HejuruUbuyobozi bukurikira
Kwangwa Gutoranya Cyiza-Guhuza
Ururimi AI
Ururimi AI
Speculative sampling uses a draft model to propose tokens and a target model to verify them.
Umva iki gitaboAI-byavuzwe · ~2 min
With the exact rejection-sampling correction, it preserves the target distribution; speedup depends on draft cost, acceptance, target verification and workload.
Autoregressive generation is slow because each token needs a full forward pass of a huge model. Speculative sampling fixes this by pairing a cheap draft model with the expensive target model. The draft proposes a short run of tokens (say 4-8); the target then scores all of them in one parallel forward pass. A modified rejection-sampling rule accepts the longest prefix that is consistent with the target's own distribution and resamples at the first rejected position. Because acceptance is probabilistic and corrected, the final token stream is provably distributed exactly as if the target had generated alone, no quality loss. Speedups vary with draft cost, acceptance rate, target verification, batching and hardware. A parallel verification pass can confirm multiple tokens, but is not free.
Gukora moderi ya 70B yo kuganira hamwe na 7B yerekana icyitegererezo kugirango ugabanye gutinda gusubiza hafi kabiri hamwe nubwiza busohoka.
Imiterere ya Medusa kumutwe umwe uhanura ibimenyetso byinshi bizaza, hanyuma ukabigenzura udafite umushinga wihariye.
Igiti gishingiye ku gutondeka ibiti bitanga amashami menshi kandi bikagenzurwa byose muburyo bumwe.
Kwihutisha abafasha-kurangiza abafasha aho umushinga wikitegererezo ukemura ibyateganijwe byavuzwe ko moderi nini yemeza vuba.
Gukoresha decoding yibintu birahinduka mubisanzwe. Impinduka nshya zireka umushinga wintangarugero: kwiyitirira gukoresha hakiri kare-gusohoka cyangwa imitwe yinyongera yo guhanura (Medusa, EAGLE), igishushanyo mbonera cyibiti kigenzura abakandida benshi icyarimwe, kandi decodehead decoding igereranya n-garama. Tegereza kwishyira hamwe hamwe nogucunga hamwe na KV-cache, gucunga ibyuma bingana, hamwe no gukoresha mugukoresha ibicuruzwa bitinda cyane nkabafasha kuganira hamwe nibikoresho bya code aho buri milisegonda ibara.
Ibaruwa y'amakuru y'ubuntu
Inkuru eshatu zizewe za AI buri gitondo cy'iminsi y'akazi, zanditse mu rurimi rworoshye. Ubuntu iteka ryose, nta kwamamaza.
Imeri imwe buri munsi y'akazi. Kwiyimura mu buryo bumwe. Ntitwigera tugurisha cyangwa ngo dusangire aderesi yawe.
Isuzume
Ibisubizo ako kanya kuri buri kibazo, n'icyemezo gishobora gusangirwa gifite ID yagenzurwa umaze gutsinda isomo.
Fasha uburezi bwa AI bw'ubuntu. AI Understanding ni umuryango utegamiye kuri leta wa 501(c)(3) — nta kwamamaza, nta kwishyura, iteka ryose. Tanga impano
Speculative sampling uses a draft model to propose tokens and a target model to verify them. With the exact rejection-sampling correction, it preserves the target distribution; speedup depends on draft cost, acceptance, target verification and workload.
Inyandiko ihendutse yerekana urugero rwerekana ibimenyetso byinshi imbere, kandi intego ihenze yerekana igenzura byose muburyo bumwe bubangikanye imbere.
Guhindura kwangwa-gutoranya gukosora byemeza ibimenyetso byemewe bitangwa neza nkuko intego yabigenewe yabyaye wenyine.
Exact speculative sampling uses a rejection-sampling correction and discards the later draft suffix.
The rejection-sampling rule preserves the target distribution; this does not guarantee the same sampled string for each random seed.
The target pass scores the draft proposals and supplies target probabilities used by the acceptance rule and residual correction; it has a computation cost.
Komeza wige
Abandi bayobozi batoranijwe kuriyi ngingo
HejuruUbuyobozi bukurikira
Kwangwa Gutoranya Cyiza-Guhuza
Ururimi AI