.`ibABddlmZddlZddlZddlZddlZddlmZddl Z ddl Z ddl Z ddlmZddlmZddlmZddlmZddlmZddlmZmZer!ddlZddlmZddlm Z ddl!Z"dd l#m$Z$dd l%m&Z&nTed e'd Z"ed e'dZede'dZ ede'dZee(Z)dZ*d4dZ+GddZ,d5d!Z-d"Z.e j/d#Z0e j/d$Z1d6d*Z2d7d,Z3d8d/Z4d9d0Z5d:d3Z6dS);) annotationsN) TYPE_CHECKING)LRUCache)Cache) init_logger) LazyLoader) GrammarOutputSchedulerOutput) TokenizerLike) InputBatchxgrxgrammaroc outlines_core file_utilsztransformers.file_utilsconvert_slow_tokenizerz#transformers.convert_slow_tokenizerscheduler_outputr grammar_outputr input_batchr logits torch.TensorreturnNonec|j}i}d}|j}t|j}t |jD]9\} } | |z} |t || dz }| |vr| || <:g} tj |j d|j dfd|j } d}|jD]} t || d}|| x}=td|zD]*}||z}|||z| |<| |+|d|zz }tj| |jd}t | |j dk}d}|s>tj| tjd d }||jd}t)j||| dS) a- Apply grammar bitmask to output logits of the model with xgrammar function. Args: scheduler_output (SchedulerOutput): The result of engine scheduling. input_batch (InputBatch): The input of model runner. logits (torch.Tensor): The output logits of model forward. r)shape fill_valuedtypeNT) non_blockingcpu)r device pin_memory)indices)grammar_bitmaskscheduled_spec_decode_tokenssetstructured_output_request_ids enumeratereq_idslengetnpfullrr rangeappendtorch from_numpytor#tensorint32r apply_token_bitmask_inplace)rrrrr&struct_out_req_batch_indicescumulative_offset spec_tokensstruct_out_req_ids batch_indexreq_id logit_index out_indicessorted_bitmaskcumulative_indexnum_spec_tokens logit_idxi bitmask_indexskip_out_indices index_tensors s/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/vllm/v1/structured_output/utils.pyapply_grammar_bitmaskrI,s6 %4O46 "?K^IJJ()<==?? V!$55 S! ( 0KW|A 5a 89#N  >00koofb99::599&AA AI N1.// 2 2 )A 0?@PST@T0U}-""=1111A//&~6699 D:O;''6<?:L I| u{5T   $v}4HH #FO\RRRRRRceZdZdZddZdS) OutlinesVocabularyzo Wrapper class for `outlines_core.Vocabulary`, which allows us to store a hash with the vocabulary vocabulary oc.Vocabularyrrc||_tj|d}t |d}||_dS)Nutf-8)innerhashlibsha256__repr__encode hexdigestint_hash)selfrMhex_strhash_ints rH__init__zOutlinesVocabulary.__init__sZ .!4!4!6!6!=!=g!F!FGGQQSSw## rJN)rMrNrr)__name__ __module__ __qualname____doc__r]rrJrHrLrLzs2 rJrLstrctjd}tjd}tjd}|r|S|r!tj|ddStj|r'|dkr!tj|ddSt j}tj|ddS)zFGet the context object that contains previously-computed return valuesOUTLINES_CACHE_DIRXDG_CACHE_HOME~z.cacheoutlines/)osgetenvpath expanduserjoinisdirtempfile gettempdir)outlines_cache_dirxdg_cache_homehome_dirtempdirs rHget_outlines_cache_pathrus#788Y/00Nw!!#&&H"!!Bw||NHjAAA  w}}X<8s??w||Hh ;;;!##G 7<<: 6 66rJc|t}tjrtdt |dd}t jd}| dd}||kr| | d||Std S) z3Get the Cache instance to be used for index cachingzEnabling outlines cache. This is an unbounded on-disk cache. It may consume a lot of disk space and should not be used with untrusted clients.noner)eviction_policy cull_limitr __version__N)maxsize) ruenvsVLLM_V1_USE_OUTLINES_CACHEloggerwarningr importlibmetadataversionr-clearr(r) cache_dircacheoutlines_versioncached_versions rHget_outlines_cachers())I &  2   iAFFF$-55oFF=$77 - - - KKMMM -!1222 C  rJz^<0x[0-9A-F]{2}>$u^.{0,6}�+.{0,6}$ tokenizerr eos_token_idrXdict[bytes, list[int]]c ` dtjD dfd }i}g}D]J\}}|jvr||}|rt |t tfrt |}n|dkr|dksd|vrt |s{t |r't t|ddd g}nO fd |D} d | vrtd |d |d|t | }n| d}||kr)||g|5||L|S)zCreate a map from vocabulary tokens to lists of equivalent token ids. Returns: A Dict of token string -> equivalent token ids ci|]\}}|| Srr).0kvs rH z'_reduced_vocabulary..s+A1rJtokenrbrc|g}t|tur|tjs|dkrd|zS|S)Nz<0x20> )convert_tokens_to_stringtyperb startswithrSPIECE_UNDERLINE)rstringrs rHconvert_token_to_stringz4_reduced_vocabulary..convert_token_to_strings^33UG<< KK3    !<==   <  rJu�rQc:g|]}|Sr)r-)rcunicode_to_bytess rH z'_reduced_vocabulary..s( H H HQ!1!5!5a!8!8 H H HrJNzCannot convert token `z` (z ) to bytes: rP)rrbrrb)rbytes_to_unicodeitems get_vocaball_special_tokens isinstancebytes bytearrayre_replacement_seqmatchre_llama_byte_tokenrX RuntimeErrorrV setdefaultr1) rrrrMempty_token_idsr token_idx token_str token_bytes byte_valsrs ` @rH_reduced_vocabularyrs+/@BBHHJJ      *,J!#O%//117799%.%.y I0 0 0 ++E22  .%%!344 8 $I.. x''EX,=,=I%%.@.F.Fy.Q.Q%',,U33 3"'U1Q3Z)<)<(=">">KK!I H H H% H H HIy((*DUDD!*DD8ADD #( "2"2KK'..w77 L((%%k266==iHHH  " "9 - - - - rJrNct|dr|jS t|dr|j|j}n tdt |dt ||}t tj||}||_|S#t$r&}tdt |d|d}~wwxYw)z2Get the `Vocabulary` object for a given tokenizer._outlines_vocabularyrNz?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.) hasattrrr ValueErrorrrrLr VocabularyAttributeError)rr reduced_vocabrMes rHget_outlines_vocabularyrs y011.-- 9n - - )2H2T$1LLVOOVVV  ,    ( lM(R(RSS )3 &  Y       sA9B C!B>>C grammar_strboolc|rt|tsdS|dD]4}tjdd|}|s-d|vrdS5dS)aV Check if grammar appears to use Lark syntax. Args: grammar_str: Input grammar string Returns: bool: True if grammar appears to be in Lark format, False otherwise Examples: >>> grammar_is_likely_lark("rule: 'abc'") True >>> grammar_is_likely_lark("rule ::= 'abc'") False F (#|//).*$z::=T)rrbsplitresubstrip)rlines rHgrammar_is_likely_larkr!s jc::u!!$''vlB--3355   D==55  4rJc t|tstdt||stdt }t }g}d*dd+d }d,d}fd|dD}d}t|dD]\}} | r| drd| vr | dddd} | | || }| dkrd}#t$r} td|d| d} ~ wwxYw|std| d|d} g} t|dD]\}} | s d| vr| ds| r-| | dd | | dd\} }| d} ||d| d|tjd d!|}||||g} n| dr| std"|d#| dd}||d$| d|tjd d!|}|||| |#t$r(} td%|d&t| | d} ~ wwxYw| r-| | dd | ||z d'hz }|r2td(d) t!|d |S)-a 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 emptyrrbrcRtjdd|S)z)Remove comments and whitespace from line.rr)rrr)rs rH clean_linez(convert_lark_to_ebnf..clean_line^s"vlB--33555rJtext rule_nameline_numrXrc|ddzdks|ddzdkrtd|d|dS)z Validate quote matching in text.'r"zMismatched quotes in z on line N)countr)rrrs rH check_quotesz*convert_lark_to_ebnf..check_quotesbs] ::c??Q ! # #tzz#':a'?'?SYSSSSTT T(@'?rJset[str]ctjdd|}tjdd|}ttjd|S)z"Extract rule references from text.z"[^"]*"rz[+*?()|\[\]{}]rz\b[a-zA-Z_][a-zA-Z0-9_]*\b)rrr(findall)rs rHextract_referencesz0convert_lark_to_ebnf..extract_referencesgsFvj"d++v'd332:;TBBCCCrJc&g|] }|Srr)rrrs rHrz(convert_lark_to_ebnf..os# B B B$ZZ   B B BrJrNr|:r?startzInvalid rule format on line z". Expected 'rule_name: definition'zNo valid rules found in grammar root ::= z ::=  | zrule 'rz '([^']*)'z"\1"zAlternative '|' on line z$ without a preceding rule definitionzalternative for rule 'zError on line z: rootz"Referenced rules are not defined: z, )rrbrrb)rrbrrbrrXrr)rrbrr)rrbrrrr(rr*radd IndexErrorr1rmrrupdatesorted)r defined_rulesreferenced_rules output_linesrrlines first_rulerrnamer current_rulecurrent_definition definitionalt_defundefined_rulesrs @rHconvert_lark_to_ebnfrAs( k3 ' 'OM${:K:KMMNNN     ;9:::EEMuuL6666UUUU DDDD C B B B+*;*;D*A*A B B BEJ#E1--$ ts++   $;; zz#q))!,2244::3??!!$'''%!%J7??!(J    78777   <:;;;0J00111L#E1--%K%K$  ! Kd{{4??3#7#7{ '''NNejj9K.L.LNN $(::c1#5#5 j#zz||11#66  Z)A,)A)A)A8LLLVL':FF  ''(:(::(F(FGGG&0&6&6&8&8%9""%% 3#$>8>>> qrr(..** ElEEEx&w@@ ''(:(:7(C(CDDD"))'222 K K KBhBB#a&&BBCC J KT|RR%**=O2P2PRRSSS'6&AO  U6/;R;R1S1S U U    99\ " ""s2A"D;; EEE)FL.. M 8#MM choice list[str]cpd dfd|D}ddd|Dz}|S) Nsrbrc.tjdd|S)z+Escape special characters in a EBNF string.z(["\\])z\\\1)rr)rs rHescape_ebnf_stringz-choice_as_grammar..escape_ebnf_stringsvj'1---rJc3.K|]}|VdS)Nr)rrrs rH z$choice_as_grammar..s/==))!,,======rJrrc3"K|] }d|dV dS)rNr)rrs rHrz$choice_as_grammar..s*&I&IAx1xxx&I&I&I&I&I&IrJ)rrbrrb)rm)rescaped_choicesgrammarrs @rHchoice_as_grammarrs\.... >===f===OEJJ&I&I&I&I&IIIIG NrJ) rr rr rr rrrr)rrb)rr rrXrr)rr rrN)rrbrr)rrbrrb)rrrrb)7 __future__rrSimportlib.metadatarrirotypingrnumpyr.regexrr2 cachetoolsr diskcacher vllm.envsr} vllm.loggerrvllm.utils.import_utilsrvllm.v1.core.sched.outputr r rr#transformers.convert_slow_tokenizerrtransformers.file_utilsrrr vllm.tokenizersr vllm.v1.worker.gpu_input_batchr globalsr^rCACHErIrLrurcompilerrrrrrrrrJrHrs#"""""   ######......DDDDDDDDHHHHHH000000------9999999 *UGGIIz 2 2C D''))_ 5 5BL''))5NOOJ'Z ''))-R X   KSKSKSKS\         77770!!!,!bj!566RZ 566CCCCL>@####DrJ