§ .`ƒi·*ãó0—ddlZddlZddlZddlmZmZmZmZm Z ddl m Z Gd„d¦«Z ed¬¦«de d ejd ejd ed ed ededejdejfd„¦«Zed¬¦«dejdeded ededejf d„¦«ZdS)éN)Úget_num_threadsÚjitÚnjitÚprangeÚset_num_threads)Ú VllmConfigc ó—eZdZdefd„Zdededejdejdeeef d„Z dd eeedejdejd e e e j fee e e j fzd zdeeef d „Zd „Zd S)Ú NgramProposerÚ vllm_configcóþ—|j€J‚|jj€J‚|jj€J‚|jj|_|jj|_|jj|_|jj|_|j j }tj ||jftj ¬¦«|_tj |tj ¬¦«|_d|_|jj}t'j¦«}|r)t+d|dz¦«|_|xj|zc_nd|_| ggdztj dtj ¬¦«tj d|jftj ¬¦«¦«dS)N©Údtypei ééi)Úspeculative_configÚprompt_lookup_minÚprompt_lookup_maxÚmin_nÚmax_nÚnum_speculative_tokensÚkÚ model_configÚ max_model_lenÚscheduler_configÚ max_num_seqsÚnpÚzerosÚint32Úvalid_ngram_draftÚvalid_ngram_num_draftsÚnum_tokens_thresholdÚparallel_configÚtensor_parallel_sizeÚosÚ cpu_countÚminÚnum_numba_thread_availableÚpropose)Úselfr rÚtp_sizer%s úv/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/vllm/v1/spec_decode/ngram_proposer.pyÚ__init__zNgramProposer.__init__ sp€ØÔ-Ð9Ð9Ð9ØÔ-Ô?ÐKÐKÐKØÔ-Ô?ÐKÐKÐKð!Ô3ÔEˆŒ à Ô3ÔEˆŒ ðÔ/ÔFˆŒà(Ô5ÔCˆÔð#Ô3Ô@ˆ Ý!#¤¨<¸¼Ð*@ÍÌÐ!QÑ!QÔ!QˆÔÝ&(¤h° ÅRÄXÐ&NÑ&NÔ&NˆÔ#ð%)ˆÔ!ØÔ-ÔBˆÝ”L‘N”Nˆ à ð 0õ/2°!°iÀ1±nÑ.FÔ.FˆDÔ +ð Ð +Ô +°Ñ 7Ð +Ô +Ð +à./ˆDÔ +ð Š Ø ˆD4‰KÝ ŒHT¥¤Ð *Ñ *Ô *Ý ŒHd˜DÔ.Ð/µr´xÐ @Ñ @Ô @ñ ô ð ð ð óÚ num_requestsÚvalid_ngram_requestsÚnum_tokens_no_specÚ token_ids_cpuÚreturnc óŽ—g}t|¦«x}r³t¦«}tj|¦«}||jkr3t dt |j|¦«¦«} t| ¦«ntd¦«t||||j |j |j |j |j|j¦ « t|¦«t!|¦«D]n} | |vrS|j| dkrB| |j| d|j| …f ¦«¦«ŒY| g¦«Œo|S)a°Batch version of ngram proposer using numba for acceleration. Args: valid_ngram_requests: Set of indices of requests that need ngram proposals. num_tokens_no_spec: Numpy array of shape (batch_size,) representing the number of tokens without speculative tokens for each request. token_ids_cpu: Numpy array of shape (batch_size, max_model_len) representing the token IDs for each request. Returns: list[list[int]]: A list where each element is a list of proposed token IDs for the corresponding request. rrN)ÚlenrrÚsumr!Úmaxr&r'rÚbatch_propose_numbarrrrrr ÚrangeÚappendÚtolist) r)r.r/r0r1Údraft_token_idsÚnum_ngram_requestsÚoriginal_num_numba_threadsÚ total_tokensÚfinal_num_threadsÚis r+Ú batch_proposezNgramProposer.batch_propose?s{€ð0,.ˆõ "%Ð%9Ñ!:Ô!:Ð :Ð ð 8Ý)8Ñ):Ô):Ð &õœ6Ð"4Ñ5Ô5ˆLؘtÔ8Ò8Ð8Ý$'Ø•s˜4Ô:Ðrqs«ðð € € € àÐÐÐØ € € € ØEÐEÐEÐEÐEÐEÐEÐEÐEÐEÐEÐEÐEÐEà"Ð"Ð"Ð"Ð"Ð"ðZ ðZ ðZ ðZ ðZ ñZ ôZ ðZ ðz€ˆtÐÑÔðOØðOàœ ðOð”:ðOð ð Oð ð Oð ð Oð ðOð”zðOðœJðOðOðOñÔðOð8€ˆdÐÑÔðV>Ø”:ðV>àðV>ððV>ðð V>ð  ð V>ð „Zð V>ðV>ðV>ñÔðV>ðV>ðV>r-