o )i3@sRddlmZddlZddlZddlZddlmZddlZ ddl m Z ddl m Z ddlmZddlmZddlmZerYddlZddlmZddlmmmZddlmZned ed Zed ed Zed edZeeZ dZ!GdddZ"d-ddZ#ddZ$e %dZ&e %dZ'd.ddZ(d/d!d"Z)d0d%d&Z*d1d'd(Z+d2d+d,Z,dS)3) annotationsN) TYPE_CHECKING)LRUCache)Cache) init_logger) LazyLoader) AnyTokenizeroc outlines_core file_utilsztransformers.file_utilstokenization_gpt2z*transformers.models.gpt2.tokenization_gpt2c@seZdZdZd ddZdS) OutlinesVocabularyzo Wrapper class for `outlines_core.Vocabulary`, which allows us to store a hash with the vocabulary vocabulary oc.VocabularyreturnNonecCs2||_t|d}t|d}||_dS)Nutf-8)innerhashlibsha256__repr__encode hexdigestint_hash)selfrhex_strZhash_intrk/home/app/PaddleOCR-VL-test/.venv_paddleocr/lib/python3.10/site-packages/vllm/v1/structured_output/utils.py__init__-s   zOutlinesVocabulary.__init__N)rrrr)__name__ __module__ __qualname____doc__r rrrrr 'sr rstrcCstd}td}tjd}|r|S|rtj|ddStj|r0|dkr0tj|ddSddl}|}tj|ddS) zFGet the context object that contains previously-computed return valuesZOUTLINES_CACHE_DIRXDG_CACHE_HOME~z.cacheZoutlines/rN)osgetenvpath expanduserjoinisdirtempfile gettempdir)Zoutlines_cache_dirZxdg_cache_homeZhome_dirr/tempdirrrrget_outlines_cache_path8s   r2cCsft}tjr.tdt|ddd}tjd}| dd}||kr&| | d||St dd S) z3Get the Cache instance to be used for index cachingzEnabling outlines cache. This is an unbounded on-disk cache. It may consume a lot of disk space and should not be used with untrusted clients.noner)Zeviction_policyZ cull_limitr __version__N)maxsize) r2envsZVLLM_V1_USE_OUTLINES_CACHEloggerwarningr importlibmetadataversiongetclearsetr) cache_dircacheZoutlines_versionZcached_versionrrrget_outlines_cacheRs     rBz^<0x[0-9A-F]{2}>$u^.{0,6}�+.{0,6}$ tokenizerr eos_token_idrdict[bytes, list[int]]c sddtDdfdd }i}g}D]n\}}|jvr&q||}|rt|ttfr8t|}n?d|vrrt |srt |rStt |d d d g}n$fd d |D} d| vrmt d|d|d|t| }n| d}||kr||g|q||q|S)zCreate a map from vocabulary tokens to lists of equivalent token ids. Returns: A Dict of token string -> equivalent token ids cSsi|]\}}||qSrr).0kvrrr tsz'_reduced_vocabulary..tokenr%rcs8|g}t|tur|tjs|dkrd|S|S)Nz<0x20> )Zconvert_tokens_to_stringtyper% startswithr ZSPIECE_UNDERLINE)rJstring)rCrrconvert_token_to_stringys   z4_reduced_vocabulary..convert_token_to_stringu�rcsg|]}|qSr)r=rFc)unicode_to_bytesrr sz'_reduced_vocabulary..NzCannot convert token `z` (z ) to bytes: r)rJr%rr%)r Zbytes_to_unicodeitemsZ get_vocabZall_special_tokens isinstancebytes bytearrayre_replacement_seqmatchre_llama_byte_tokenr RuntimeErrorr setdefaultappend) rCrDrOrZempty_token_idsrJZ token_idxZ token_strZ token_bytesZ byte_valsr)rCrTr_reduced_vocabularyjsD         r`rc Cst|dr|jSz+t|dr|jdur|j}n tdt|dt||}tt||}||_|WSt yJ}z tdt|d|d}~ww)z7Get the `Vocabulary` object for a given tokenizer. _outlines_vocabularyrDNz?Error during structured outputs setup for outlines: Tokenizer (zi) has no `eos_token_id` property, but `eos_token_id` is required for structured outputs to work properly.z,Cannot get the vocabulary of the tokenizer (z0). The tokenizer should have a get_vocab method.) hasattrrarD ValueErrorrLr`r r Z VocabularyAttributeError)rCrDZ reduced_vocabrerrrget_outlines_vocabularys:   rf grammar_strboolcCsL|rt|ts dS|dD]}tdd|}|sqd|vr#dSqdS)aV Check if grammar appears to use Lark syntax. Args: grammar_str: Input grammar string Returns: bool: True if grammar appears to be in Lark format, False otherwise Examples: >>> grammar_is_likely_lark("rule: 'abc'") True >>> grammar_is_likely_lark("rule ::= 'abc'") False F (#|//).*$z::=T)rWr%splitresubstrip)rglinerrrgrammar_is_likely_larksrqc st|tstdt||stdt}t}g}d.ddd/d d}d0dd}fdd|dD}d}t|dD]F\}} | rK| drLq@d| vrz | dddd} | | |duri| }| dkrod}Wq@t y} z td|d| d} ~ wwq@|std| d|d} g} t|dD]\}} | sqz|d| vr| ds| r| | d d! | | dd\} }| d} ||d"| d#|t d$d%|}||||g} n5| dr | std&|d'| dd}||d(| d#|t d$d%|}|||| |Wqty;} z td)|d*t| | d} ~ ww| rL| | d d! | ||d+h}|rbtd,d- t|d |S)1a Convert a Lark grammar string to EBNF format. EBNF reference: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md Lark grammar reference: https://lark-parser.readthedocs.io/en/latest/grammar.html Args: grammar_str: Input grammar in Lark format Returns: str: Converted grammar in EBNF format Examples: >>> print(convert_lark_to_ebnf("rule: 'hello'")) root ::= rule rule ::= "hello" zGrammar must be a string, got zGrammar string cannot be emptyrpr%rcSstdd|S)z)Remove comments and whitespace from line.rjrk)rmrnro)rprrr clean_line sz(convert_lark_to_ebnf..clean_linetext rule_nameline_numrrcSs<|dddks|dddkrtd|d|dS)z Validate quote matching in text.'r"zMismatched quotes in z on line N)countrc)rsrtrurrr check_quotess $z*convert_lark_to_ebnf..check_quotesset[str]cSs,tdd|}tdd|}ttd|S)z"Extract rule references from text.z"[^"]*"rkz[+*?()|\[\]{}]rKz\b[a-zA-Z_][a-zA-Z0-9_]*\b)rmrnr?findall)rsrrrextract_referencessz0convert_lark_to_ebnf..extract_referencescsg|]}|qSrr)rFrprrrrrUsz(convert_lark_to_ebnf..riN|:r?startzInvalid rule format on line z". Expected 'rule_name: definition'zNo valid rules found in grammar root ::= z ::=  | zrule 'rvz '([^']*)'z"\1"zAlternative '|' on line z$ without a preceding rule definitionzalternative for rule 'zError on line z: rootz"Referenced rules are not defined: z, )rpr%rr%)rsr%rtr%rurrr)rsr%rr{)rWr%rcrLror?rl enumeraterMadd IndexErrorr_r-rmrnupdatesorted)rgZ defined_rulesZreferenced_rules output_linesrzr}linesZ first_rulerurpnamereZ current_ruleZcurrent_definitionZ definitionZalt_defZundefined_rulesrr~rconvert_lark_to_ebnfs           rchoice list[str]cs8d ddfdd|D}dd d d|D}|S) Nsr%rcSstdd|S)z+Escape special characters in a EBNF string.z(["\\])z\\\1)rmrn)rrrrescape_ebnf_stringnsz-choice_as_grammar..escape_ebnf_stringc3s|]}|VqdS)NrrRrrr ssz$choice_as_grammar..rrcss|] }d|dVqdS)rxNrrRrrrrts)rr%rr%)r-)rZescaped_choicesZgrammarrrrchoice_as_grammarls r)rr%)rCrrDrrrE)rCrrr)rgr%rrh)rgr%rr%)rrrr%)- __future__rrimportlib.metadatar:r)typingrregexrmZ cachetoolsrZ diskcacherZ vllm.envsr7Z vllm.loggerrZ vllm.utilsrr r Ztransformers.file_utilsr Z*transformers.models.gpt2.tokenization_gpt2modelsZgpt2r Z!vllm.transformers_utils.tokenizerrglobalsr!r8ZCACHEr r2rBcompiler\rZr`rfrqrrrrrrsF             E }