.`izddlZddlZddlmZddlmZmZmZddlm Z ddl m Z ddl m Z mZmZmZmZddlmZmZddlmZdd lmZdd lmZdd lmZmZdd lmZm Z m!Z!dd l"m#Z#m$Z$ddl%m&Z&ddl'm(Z(ddl)m*Z*ddl+m,Z,ddl-m.Z.m/Z/ddl0m1Z1ddl2m3Z3ddl4m5Z5m6Z6ddl7m8Z8ddl9m:Z:ddl;mZ>m?Z?ddl@mAZAddlBmCZCddlDmEZEeeFZGGddZHdS)N)Mapping)AnyLiteralcast) VllmConfig)VLLMValidationError)ProcessorInputs PromptTypeSingletonInputsSingletonPrompt TextPrompt)"is_explicit_encoder_decoder_promptsplit_enc_dec_inputs)InputPreprocessor) init_logger) LoRARequest)MULTIMODAL_REGISTRYMultiModalRegistry)MultiModalDataDictMultiModalFeatureSpecMultiModalUUIDDict)ModalityDataItemsMultiModalDataItems)set_request_id)argsort_mm_positions) PoolingParams) RendererLike) _SAMPLING_EPSSamplingParams) TokenizerLike)MistralTokenizer)&length_from_prompt_token_ids_or_embeds random_uuid)set_default_torch_num_threads)EngineCoreRequest)MultiModalCacheStats)&has_guidance_unsupported_json_featuresvalidate_guidance_grammar)5validate_structured_output_request_lm_format_enforcer)+validate_structured_output_request_outlines)validate_xgrammar_grammarcpeZdZefdededdfdZededzfdZ defdZ ede fdZ d e ddfd Zd e ddfd Zd e ddfd Zd e ddfd Zd e ezfdZdedefdZdeddfdZdeddfdZdedzddfdZd e ddfdZdededzfdZdededzfdZde dede!dzfdZ"de dedzde fdZ#e$de%fdZ& d2de ded e ezd!e'dzdedzd"e(e e)fdzd#e*e e fdzd$e+d%e+dzd&e,de%fd'Z-d(e.dzd)e.fd*Z/d+e.d,e0d-fd.Z1de2dzfd/Z3d3d0Z4d3d1Z5dS)4InputProcessor vllm_config mm_registryreturnNc@||_|j|_|j|_|j|_|j|_|j|_||_|||_ t|j|j ||j |_ dS)N)mm_processor_cache) r. model_config cache_config lora_configstructured_outputs_configtry_get_generation_configgeneration_config_fieldsr/processor_cache_from_configr2robservability_configinput_preprocessor)selfr.r/s r/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/vllm/v1/engine/input_processor.py__init__zInputProcessor.__init__8s ''4'4&2)4)N&(,(9(S(S(U(U%&"-"I"I+"V"V"3    , #6 # # # c|jjSN)r; tokenizerr<s r=rBzInputProcessor.tokenizerOs&00r?c4|jSrA)r; get_tokenizerrCs r=rEzInputProcessor.get_tokenizerS&44666r?c|jjSrA)r;rendererrCs r=rHzInputProcessor.rendererVs&//r?paramsc|jj}|dkr|j}|jrD|j}|dkr|j}||krt d|d|d||jrD|j}|dkr|j}||krt d|d|d|dSdS)NzRequested sample logprobs of z%, which is greater than max allowed: logprobs parametervaluezRequested prompt logprobs of prompt_logprobs)r3 max_logprobsget_vocab_sizerLrrP)r<rIrQ num_logprobsnum_prompt_logprobss r=_validate_logprobsz!InputProcessor._validate_logprobsZs?(5 2  ,;;==L ? !?Lr!!#0??AA l**)ILII:FII(&   ! "("8 "b((&*&7&F&F&H&H#"\11)I4GII:FII/-   21r?c2|||||jdS|jstd|jdSt |jt fd|jDstddS)Nz(allowed_token_ids is not None and empty!c38K|]}d|cxkokncVdS)rN).0tid vocab_sizes r= z;InputProcessor._validate_sampling_params..s?MMS1((((j((((MMMMMMr?z1allowed_token_ids contains out-of-vocab token id!)_validate_structured_output_validate_logit_biasallowed_token_ids ValueErrorrBlenall)r<rIr[s @r=_validate_sampling_paramsz(InputProcessor._validate_sampling_params|s ((000 !!&)))  # + F' IGHH H > ! F(( MMMMF4LMMMMM RPQQ Q R Rr?c|jsdS|j}g}|jD]#}|dks||kr||$|rt d|d|d|dS)z:Validate logit_bias token IDs are within vocabulary range.Nrz token_id(s) z@ in logit_bias contain out-of-vocab token ids. Vocabulary size: logit_biasrM)rer3rRappendr)r<rIr[invalid_token_idstoken_ids r=r^z#InputProcessor._validate_logit_biass    F&5577 ) 3 3H!||x:55!((222  %I0II>>r?promptc t|trt|}tt|dpi}tt |dpi}|s|sdS|d|D}d|D}| | z}|D]s}ttttz||g}tttdz||g} t|dkrt| dkrVt|t| kr6td|d|d t| d t|d t|D]F\} } | ?| std |d | d|d| | td |d | d|d | d GJt| dkrtd |d|dudS)N)r{multi_modal_datamulti_modal_uuidsci|] \}}||| SrArXrYkvs r= z?InputProcessor._validate_singleton_mm_uuids.. ? ? ?daQr?cNi|]"\}}||t|tr|gn|#SrA)rrstrrs r=rz?InputProcessor._validate_singleton_mm_uuids..s@   1} jC((/ssa}}r?rzIf given, multi_modal_uuids[z,] must have same length as multi_modal_data[z ], but got z vs .zmulti_modal_data[z][z!] is empty but multi_modal_uuids[z ] is missing.)rrrr rrgetrrzitemskeysrlistrrar` enumerate) r<r{rtmm_uuidsmm_data_parsedmm_uuids_parsed modalitiesmodality data_items uuid_itemsiitems r=_validate_singleton_mm_uuidsz+InputProcessor._validate_singleton_mm_uuidss+ fc " " /v...F)6::6H+I+I+ORPP*FJJ7J,K,K,QrRR x  F-- ? ?gmmoo ? ? ?     ((   \\^^hmmoo5 "  H!DI-~/A/A(B/O/OJd3:.0C0CHb0Q0QRRJ:""z??Q&&3z??c*oo+M+M$GxGG;CGG":GG47 OOGGG )44  GAt|)",!OH!O!O!!O!O5=!O!O!O## &a=0",!TH!T!T!!T!T5=!T!TCD!T!T!T## z??a''$GHGG-5GGG(9  r?ct|r>||d|dx}||dSdS||dS)z Validate that user-provided multi_modal_uuids align with multi_modal_data in the incoming request prompt(s). Only checks lengths; `None` entries are allowed and will be auto-hashed downstream. encoder_promptdecoder_promptN)rr)r<r{ dec_prompts r=_validate_mm_uuidsz!InputProcessor._validate_mm_uuidss~ .f 5 5 6  - -f5E.F G G G$%566 C11*=====DC  - -f 5 5 5 5 5r? lora_requestc|dS|jstd|d|jtddSdS)NzGot lora_request z but LoRA is not enabled!zvLLM has deprecated support for supporting different tokenizers for different LoRAs. By default, vLLM uses base model's tokenizer. If you are using a LoRA with its own tokenizer, consider specifying `--tokenizer [lora_path]` to use the LoRA tokenizer.)r5r`rBlogger warning_once)r<rs r=_validate_lorazInputProcessor._validate_lorass   F KLKKK  > %   :      & %r?c|jr|jsdS|jjr|jrt d|jj}|jjx}r/||kr(|dkr |jjst d|d|dn ||j_t|jj tr)|jj st d|jj dt|jj tr1|jj dkrt d |d rt|n|d r > -4  V64;VVV  v08# > > U)17799R??STT T   j ) )@ ? %f - - - -    + += ? $.*:;;  ? &f = = = = =  " " 7 ? ? ? ? , , ,$.*:;;  ? B& I I I I D)&1115?)22 D D D #5 % >S!).#660# !%IN!;!;!*$J6$R$RMdn.>??D=D@GGG9CF-66.fEEEE9CF-63 D6;?F % 7 !//11111s.H B>K  K cZt|trdS|dS)Nr})rrrrr<r{s r=_extract_singleton_mm_dataz)InputProcessor._extract_singleton_mm_datas. fc " " 4zz,---r?ct|r||dS||S)Nr)rrrs r=_extract_mm_datazInputProcessor._extract_mm_datas@ -f 5 5 ;226:J3KLL L226:: :r? request_idc||}|sdS|d|D}fd|DS)aHBuild per-item multimodal hash overrides when enabled. In this case, multimodal data items are identified by their request id, modality and index rather than their content. Returns a dictionary of modality -> list[str] of overrides, or None if disabled or no multimodal data is present. Nci|] \}}||| SrArXrs r=rz8InputProcessor._maybe_build_mm_uuids..rr?cRi|]"\}fdt|D#S)c$g|] }dd| S)-rX)rYrrrs r= zCInputProcessor._maybe_build_mm_uuids...s/QQQ!*55x55!55QQQr?)range)rY data_countrrs @r=rz8InputProcessor._maybe_build_mm_uuids..sP   $* QQQQQuZ?P?PQQQ   r?)rrzrget_all_counts)r<rr{rtmm_itemss ` r=_maybe_build_mm_uuidsz$InputProcessor._maybe_build_mm_uuidss''// 4'' ? ?gmmoo ? ? ?      (0(?(?(A(A(G(G(I(I    r?mm_hashcH||j |jjs|S|jd|S)z When enable_tower_connector_lora is True, multi-modal embeddings vary depending on the LoRA request. Therefore, the mm_hash must be generated based on the LoRA request to prevent incorrect cache hits. N:)r5enable_tower_connector_lora lora_name)r<rrs r=_get_mm_identifierz!InputProcessor._get_mm_identifiers?  '#?(N(447444r?requestc|jtd|j|_|jdtd|_dS)zReplace the externally supplied request ID with an internal request ID that adds 8 random characters in order to ensure uniquness. NzkThe external_req_id field should not be set on EngineCoreRequests passed to vLLM; use the request_id field.rz.8)external_req_idr`rr#)rs r=assign_request_idz InputProcessor.assign_request_idsW  " .= #*"4 ' 7LL+--LLLr?rF arrival_timetokenization_kwargs trace_headersprioritydata_parallel_rank resumablec |||||jj} | j} | j} | jr| n| }| %d| cxkr|ksntd| d|d|tj}|j j r8|j j j dkr#|j j s|||}nX||t!|t"r,t%t&dz|d}nd}t+t,jdd}dt,jvrt0d|t5|5t7|5|j||| }dddn #1swxYwYdddn #1swxYwYdd lm}| ||| |j!}tE|\}}|#|||d d kr d}|d}n |d}d}d}d}t!|tHr||%}|j&$tO||}|j j(|z |_&|)|j*||j+|,|j+n|%}d}|d dkr|d}|d}|d}t[|}g}|D]f\} }!|| |!}"|.t_|| |!| |0|"||| |!|"gtc||||||||||d|| || S)Nrzdata_parallel_rank z is out of range [0, z).r~OMP_NUM_THREADS1zSOMP_NUM_THREADS is not set; defaulting Torch threads to %d for input preprocessing.)rr)current_platform)r{rIprocessed_inputstypeembeds prompt_embedsprompt_token_ids multimodal mm_kwargsmm_placeholders mm_hashes)datar identifier mm_positionr cache_salt)rrr mm_featuressampling_paramspooling_params eos_token_idrrrrrrr)2rrsr.parallel_configdata_parallel_sizedata_parallel_size_locallocal_engines_onlyr`timer3multimodal_configmm_processor_cache_gbr4enable_prefix_cachingrrrrdictrrrintosenvironr debug_oncerr$r; preprocessvllm.platformsrvalidate_requestget_eos_token_idr_validate_model_inputsrclone max_tokensr" max_model_lenupdate_from_generation_configr8rBupdate_from_tokenizerrrfrrr%)#r<rr{rIrrrrrrrrdp_size dp_local_size num_ranksr num_threadsrrrencoder_inputsdecoder_inputsrrrrseq_lenrdecoder_mm_inputsdecoder_mm_positionsdecoder_mm_hashessorted_mm_idxsridx base_mm_hashs# r=process_inputszInputProcessor.process_inputss( L))) f%%%*:!4'@ %4%GTMMW  )18J3V3V3V3VY3V3V3V3V5&855'0555   9;;L   / !3IQNN%;O11*fEEHH  # #F + + +&$'' &-vzz:M/N/N  "*..):C@@AA BJ . .   '    J ' '  )F{)S)S  040G0R0R$7!1S11                                433333))- *   .??AA )=>N)O)O& ##NNCCC & !X - -# *?;MM-.@A  M fn - - ,$llnnO)1@$m.2->-Lw-V*  9 9-|   ~)55dnEEE#\\^^N;? & !\ 1 1 .{ ; #12C#D .{ ;  22FGGNK!/   #0:3? "").x8=!)#'#:#:(($$%9$B3$G ,       !!-'#+)%%%%)),771'    s6 GG; GG GG GG"%G"rr cf|||d||ddS)Nencoder) prompt_typedecoder)_validate_model_input)r<rr s r=rz%InputProcessor._validate_model_inputs~sB  %  & &~9 & M M M "">y"IIIIIr? prompt_inputsr)rrc v|j}|ddkrdn|d}|ddkr|dnd}t||}|s.|dkr|jrn |ddkrntd|d|j}|Yt |pdd }|t |j|jd z krtd |d |jj} || kr|jr]|j j } | |} | j } | ||} | | }|jrdS|jrd}nd}td|d|d| d||| kr:|dkr6|js1|jjdkr#d}td|d|d| d|dSdSdSdS)NrrrrrzThe z prompt cannot be emptyrXr)defaultrjz Token id z is out of vocabularyrzMake sure that `max_model_len` is no smaller than the number of text tokens plus multimodal tokens. For image inputs, the number of image tokens depends on the number of images, and possibly their aspect ratios as well.zLMake sure that `max_model_len` is no smaller than the number of text tokens.z prompt (length z-) is longer than the maximum model length of z. rpoolingzoMake sure that `max_model_len` is no smaller than the number of text tokens (prompt + requested output tokens).ze) plus the number of requested output tokens (at least 1) is longer than the maximum model length of )r3r"is_multimodal_modelr`rBmax max_token_idrRrr;r/_get_model_cls_processor_factory_create_processing_ctxinfoskip_prompt_length_check runner_type)r<rrr3 prompt_idsr prompt_lenrB max_input_idmax_prompt_lenr/ model_cls factoriesctxmm_info suggestions r=rz$InputProcessor._validate_model_inputs ( V$00 D12 V$00 / * *  propertyr rBrErrHrrUrcr^rprrsrrrzr rr rrrr]rrrrrr staticmethodr%rfloatrrrrboolrr rrrr&r.r0r2rXr?r=r-r-7s+>   (       .1=4/111X17}77770,000X0       DRR RRRR& . &9.9999 ?'9?>Q????5?5t5555n666666 ;+=$&p2.p2Tp2p2p2p2d.%. d "....;z;6H46O;;;;     d "     255"D(5 5555$ M#4 M M M\ M"&*+/5926)-] ] ] ] . ] dl ] "D( ] "#s(^d2] sCx(4/] ]  $J] ]  ] ] ] ] ~J-4JFUJJJJb&b12 bbbbH73d:77771111,,,,,,r?r-)Irrcollections.abcrtypingrrr vllm.configrvllm.exceptionsr vllm.inputsr r r r r vllm.inputs.parserrvllm.inputs.preprocessr vllm.loggerrvllm.lora.requestrvllm.multimodalrrvllm.multimodal.inputsrrrvllm.multimodal.parserr"vllm.multimodal.processing.contextrvllm.multimodal.utilsrvllm.pooling_paramsrvllm.renderersrvllm.sampling_paramsrrvllm.tokenizersr vllm.tokenizers.mistralr! vllm.utilsr"r#vllm.utils.torch_utilsr$vllm.v1.enginer%vllm.v1.metrics.statsr&*vllm.v1.structured_output.backend_guidancer'r(4vllm.v1.structured_output.backend_lm_format_enforcerr)*vllm.v1.structured_output.backend_outlinesr**vllm.v1.structured_output.backend_xgrammarr+r3rr-rXr?r=rUs ######%%%%%%%%%%""""""//////WVVVVVVV444444######))))))CCCCCCCC IHHHHHHH======666666------''''''>>>>>>>>))))))444444JJJJJJJJ@@@@@@,,,,,,666666QPPPPP X  { ,{ ,{ ,{ ,{ ,{ ,{ ,{ ,{ ,{ ,r?