.verbose,)def__setstate__(self,state):self.__init__(**state)defsave_state(self)->LlamaState:ifself.verbose:print("Llama.save_state: saving llama state",file=sys.stderr)state_size=llama_cpp.llama_get_state_size(self._ctx.ctx)ifself.verbose:print(f"Llama.save_state: got state size:{state_size}",file=sys.stderr)llama_state=(ctypes.c_uint8*int(state_size))()ifself.verbose:print("Llama.save_state: allocated state",file=sys.stderr)n_bytes=llama_cpp.llama_copy_state_data(self._ctx.ctx,llama_state)ifself.verbose:print(f"Llama.save_state: copied llama state:{n_bytes}",file=sys.stderr)ifint(n_bytes)>int(state_size):raiseRuntimeError("Failed to copy llama state data")llama_state_compact=(ctypes.c_uint8*int(n_bytes))()llama_cpp.ctypes.memmove(llama_state_compact,llama_state,int(n_bytes))ifself.verbose:print(f"Llama.save_state: saving{n_bytes} bytes of llama state",file=sys.stderr,)returnLlamaState(scores=self._scores.copy(),input_ids=self.input_ids.copy(),n_tokens=self.n_tokens,llama_state=bytes(llama_state_compact),llama_state_size=n_bytes,seed=self._seed,)defload_state(self,state:LlamaState)->None:# Only filling in up to `n_tokens` and then zero-ing out the restself.scores[:state.n_tokens,:]=state.scores.copy()rest=self.scores[state.n_tokens:,:]rest[rest>0]=0.0self.input_ids=state.input_ids.copy()self.n_tokens=state.n_tokensself._seed=state.seedstate_size=state.llama_state_sizeLLamaStateArrayType=ctypes.c_uint8*state_sizellama_state=LLamaStateArrayType.from_buffer_copy(state.llama_state)ifllama_cpp.llama_set_state_data(self._ctx.ctx,llama_state)!=state_size:raiseRuntimeError("Failed to set llama state data")defn_ctx(self)->int:"""Return the context window size."""returnself._ctx.n_ctx()defn_embd(self)->int:"""Return the embedding size."""returnself._model.n_embd()defn_vocab(self)->int:"""Return the vocabulary size."""returnself._model.n_vocab()deftokenizer(self)->LlamaTokenizer:"""Return the llama tokenizer for this model."""returnLlamaTokenizer(self)deftoken_eos(self)->int:"""Return the end-of-sequence token."""returnself._model.token_eos()deftoken_bos(self)->int:"""Return the beginning-of-sequence token."""returnself._model.token_bos()deftoken_nl(self)->int:"""Return the newline token."""returnself._model.token_nl()defpooling_type(self)->str:"""Return the pooling type."""returnself._ctx.pooling_type()defclose(self)->None:"""Explicitly free the model from memory."""self._stack.close()def__del__(self)->None:self.close()@staticmethoddeflogits_to_logprobs(logits:Union[npt.NDArray[np.single],List],axis:int=-1)->npt.NDArray[np.single]:# https://docs.scipy.org/doc/scipy/reference/generated/scipy.special.log_softmax.htmllogits_maxs:np.ndarray=np.amax(logits,axis=axis,keepdims=True)iflogits_maxs.ndim>0:logits_maxs[~np.isfinite(logits_maxs)]=0elifnotnp.isfinite(logits_maxs):logits_maxs=0subtract_maxs=np.subtract(logits,logits_maxs,dtype=np.single)exp=np.exp(subtract_maxs)# Suppress warnings about log of zerowithnp.errstate(divide="ignore"):summed=np.sum(exp,axis=axis,keepdims=True)out=np.log(summed)returnsubtract_maxs-out@staticmethoddeflongest_token_prefix(a:Sequence[int],b:Sequence[int]):longest_prefix=0for_a,_binzip(a,b):if_a==_b:longest_prefix+=1else:breakreturnlongest_prefix@classmethoddeffrom_pretrained(cls,repo_id:str,filename:Optional[str],additional_files:Optional[List]=None,local_dir:Optional[Union[str,os.PathLike[str]]]=None,local_dir_use_symlinks:Union[bool,Literal["auto"]]="auto",cache_dir:Optional[Union[str,os.PathLike[str]]]=None,**kwargs:Any,)->"Llama":"""Create a Llama model from a pretrained model name or path.        This method requires the huggingface-hub package.        You can install it with `pip install huggingface-hub`.        Args:            repo_id: The model repo id.            filename: A filename or glob pattern to match the model file in the repo.            additional_files: A list of filenames or glob patterns to match additional model files in the repo.            local_dir: The local directory to save the model to.            local_dir_use_symlinks: Whether to use symlinks when downloading the model.            **kwargs: Additional keyword arguments to pass to the Llama constructor.        Returns:            A Llama model."""try:fromhuggingface_hubimporthf_hub_download,HfFileSystemfromhuggingface_hub.utilsimportvalidate_repo_idexceptImportError:raiseImportError("Llama.from_pretrained requires the huggingface-hub package. ""You can install it with `pip install huggingface-hub`.")validate_repo_id(repo_id)hffs=HfFileSystem()files=[file["name"]ifisinstance(file,dict)elsefileforfileinhffs.ls(repo_id,recursive=True)]# split each file into repo_id, subfolder, filenamefile_list:List[str]=[]forfileinfiles:rel_path=Path(file).relative_to(repo_id)file_list.append(str(rel_path))# find the only/first shard file:matching_files=[fileforfileinfile_listiffnmatch.fnmatch(file,filename)]# type: ignoreiflen(matching_files)==0:raiseValueError(f"No file found in{repo_id} that match{filename}\n\n"f"Available Files:\n{json.dumps(file_list)}")iflen(matching_files)>1:raiseValueError(f"Multiple files found in{repo_id} matching{filename}\n\n"f"Available Files:\n{json.dumps(files)}")(matching_file,)=matching_filessubfolder=str(Path(matching_file).parent)filename=Path(matching_file).name# download the filehf_hub_download(repo_id=repo_id,filename=filename,subfolder=subfolder,local_dir=local_dir,local_dir_use_symlinks=local_dir_use_symlinks,cache_dir=cache_dir,)ifadditional_files:foradditonal_file_nameinadditional_files:# find the additional shard file:matching_additional_files=[fileforfileinfile_listiffnmatch.fnmatch(file,additonal_file_name)]iflen(matching_additional_files)==0:raiseValueError(f"No file found in{repo_id} that match{additonal_file_name}\n\n"f"Available Files:\n{json.dumps(file_list)}")iflen(matching_additional_files)>1:raiseValueError(f"Multiple files found in{repo_id} matching{additonal_file_name}\n\n"f"Available Files:\n{json.dumps(files)}")(matching_additional_file,)=matching_additional_files# download the additional filehf_hub_download(repo_id=repo_id,filename=matching_additional_file,subfolder=subfolder,local_dir=local_dir,local_dir_use_symlinks=local_dir_use_symlinks,cache_dir=cache_dir,)iflocal_dirisNone:model_path=hf_hub_download(repo_id=repo_id,filename=filename,subfolder=subfolder,local_dir=local_dir,local_dir_use_symlinks=local_dir_use_symlinks,cache_dir=cache_dir,local_files_only=True,)else:model_path=os.path.join(local_dir,filename)# loading the first file of a sharded GGUF loads all remaining shard files in the subfolderreturncls(model_path=model_path,**kwargs,)