Ndaju nke ekewapụrụ na Decode ije ozi
Ihe owuwu na-ekesa nnukwu ihe nlere asụsụ n'ime akụkụ abụọ dị iche iche — prefill na decode — ma na-agba ha na ọdọ mmiri GPU dị iche iche.
Nchịkọta
It matters because these two phases have opposite hardware appetites, and forcing them onto the same machines wastes capacity and hurts latency.
Ime miri emi
Mgbe LLM zara, ọ na-arụ ọrụ na nkebi abụọ. Prefill na-agụ ozugbo ozugbo wee wuo oghere igodo-uru (KV); nke a bụ nnukwu myirịta, gbawara agbakọ nke na-eju nkeji mgbakọ na mwepụ GPU. Decode wee wepụta akara n'otu n'otu, usoro nke ọ bụla na-agụ cache KV dum — ihe na-ejikọta bandwidth ebe nchekwa, na-agbakọ ngwa ngwa. Na-agba ọsọ ọnụ, ogologo prefill na-egbochi onye ọ bụla decode (mchichi isi nke ahịrị), na ịchịkọta ha abụọ na-emepụta nnyonye anya. Nkewa na-etinye prefill n'otu ọdọ mmiri GPU wee depụta koodu na nke ọzọ, na-ebufe cache KV n'etiti ha na njikọ ngwa ngwa dị ka NVLink ma ọ bụ InfiniBand. A na-edobe ọdọ mmiri ọ bụla ma tụkwasa ya n'onwe ya, na-eme ka ihe dị mma na-eme ka ọ dị mma, na-eme ka nkwụsị ọdụ dị nro, na ikwe ka ndị na-arụ ọrụ kụrie ihe nleba anya oge-na-mbụ-mbụ na oge-otu mmepụta-token n'otu oge.
Nghọta nka nka
Usoro abụọ a dị iche na njedebe ha. Na-emeju ihe nrịbama niile n'otu ntabi anya, ya mere ọ̀tụ̀tụ̀ FLOP ya nwere ogologo ọsọ ọsọ na ọ na-ebupụ cores tensor. Decode bụ autoregressive: ọhụrụ token ọ bụla chọrọ otu ngafe na-aga n'ihu nke na-agụgharị cache KV zuru ezu sitere na HBM, yabụ, a na-eji bandwidth ebe nchekwa na-enweta mmepụta, ọ bụghị ịgbakọ. Nkewa nkewa na-erigbu nke a site n'ịha nha, nhazi, na ọbụna họrọ myirịta dị iche iche maka ọdọ mmiri ọ bụla, wee bupu cache KV site na ndị ọrụ prefill iji depụta ndị ọrụ.
Mmetụta atụmatụ
Ọnụ ego na mmefu ego
Mkpebi ihe owuwu ụlọ na-akwalite arụmọrụ yana ọnụ ahịa ọrụ ruo ọtụtụ afọ.
Mkpebi doro anya
Nkà mmụta nka na-enyere ndị otu egwuregwu aka ịhọrọ nchịkọta ziri ezi, ọ bụghị naanị nke kachasị ọhụrụ.
Quality akara
Nhọrọ injinia ka mma na-ebelata ihe omume ntụkwasị obi na mmepụta.
Ọdịnihu nke njuju ekewapụrụ yana dekọọ ozi
Na-atụ anya ikewasa ka ọ bụrụ ndabara na nchịkọta mmepụta. Sistemu dị ka DistServe, Splitwise, na Mooncake mere ka ọ gbasaa, yana vLLM na NVIDIA Dynamo na-ebuga ụdịdị ekewapụrụ ugbu a. Nchọpụta na-akwalite nkwalite mbufe cache nke KV-cache, ịkwakọba cache na ijigharị ihe n'ofe arịrịọ, nhazigharị ike nke prefill/decode ratios n'okpuru okporo ụzọ na-agbanwe, yana njikọta siri ike na caching prefix na prefill chunked. Ka windo ndị gbara ya gburugburu na-etolite n'ime nde token, ikewa usoro ndị a na-aghọwanye ihe dị mkpa maka ozi dị ọnụ ala, dị ala.
Mmejuputa n'ezie n'ụwa
Onye enyemaka nkata na-aga akwụkwọ dị ogologo na-akpalite ụyọkọ prefill dị arọ, wee zaghachi iyi sitere na ụyọkọ decode kachasị mma nke ebe nchekwa ka ọ nọgide na-ede ederede nke ọma.
NVIDIA Dynamo na vLLM na-ahapụ ndị na-arụ ọrụ ka ha wepụta prefill dị iche iche na decode otu ndị ọrụ ka mwapu nke ogologo mkpali ghara ifriizi ọgbọ na-aga n'ihu.
Mooncake (nke Moonshot AI's Kimi ji mee ihe) na-ekewapụta prefill na decode ma gbakwụnye ọdọ mmiri KV-cache ekesara iji belata ngụkọ ngwa ngwa n'ọtụtụ.
Ọrụ mmecha koodu na-arara obere ọdọ mmiri prefill maka mkpali dị mkpirikpi yana nnukwu ọdọ mmiri decode, ebe ọ bụ na ọtụtụ ọnụ ahịa na-abịa site na ịkwanye ọtụtụ akara ngosi mmepụta.
Ihe ize ndụ & okporo ụzọ nche
Ịkwalite otu akara ngosi nwere ike zoo adịghị ike sistemụ sara mbara.
A na-eledakarị ihe akụrụngwa na ụgwọ ọrụ anya.
Ọdịiche nchekwa na nleba anya nwere ike itolite ka sistemu na-adịwanye mgbagwoju anya.
Map mmejuputa
Kọwaa latency, ịdịmma na ebumnuche ọnụ ahịa tupu mmejuputa ya.
Benchmark n'okpuru ibu dị adị na ọnọdụ data.
Nleba anya akụrụngwa maka mperi, ịkpafu na mmetụta onye ọrụ.
Kwadebe ụzọ nzaghachi azụghachi azụ na ihe omume tupu ịchachaa.
Nọgide na-eme nchọpụta
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Disaggregated Prefill and Decode Serving quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Ntuziaka na-esote
KServe na Model Na-eje ozi na Kubernetes
Ajụjụ a na-ajụkarị
What is Disaggregated Prefill and Decode Serving?
Ihe owuwu na-ekesa nnukwu ihe nlere asụsụ n'ime akụkụ abụọ dị iche iche — prefill na decode — ma na-agba ha na ọdọ mmiri GPU dị iche iche. Ọ dị mkpa n'ihi na usoro abụọ a nwere agụụ ngwaike dị iche, na ịmanye ha n'otu igwe na-emebi ikike ma na-emerụ ahụ ike.
Kedu ihe bụ isi ngwaike kpatara ikewapụta prefill na decode na ọdọ mmiri GPU dị iche iche?
Prefill na-ahazi ngwa ngwa n'otu n'otu yana saturates na-agbakọ, ebe decode na-agụ cache KV nke ọ bụla ma na-ejedebe site na bandwit ebe nchekwa - agụụ na-emegide nke na-akwado ọdọ mmiri ndị a na-emegharị onwe ha.
Kedu usoro data a ga-ebufe site na ndị ọrụ prefill iji depụta ndị ọrụ?
Prefill na-ewuli oghere KV maka ngwa ngwa; decode chọrọ cache ahụ ka ọ gaa n'ihu na-emepụta, yabụ na-ebufe cache ahụ n'otu njikọ ngwa ngwa na ọdọ mmiri decode.
Kedu nsogbu nkewasa na-ebelata kpọmkwem na nhazi GPU nkekọrịta?
Na GPU ndị na-ekerịta ogologo njuju njuju nwere ike igbochi usoro ngbanwe na-aga n'ihu; ikewapụ ha na-egbochi nnyonye anya ahụ ma na-edozi ọdụ ọdụ.
Kedu ihe kpatara enwere ike ijikọ ihe njuju nke ọma mana dekọọ uru site na nlegharị anya dị iche iche?
Mejupụta usoro ihe niile na-akpalite ọnụ, yabụ nnukwu batches na-eri nri nke ọma; decode na-emepụta otu token n'otu oge ma na-echekwa ya site na ebe nchekwa, ya mere ọ na-arị elu dị iche iche.
Kedu njikọ njikọ a na-ejikarị ibugharị cache KV n'etiti ọdọ mmiri ekewapụrụ?
Njikọ bandwit dị elu, njikọ dị ala dị ka NVLink (intra-node) na InfiniBand (inter-node) chọrọ ka mbufe KV-cache ghara ịghọ ihe ọhụrụ.