Wenxin 5.5 Flash Base Uyasebenza

I-Wenxin 5.5 Flash Base iyimodeli yakamuva eyisisekelo ekhishwe yi-Baidu njengomthombo ovulekile. Isebenzisa ukwakheka kwe-sparse MoE, inamaphamethara ayizigidi eziyizinkulungwane ezingama-280 esewonke, ivuse amaphamethara ayizigidi eziyizinkulungwane eziyi-14 kuphela ngethokheni ngayinye, yaqeqeshwa kusengaphambili ngamathokheni ayizigidigidi ezingama-33.5, futhi isekela umongo wamathokheni ayisigidi esi-1. I-Wenxin 5.5 Flash Base igxile ekusunguleni ukusebenza kahle: i-sparse attention yehlisa izindleko zemibhalo emide, ama-router amaningi e-MoE athuthukisa umzila, kanti i-ECHO iphinda isebenzise ungqimba lokugcina lomgogodla ukubikezela amathokheni alandelayo. Ngokusho kwe-Baidu, ithola i-SOTA phakathi kwamamodeli anosayizi ofanayo kuma-benchmark amaningi, futhi isiyatholakala ku-HuggingFace.

Isisekelo sombhalo esisebenza kahle kakhulu: i-sparse MoE enamaphamethara ayizigidi eziyizinkulungwane ezingama-280 esewonke, lapho kuvuswa ayizigidi eziyizinkulungwane eziyi-14 kuphela ngethokheni ngayinye, inikeza amakhono ombhalo ajwayelekile asezingeni le-SOTA phakathi kwamamodeli anosayizi ofanayo.
Umongo omude kakhulu: umongo wokuqeqeshwa kwangaphambili ungamathokheni angu-128K, kanti ukulungisa kahle nokucabanga kusekela kufika kumathokheni ayisigidi esi-1.
Ukumbozwa kwamakhono okubanzi: ifinyelela ezingeni eliphezulu phakathi kwamamodeli anosayizi ofanayo kuma-benchmark olwazi, ukucabanga, izibalo, ikhodi, izilimi eziningi nomongo omude.
Umthombo ovulekile ongenziwa ngokwezifiso: izisindo ze-FP8 ne-BF16 ziyatholakala ku-HuggingFace, zinika onjiniyela namabhizinisi ithuba lokulungisa kahle nokwakha amamodeli ochwepheshe.
Isimiso sokucabanga esiphelele: i-Baidu inikeza i-vLLM plugin, i-SGLang Cookbook nezibonelo zokulungisa kahle ze-Megatron-Bridge.

Landela i-WeChat bese uphendula ngokuthi “umthombo ovulekile” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI omthombo ovulekile
Latent-Gated Sparse Attention: ihlanganisa i-heavy-compression attention (HCA) ne-multi-head latent attention (MLA) ibe ukwakheka okunokwabiwa kwemisebenzi. Izingqimba ezingama-37 zicutshungula ulwazi lomlando ochungechungeni olucindezelwe izikhathi ezingu-128, kanti izingqimba eziyisi-6 kuphela zibhekele ukubuyisa imininingwane emincane, ngaleyo ndlela kwehliswe kakhulu izindleko ze-KV Cache kumongo omude; i-low-rank gated attention iphinda isebenzise i-low-rank query latent ekhiqizwa ungqimba lwe-attention ngokwalo ukubala isango lokuphumayo, yehlise inani lamapharametha e-gated projection ngo-75% futhi ithuthukise imiphumela; i-virtual-head attention (VHA) ithuthukisa amandla okuveza ngesikhathi sokuqeqeshwa, bese ihlanganiswa ne-attention operator ngesikhathi sokucabanga, ngezindleko ezengeziwe ezincane kakhulu.
Multi-Router Latent MoE: ama-router amaningi azimele abamba iqhaza ngokuhlanganyela ekukhethweni kochwepheshe, athuthukise ukuhlobana phakathi kwamaphuzu omzila nokusetshenziswa kwangempela kochwepheshe, anciphise umthelela wokuchema kwamaphuzu owodwa ekukhethweni komzila, kuyilapho kugcinwa kungashintshiwe inani lochwepheshe abavuswayo, umthamo wokubala kochwepheshe kanye nenani lokuxhumana.
Inhloso yokuqeqeshwa ye-ECHO: ayengezi enye imojuli ye-MTP, kodwa iphinda isebenzise ungqimba lokugcina lomgogodla ukuze imodeli ifunde ngokwayo ukubikezela amathokheni alandelayo, ihlele kusenesikhathi indlela yokulandelana futhi ithuthukise amakhono okucabanga nawokubhala ikhodi; ungqimba oluphinde lusetshenziswe lungasebenza ngokuqondile njengemodeli esalungisa ukuqagela, futhi ukuqeqeshwa ngesinyathelo esisodwa kungabonisa ikhono lokwandisa izinyathelo eziningi, ikhwalithi yokusalungiswa ibe ngcono kakhulu kunemojuli esizayo ye-MTP.
Ukuthuthukiswa kohlelo lokuqeqesha: i-distributed Muon optimizer ihlanganisa ukuxhumana kwamapharametha e-Dense nawe-MoE phakathi kwamakhompyutha kube ukuxhumana okuqoqiwe okukodwa, kuthi i-FP8 mixed precision ithuthukise ukusebenza kahle kokuqeqeshwa okukhulu; i-FlashMask-4 attention operator eyakhiwe ngaphakathi isekelwe ku-CuTe-DSL, isekela izakhiwo ze-SM90/SM100 kanye ne-head dim efinyelela ku-576, futhi ivumelana ngokuqondile nokubala kwe-wide-head kongqimba lwe-HCA.

Lungisa indawo yehardware: imodeli inamaphamethara ayizigidi eziyizinkulungwane ezingama-280, futhi ngisho nezisindo ze-FP8 quantization zicishe zibe ngu-280GB. Kudingeka iseva enama-GPU amaningi asezingeni eliphezulu, njenge-H-series enezakhiwo ze-SM90/SM100; ikhadi elilodwa alanele.
Landa izisindo zemodeli: donsa indawo yokugcina esemthethweni ku-HuggingFace, ukhethe i-FP8 (baidu/ERNIE-5.5-Flash-Base) noma i-BF16 ngokwesidingo, isibonelo somyalo sithi huggingface-cli download baidu/ERNIE-5.5-Flash-Base --local-dir ./ernie-base.
Faka uhlaka lokucabanga: faka i-vLLM plugin esemthethweni ye-Baidu (baidu/vLLM-ERNIE) ukuze ivumelane nokwakheka kwayo kwe-sparse attention, noma usebenzise i-SGLang Cookbook esemthethweni.
Qalisa isevisi yokucabanga: sebenzisa i-vllm serve ukhombise indlela yezisindo zasendaweni, setha --tensor-parallel-size ngokuya ngenani lamakhadi bese uvula --enable-expert-parallel. Ngemva kokuqalisa kwesevisi, kuzovela i-interface ehambisana ne-OpenAI.
Qalisa ucingo: thumela izicelo embobeni yasendaweni usebenzisa ifomethi evamile ye-OpenAI API, futhi ulungise umongo omude ofinyelela kumathokheni ayisigidi esi-1 ngokulungisa --max-model-len kuye ngememori ye-GPU.
Yenza ukulungisa ngokwezifiso: sebenzisa isibonelo esisemthethweni se-Megatron-Bridge ukuze uqhubeke nokuqeqeshwa kwangaphambili kwedomeyini noma wenze i-SFT.

Ukusebenza kahle kokucabanga okwedlulele: ukwakheka kwe-sparse MoE okunamaphamethara ayizigidi eziyizinkulungwane ezingama-280 kuvusa amaphamethara ayizigidi eziyizinkulungwane eziyi-14 kuphela ngethokheni ngayinye, kusekela umthamo wemodeli enkulu kakhulu ngezindleko zokubala eziphansi kakhulu.
Ikhono lomongo omude kakhulu: umongo wokuqeqeshwa kwangaphambili ungamathokheni angu-128K, kanti ukulungisa kahle nokucabanga kusekela kufika kumathokheni ayisigidi esi-1, kunciphise kakhulu izindleko zezinsiza zokucubungula imibhalo emide.
Ukusebenza okuhamba phambili ngosayizi ofanayo: ukuhlolwa okusemthethweni kukhombisa ukuthi ithole i-SOTA phakathi kwamamodeli anosayizi ofanayo kuma-benchmark angu-30 olwazi, ukucabanga, izibalo, ikhodi, izilimi eziningi nomongo omude.
Ukusungula ukusebenza kahle ngendlela ehlelekile: i-Latent-Gated sparse attention, i-Multi-Router MoE kanye nenhloso yokuqeqeshwa ye-ECHO kuthuthukisa ngokubambisana izigaba ezintathu ze-attention, umzila nenhloso yokuqeqesha.
Ingqalasizinda yobunjiniyela ephelele: i-distributed Muon optimizer, i-FP8 mixed precision kanye ne-FlashMask-4 operator eyakhiwe ngaphakathi kuqinisekisa ukusebenza kahle kokuqeqesha nokucabanga ngezinga elikhulu.

Iwebhusayithi esemthethweni yephrojekthi: https://ernie.baidu.com/zh/release/ernie-5-5-flash-base/
Indawo yamamodeli ye-HuggingFace: https://huggingface.co/baidu/ERNIE-5.5-Flash-Base

Ukucubungula okujulile kwemibhalo emide: ngenxa yomongo omude ofinyelela kumathokheni ayisigidi esi-1, ingacubungula ngesikhathi esisodwa inkontileka yezomthetho ephelele, umbiko wezezimali, iqoqo lamaphepha noma indawo yokugcina ikhodi, yenze isifinyezo esiphelele, imibuzo nezimpendulo eziwela izahluko kanye nokukhipha ulwazi.
Usizo ekuthuthukisweni kwekhodi: ngenxa yamakhono ayo avelele ekhodi nomongo omude kakhulu, ingaqonda isakhiwo sendawo yokugcina yonke ukuze yenze ikhodi, iphakamise ukwakhiwa kabusha futhi ihlaziye amaphutha awela amafayela.
Ukwenza amamodeli ezindima athile: njengesisekelo somthombo ovulekile esingakaze sidlule ekuqeqeshweni kwangemva kwalokho, amabhizinisi angasilungisa kahle esebenzisa idatha yemboni yawo ukuze akhe amamodeli ochwepheshe bezomthetho, ezezimali, ezokwelapha nezinye izindawo.
Ukucubungula okuqukethwe ngezilimi eziningi: amakhono ezilimi eziningi afinyelela ku-SOTA phakathi kwamamodeli anosayizi ofanayo, ngakho afanele ukuhumusha imibhalo yezilimi ezahlukene, ukubuyekeza okuqukethwe kwamazwe ngamazwe kanye nemibuzo nezimpendulo ezisekelwe ezinqolobaneni zolwazi zomhlaba wonke.