I-LLaMA Factory: Ukulungisa kahle amamodeli amakhulu angu-100+ ngaphandle kwekhodi

I-LLaMA Factory (manje ebizwa ngokuthi i-LlamaFactory) iwuhlaka oludidiyelwe lokuqeqesha kabusha amamodeli amakhulu olimi. Le phrojekthi yaqalwa ithimba likaDkt. Zheng Yaowei wase-Beihang University, futhi isekelwa ngokwezifundo yiphepha le-ACL 2024. Uhlaka lusebenzisa umklamo othi “CLI engenakhodi + Web UI”, ngenhloso yokwenza inqubo yokuqeqesha kabusha ibe lula. I-LlamaFactory isiqongelele ama-Stars angaphezu kuka-72 000 ku-GitHub, futhi isiphenduke ireferensi emkhakheni wokuqeqesha kabusha amamodeli amakhulu.

>

I. Uhlolojikelele lwephrojekthi

• Isigaba｜Imininingwane

• Igama eligcwele lephrojekthi｜LlamaFactory (eyayibizwa nge-LLaMA-Factory)

• Ikheli le-GitHub｜github.com/hiyouga/LlamaFactory

• Stars｜72,476 ⭐

• Forks｜8,861

• Ulimi oluyinhloko｜Python (99.7%)

• Ilayisensi｜Apache-2.0

• Isikhathi sokudalwa｜2023-03-21

• Inguqulo yakamuva｜v0.9.5 (2026-05-30)

• Inani eliphelele lama-commit｜3,042 commits

• Umnikeli oyinhloko｜hiyouga (Zheng Yaowei, 2,188 commits, u-72% wawo wonke ama-commit)

• Iphepha lezemfundo｜ACL 2024 System Demonstrations

• Usayizi weriphozithri｜~412.9 MB

• Izinkinga ezivulekile｜57

Umsunguli uZheng Yaowei (hiyouga) ungumfundi we-PhD e-Beihang University. Njengombhali wokuqala, ushicilele amaphepha amaningi ezingqungqutheleni ezihamba phambili emhlabeni ezifana ne-ACL, CVPR kanye ne-WWW, futhi uke waba umbuyekezi wezingqungquthela ezifana ne-AAAI ne-EMNLP. Imiklomelo ayitholile ihlanganisa i-Huawei Cup, umklomelo wokuqala kazwelonke we-Lanqiao Cup, i-Beihang University Student of the Year, kanye ne-Ascend Open Source Excellence Contribution Award. Amalungu ethimba asebenzisana naye ahlanganisa oprofesa base-Beihang uZhang Richong, uZhang Junhao, uYe Yanhan, uLuo Zheyan, uFeng Zhangchi, uMa Yongqiang nabanye.

II. Indawo yephrojekthi nenani layo eliyinhloko

2.1 Incazelo ngomusho owodwa

Uchungechunge olujwayelekile lokuqeqesha kabusha amamodeli amakhulu: inqubo ehlanganisiwe esebenza kunoma iyiphi imodeli, ehlanganisa ukulungiswa kwedatha, ukuqeqeshwa kwemodeli, ukuqinisekiswa kokucabanga kanye nokuthunyelwa, kusukela ekuqaleni kuya ekugcineni.

2.2 Izinkinga eziyinhloko elizixazululayo

• Inkinga｜Isixazululo se-LLaMA Factory

• Izinqubo zokuqeqesha kabusha ziyahlukahluka kumodeli ngamunye, okwenza ukushintsha kube nezindleko eziphezulu｜Umklamo “owahlukanisa” i-algorithm, imodeli nedatha, onendawo eyodwa yokufinyelela amamodeli angaphezu kuka-100

• Ikhodi yokuqeqesha kabusha inzima ukuyisebenzisa futhi idinga ulwazi olujulile lokufunda komshini｜CLI engenakhodi + i-LLaMA Board Web UI (Gradio)

• Imemori ye-GPU ayanele, ama-GPU asezingeni labathengi awakwazi ukuqeqesha amamodeli amakhulu｜QLoRA 4-bit: i-7B idinga u-6GB kuphela, i-70B idinga u-48GB kuphela

• Umphakathi wemithombo evulekile awunalo uchungechunge lwamathuluzi oluhlanganisa konke endaweni eyodwa｜Ihlanganisa wonke umjikelezo wempilo kusukela ku-PT→SFT→RM→PPO→DPO

• Ukuthunyelwa kokucabanga kuyinkimbinkimbi｜Ukuthekelisa ngokuchofoza okukodwa + i-API yesitayela se-OpenAI + vLLM/SGLang

2.3 Ifilosofi yomklamo: “ukwahlukanisa” i-algorithm, imodeli nedatha

Engxoxweni eyodwa, uZheng Yaowei wafanisa inqubo yokucabanga kwemodeli nezimoto ezihamba emgwaqweni omkhulu ojwayelekile. Imodeli ngokwayo imele izinhlobo ezahlukene zezimoto, kanti idatha imele impahla ethwelwe yilezo zimoto. Uma nje imodeli ikwazi ukusebenza kahle kulo mgwaqo omkhulu, inqubo yokuqeqesha kabusha ingaqedwa ngokulungisa amapharamitha kuphela.

III. Ukuhlaziywa okujulile kwesakhiwo sobuchwepheshe

3.1 Umklamo osebenzisa izakhiwo ezimbili ngesikhathi esisodwa

I-LLaMA Factory inikeza ngesikhathi esisodwa izakhiwo ezimbili: i-v0 (inguqulo ezinzile) ne-v1 (inguqulo yokuhlola). Kungashintshwa phakathi kwazo ngokusetha i-variable yemvelo ethi USE_V1.

v0 (okuzenzakalelayo): api/webui → chat/eval/train → data/model → hparams → extras
v1 (yokuhlola): trainers → core → accelerator/plugins/config → utils

Isakhiwo se-v0 siyindlela esetshenziswa kakhulu njengamanje, sinikeza ukusebenza okuphelele kanye nemibhalo egcwele. Isakhiwo se-v1 sisasesigabeni sokuhlola futhi sisebenzisa umklamo ohlukaniswe waba amazinga amaningi.

3.2 Incazelo eningiliziwe yamamojuli ayinhloko e-v0

src/llamafactory/
├── cli.py              # Indawo yokungena ye-CLI, isabalalisa imiyalo ye-llamafactory-cli
├── launcher.py          # Isiqalisi, sisabalalisa ngokuya nge-USE_V1
├── api/                 # REST API yesitayela se-OpenAI
│   ├── app.py           # Uhlelo lwe-FastAPI
│   ├── chat.py          # Iphoyinti lokuxhumana lengxoxo
│   └── protocol.py      # Incazelo yephrothokholi
├── chat/                # Isendlalelo senjini yokucabanga
│   ├── base_engine.py   # Isigaba esiyisisekelo esingabonakali
│   ├── hf_engine.py     # Ukucabanga komdabu kwe-HuggingFace
│   ├── vllm_engine.py   # Ukucabanga kokusebenza okuphezulu kwe-vLLM
│   └── sglang_engine.py # Ukucabanga kwe-SGLang
├── data/                # Isendlalelo sokucubungula idatha
│   ├── template.py      # Ukubhaliswa kwezifanekiso ze-Prompt zamamodeli angaphezu kuka-100
│   ├── mm_plugin.py     # Ukucubungula kwemidiya eminingi (izithombe/ividiyo/umsindo)
│   ├── loader.py        # Isilayishi sedathasethi
│   ├── processor/       # Amaphrosesa ngokwesigaba sokuqeqesha
│   └── converter.py     # Ukuguqulwa kwefomethi (Alpaca/ShareGPT)
├── model/               # Isendlalelo semodeli
│   ├── loader.py        # Ukulayishwa kwemodeli ne-tokenizer, quantization, LoRA
│   └── patcher.py       # Amapheshana okuhambisana kwemodeli
├── train/               # Isendlalelo sokuqeqesha
│   ├── sft/             # Ukuqeqesha kabusha okugadiwe
│   ├── dpo/             # Ukuthuthukisa okuncamelayo kwe-DPO
│   ├── ppo/             # Ukufunda okuqinisa kwe-PPO
│   ├── rm/              # Ukwakha imodeli yomvuzo
│   ├── pt/              # Ukuqeqeshwa kwangaphambili
│   ├── kto/             # Ukuqeqeshwa kwe-KTO
│   └── mca/             # Ukuhambisana ne-Megatron-Core
├── hparams/              # Isendlalelo sokumiswa kwamapharamitha
│   ├── model_args.py    # Amapharamitha emodeli
│   ├── data_args.py     # Amapharamitha edatha
│   ├── finetuning_args.py # Amapharamitha okuqeqesha kabusha
│   └── training_args.py # Amapharamitha okuqeqesha (azuza njengefa ku-HF TrainingArguments)
├── extras/               # Iphakheji yamathuluzi
│   ├── constants.py     # Ama-constant/enumeration omhlaba wonke
│   └── packages.py      # Ukuhlolwa kokuncika
└── webui/                # Isixhumi esibonakalayo sewebhu se-LLaMA Board
    ├── interface.py     # Ukwakhiwa kwesixhumi esibonakalayo se-Gradio
    ├── runner.py        # Ukuqalisa imisebenzi yokuqeqesha
    └── components/      # Izingxenye ze-UI

3.3 Inqubo yokuqeqesha (v0 Training Flow)

llamafactory-cli train xxx.yaml
  → run_exp() [tuner.py]
    → read_args() → Hlaziya ukumiswa kwe-YAML/JSON
    → get_train_args() → Khiqiza ama-dataclass angu-4 anohlobo:
        ModelArguments + DataArguments + FinetuningArguments + TrainingArguments
    → Dlulisela emsebenzini othile wokuqeqesha:
        run_sft / run_dpo / run_ppo / run_rm / run_pt / run_kto
    → (ngokuzikhethela) export_model() → Hlanganisa izisindo ze-LoRA bese uthekelisa

3.4 I-matrix yezinjini zokucabanga

• Injini｜Izici｜Izimo ezifanele zokusetshenziswa

• HF Engine｜I-HuggingFace yomdabu, inokuhambisana okuhle kakhulu｜Ukuthuthukisa nokulungisa, kanye nokucabanga ngezinga elincane

• vLLM Engine｜Ukusheshisa ukucabanga ngo-270%, i-PagedAttention｜Ukucabanga okuku-inthanethi okusebenza kahle kakhulu

• SGLang Engine｜I-RadixAttention, ukukhiqiza okuhlelekile｜Ukucabanga kwenqwaba nokukhiqiza okuyinkimbinkimbi

• KT Engine｜Eyenzelwe ngokukhethekile ukudluliswa kolwazi｜Ukufingqa imodeli

IV. Uhlelo lwemisebenzi

4.1 Ama-paradigm okuqeqesha asekelwayo (8 × 6 izindlela zokuqeqesha kabusha = 48 izinhlanganisela)

• I-paradigm yokuqeqesha｜Full｜Freeze｜LoRA｜QLoRA｜OFT｜QOFT

• Ukuqeqeshwa kwangaphambili (PT)｜✅｜✅｜✅｜✅｜✅｜✅

• Ukuqeqesha kabusha okugadiwe (SFT)｜✅｜✅｜✅｜✅｜✅｜✅

• Ukwakha imodeli yomvuzo (RM)｜✅｜✅｜✅｜✅｜✅｜✅

• Ukuqeqeshwa kwe-PPO｜✅｜✅｜✅｜✅｜✅｜✅

• Ukuqeqeshwa kwe-DPO｜✅｜✅｜✅｜✅｜✅｜✅

• Ukuqeqeshwa kwe-KTO｜✅｜✅｜✅｜✅｜✅｜✅

• Ukuqeqeshwa kwe-ORPO｜✅｜✅｜✅｜✅｜✅｜✅

• Ukuqeqeshwa kwe-SimPO｜✅｜✅｜✅｜✅｜✅｜✅

4.2 Uchungechunge lwamamodeli olusekelwayo (uchungechunge olungaphezu kuka-60, izinhlobo ezingaphezu kuka-100)

Amamodeli akhethiwe amelele:

• Uchungechunge lwemodeli｜Usayizi wamapharamitha｜Izici

• Llama 3-4｜1B–402B｜Imodeli ehamba phambili ye-Meta yomthombo ovulekile

• Qwen 2-3.6｜0.5B–397B｜Uchungechunge oluphelele lwe-Tongyi Qianwen lwe-Alibaba

• DeepSeek R1｜1.5B–671B｜Inguqulo ecindezelwe yemodeli yokucabanga

• Gemma 3｜270M–27B｜Imodeli encane ye-Google

• InternVL 2.5-3.5｜1B–241B｜Imodeli yolimi nombono enamamodi amaningi

• MiniCPM｜0.5B–9B｜Ilungiselelwe ukufakwa kumadivayisi asemaphethelweni

• GLM-4/4.5｜9B–355B｜Uchungechunge lwe-ChatGLM lwe-Zhipu

• Mistral/Mixtral｜7B–8x22B｜Inhlanganisela yochwepheshe be-MoE

• Phi-3/4｜3.8B–14B｜Imodeli encane esebenza kahle ye-Microsoft

• Qwen-VL 2-3｜2B–235B｜Imodeli yolimi nombono

4.3 Ukusekelwa kwama-algorithm athuthukile

• Algorithm｜Umsebenzi｜Inguqulo yokwethulwa

• FlashAttention-2｜Ukusheshisa ukubalwa kwe-attention｜2023-09

• Unsloth｜Ukusheshisa ukuqeqeshwa ngo-170% + ukonga imemori ye-GPU ngo-50%｜2023-12

• DoRA｜Ukuhlukaniswa kwezisindo kwe-LoRA｜2024-02

• GaLore｜Ukulungiselelwa kwe-gradient low-rank｜2024-03

• BAdam｜Ukulungiselelwa kokwehla kwezixhumanisi zamabhulokhi｜2024-04

• APOLLO｜I-optimizer ethuthukile｜2025-Q1

• Liger Kernel｜Ukusheshisa kwe-Triton kernel｜2025

• Ukuqeqeshwa kwe-FP8｜Ukusekelwa kwe-Transformer Engine｜2025

4.4 I-LLaMA Board Web UI

• Isixhumi esibonakalayo esibonakalayo esingadingi ukubhala ikhodi, esakhelwe ku-Gradio

• Ukukhetha imodeli → Ukulungisa idathasethi → Amapharamitha e-fine-tuning → Qalisa ukuqeqeshwa ngokuchofoza okukodwa

• Ukuqapha kwesikhathi sangempela: ijika le-Loss, ukusetshenziswa kwe-GPU, inqubekelaphambili yokuqeqeshwa

• Usekela ama-backend amaningi okungena kwamalogi: TensorBoard, Wandb, MLflow, SwanLab

• Isixhumi esakhelwe ngaphakathi se-Chat sokwenza inference, ukuze kuhlolwe ngokushesha ngemva kokuqeqeshwa

4.5 Izindlela zokuthumela

# I-CLI fine-tuning (umyalo owodwa)
llamafactory-cli train examples/train_lora/qwen3_lora_sft.yaml

# I-Web UI
llamafactory-cli webui

# Ukuthunyelwa kwe-API ngesitayela se-OpenAI
API_PORT=8000 llamafactory-cli api examples/inference/qwen3.yaml

# Docker
docker run -it --rm --gpus=all hiyouga/llamafactory:latest

Okwesihlanu、Izidingo zehadiwe (isilinganiso sememori ye-GPU)

• Indlela｜Ukunemba｜7B｜14B｜30B｜70B

• Full (bf16)｜32-bit｜120GB｜240GB｜600GB｜1200GB

• Full (pure_bf16)｜16-bit｜60GB｜120GB｜300GB｜600GB

• Freeze/LoRA｜16-bit｜16GB｜32GB｜64GB｜160GB

• QLoRA｜8-bit｜10GB｜20GB｜40GB｜80GB

• QLoRA｜4-bit｜6GB｜12GB｜24GB｜48GB

• QLoRA｜2-bit｜4GB｜8GB｜16GB｜24GB

I-QLoRA 4‑bit quantization ingakwazi ukwenza i-fine-tuning yemodeli enamapharamitha angu-14B ku-RTX 3060 eyodwa (12 GB yememori ye-GPU), kanye nemodeli enamapharamitha angu-30B ku-RTX 4090 eyodwa (24 GB yememori ye-GPU).

Okwesithupha、Ukuthuthuka kwezinguqulo

• Inguqulo｜Usuku｜Izinguquko ezibalulekile

• v0.1.0｜2023-03｜Ukukhishwa kokuqala, kusekela i-fine-tuning ye-LLaMA

• Ukuphindaphindwa kwasekuqaleni｜2023｜SFT + LoRA/QLoRA、DPO、Web UI、FlashAttention-2、Unsloth

• Ukuphindaphindwa kwasekuqaleni｜2024 Q1-Q3｜DoRA、GaLore、BAdam、FSDP+QLoRA、vLLM、LLaVA enamamodi amaningi、ukushicilelwa kwephepha

• v0.9.1｜2025 Q1｜Llama 3.2 Vision、LLaVA-NeXT、Pixtral、Video-LLaVA

• v0.9.2｜2025 Q1｜MiniCPM-o、SwanLab、i-optimizer ye-APOLLO

• v0.9.3｜2025 Q2｜Llama 4、Gemma 3、Qwen3、InternVL3、Qwen2.5-Omni

• v0.9.4｜2025-12｜🔄 Igama le-repository lashintshwa laba yi-LlamaFactory, kwahoxiswa ukusekelwa kwe-Python 3.9-3.10

• v0.9.5｜2026-05-30｜Qwen3.5/3.6、Gemma 4、ukuhambisana ne-Transformers v5

• v1.x (yokuhlola)｜2026｜Izakhiwo ezintsha trainers → core → accelerator, ukubhalwa kabusha kwamamojula

Ukuhlaziywa kwezinyathelo ezibalulekile

1. Ukushicilelwa kwephepha le-ACL 2024 —— Ukuqinisekiswa kwezemfundo kwenza ama-Stars akhuphuke kusuka ku-10,000 kuya ngaphezu kuka-30,000

2. Ingxenye yesibili ka-2024 —— Ukusekelwa kwamamodi amaningi (LLaVA/Video/Audio) kwavula imakethe entsha

3. Unyaka ka-2025 —— Ukugcina ukukhishwa kwanyanga zonke, kusekelwa i-Qwen3/DeepSeek/Llama4, kwaba ukukhetha kokuqala komphakathi womthombo ovulekile wase-China

4. Ukushintshwa kwegama le-repository —— LLaMA-Factory → LlamaFactory, kuncishiswa ukuncika kumkhiqizo we-Meta

5. Isakhiwo sokuhlola se-v1 —— Kusikisela ukwakhiwa kabusha okukhulu okuzayo

Okwesikhombisa、I-ecosystem yomphakathi

7.1 I-matrix ye-ecosystem yamaphrojekthi

• Iphrojekthi｜Incazelo｜Stars

• LlamaFactory｜Uhlaka oluyinhloko lwe-fine-tuning｜72K+

• EasyR1｜Ukufunda okuqinisa kwe-VLM (GRPO/DAPO/Reinforce++)｜Esafufusa

• FastEdit｜Ukuhlelwa okuphumelelayo kwamaqiniso emodeli｜Kufakwe kungobo yomlando

• llamafactory-community｜Isikhungo sokuxhumana somphakathi｜-

• Easy Dataset｜Ithuluzi lokukhiqiza idatha yokwenziwa｜Owenkampani yangaphandle

• DataFlow｜Ipayipi lokucubungula idatha｜Owenkampani yangaphandle

7.2 Ukuthandwa emphakathini

• Umphakathi we-AI wase-China: I-CSDN, Zhihu ne-Blog Garden zinezifundo namakhulu ezihloko zokwabelana ngolwazi

• Ukwamukelwa yizinkampani: Kunconywa ngokusemthethweni yi-Alibaba Cloud, Huawei Ascend, Amazon AWS ne-NVIDIA

• Izimo zemfundo: Isetshenziswa yi-Fudan University namanye amanyuvesi amaningi njengethuluzi lokufundisa izifundo zamamodeli amakhulu

• Ukuhambisana nesiShayina: Isekela ngokwendabuko i-ModelScope (umphakathi we-MoDa), kanye nokulanda izibuko ze-OpenMind Hub

• LLaMA Factory Online: Yake yanikeza isevisi yesikhulumi esiku-inthanethi, yavalwa ngo-2026-06-20

7.3 Ukuqhathaniswa nezimbangi

• Uhlaka｜Stars｜Isici esiyinhloko｜Ukusekelwa kwamamodeli｜Ukusebenziseka kalula

• LlamaFactory｜72K｜I-GUI/CLI ehlanganisa konke, iqinisekiswe yiphepha le-ACL｜100+｜⭐⭐⭐⭐⭐

• HuggingFace TRL｜12K｜Uchungechunge lwamathuluzi olusemthethweni lwe-HF, luhlanganiswe ngokujulile ne-ecosystem｜Wonke amamodeli e-HF｜⭐⭐⭐

• Axolotl｜8K｜Iqhutshwa yi-YAML, ukwenza ngokwezifiso kwabasebenzisi abathuthukile｜Amamodeli ajwayelekile｜⭐⭐⭐

• Unsloth｜20K+｜Ukulungiselelwa kokusebenza okwedlulele (ukusheshisa ngo-2-5x)｜LLaMA/Mistral njll.｜⭐⭐⭐⭐

• Firefly｜5K｜Ukulungiselelwa kwamamodeli amakhulu esiShayina｜Ikakhulu amamodeli esiShayina｜⭐⭐⭐

• LLaMA-Adapter｜6K｜Iqondiswe ocwaningweni lwezemfundo｜Uchungechunge lwe-LLaMA｜⭐⭐

Izinzuzo eziyinhloko ze-LlamaFactory:

• Uhlaka okuwukuphela kwalo oluhlanganisa iphepha le-ACL + ama-Stars angu-70,000 + izibonelo zokusetshenziswa kwezentengiso + i-GUI engadingi ikhodi 

• Inobubanzi obukhulu bokumbozwa kwamamodeli (kusuka ku-GPT-2 kuya ku-Llama 4, kusuka embhalweni omsulwa kuya kumamodi amaningi)

• I-ecosystem yomphakathi wase-China enamandla kakhulu (isibuko se-HuggingFace + ukuhlanganiswa kwe-ModelScope)

Okwesishiyagalombili、Ukuhlaziywa kwe-SWOT

8.1 Amandla (Strengths)

• Amandla｜Imininingwane

• 🔬 Igunya lezemfundo｜Iphepha le-ACL 2024, elinengcaphuno yezemfundo engaphezu kuka-500+

• 🎯 Umkhawulo ophansi kakhulu｜I-fine-tuning ngomyalo owodwa, ukusebenza kwe-Web UI ngaphandle kwekhodi

• 🌐 Ukumbozwa kwamamodeli okubanzi kakhulu｜Amamodeli angu-100+, uchungechunge olungu-60+, umbhalo/umbono/ividiyo/umsindo

• 🇨🇳 I-ecosystem yomphakathi wase-China｜Izibuko ze-ModelScope/OpenMind, izifundo eziningi zesiShayina

• 🔧 Uchungechunge lwamathuluzi lwezinga lezimboni｜I-Docker/ukusatshalaliswa okusabalalisiwe/ukuthunyelwa kwe-API konke kuyatholakala

• 💾 Ukulungiselelwa kwememori ye-GPU okwedlulele｜I-QLoRA 4-bit isebenzisa i-70B ku-GPU yezinga labathengi

8.2 Ubuthakathaka (Weaknesses)

• Ubuthakathaka｜Imininingwane

• 👤 Ukuncika kumuntu oyedwa kakhulu｜U-hiyouga wenza u-72% wama-commit (2,188/3,042), Bus Factor = 1

• 📚 Imibhalo ihlakazekile｜I-README yinde kakhulu (68KB), ulwazi lwe-Wiki/ReadTheDocs/izihloko zomphakathi luhlukene

• 🐛 Ukunqwabelana kwama-Issue｜Ama-Open Issues angu-57 kuphela, kodwa izingxoxo endaweni yama-Issue ziningi kakhulu

• 🏗️ Ukwehlukana kwemigqa emibili ye-v0/v1｜Izindleko zokugcina izakhiwo ezimbili ziphezulu, i-v1 ayikazinzi isikhathi eside

• 📦 Ukuncika kwamaphakheji kuyinkimbinkimbi｜I-matrix yezinguqulo ze-CUDA/PyTorch/FlashAttn kulula kakhulu ukwenza amaphutha

• 🔒 Isimo sephrojekthi yezemfundo｜Ayinalo iqembu elisebenza ngokugcwele, ukusimama kwesikhathi eside kuyangabazeka

8.3 Amathuba (Opportunities)

• Amathuba｜Imininingwane

• 🚀 Izimo ze-Agent fine-tuning｜Isidingo se-fine-tuning ye-Function Calling / Tool Use sikhula ngokushesha

• 🎥 Ukwandiswa kwamamodi amaningi｜Imakethe engakasetshenziswa kakhulu ye-fine-tuning yamamodeli e-Video/Audio/3D

• ☁️ I-SaaS ye-cloud-native｜Nakuba i-LLaMA Factory Online ivaliwe, isidingo sokusingathwa kwezinga lebhizinisi sisekhona

• 🏢 Imakethe yokuqeqeshwa kwezinkampani｜Izifundo zokuqeqesha neziqu ze-fine-tuning yamamodeli amakhulu

• 🔗 Ukuhlanganiswa kwe-MCP｜I-backend ye-fine-tuning njengengqalasizinda ye-AI Agent

8.4 Izinsongo (Threats)

• Izinsongo｜Imininingwane

• 🏢 Izinhlaka ezakhiwe yizinkampani ezinkulu｜I-Alibaba PAI ne-Huawei ModelArts zinikeza izixazululo ezihlanganisiwe

• 🔄 Ukuncintisana okusemthethweni kwe-HF｜I-ecosystem ye-HuggingFace TRL/peft iya ngokuya iphelela

• 📉 Ukuncipha kwesidingo se-fine-tuning｜I-Prompt Engineering + RAG kuthatha indawo yezinye izimo ze-fine-tuning

• 👨‍🎓 Ukuphothula komsunguli｜Ngemva kokuphothula kukaDkt. Zheng Yaowei, ukunakekelwa kwephrojekthi kungase kwehle

Okwesishiyagalolunye、Ukuhlolwa kobungozi

• Uhlobo lwengozi｜Izinga lengozi｜Incazelo

• Bus Factor｜🔴 Ingozi ephezulu｜Unjiniyela oyedwa oyinhloko wenza u-72% wama-commit; uma ehamba, ingozi yokuma kwephrojekthi iphezulu kakhulu

• Isikweletu sobuchwepheshe｜🟡 Ingozi emaphakathi｜Izakhiwo ezimbili ze-v0/v1, i-README engu-68KB, imibhalo ehlakazekile

• Uchungechunge lwezinc依｜🟡 Ingozi emaphakathi｜Izindleko zokugcina i-matrix yezinguqulo ze-CUDA/PyTorch/FlashAttn ziphezulu

• Ukwenziwa kwemali｜🟢 Ingozi ephansi｜Ilayisensi ye-Apache-2.0, akukho mingcele yezohwebo

• Impilo yomphakathi｜🟢 Inempilo｜Ama-Stars angu-7.2万, umphakathi womhlaba wonke osebenzayo, ukusetshenziswa kabanzi ngamabhizinisi

10. Amaphuzu avelele obuchwepheshe kanye nokusungula

10.1 Ukwakhiwa “komgwaqo omkhulu osheshayo ojwayelekile”

Ekuqeqesheni okujwayelekile kokulungisa, imodeli ngayinye idinga ukujwayelaniswa ngokuhlukile. I-LLaMA Factory isebenzisa indlela yokubhalisa izifanekiso kanye nobuchwepheshe be-monkey‑patch, okunciphisa kakhulu izindleko zokujwayelanisa.

# Ukwengeza imodeli entsha kudinga izinyathelo ezi-3 kuphela:
# 1. Bhalisa isifanekiso se-Prompt kusichazamazwi se-TEMPLATES ku-template.py
# 2. Engeza isichibi sokuhambisana kwemodeli ku-patcher.py (uma kudingeka)
# 3. Engeza ukwesekwa kwe-multimodal ku-mm_plugin.py (uma kudingeka)

10.2 Uhlelo lokumisa amapharamitha olunezingqimba ezine

Ifayela lokumisa le-YAML/JSON
  → parser.py iyalihumusha
    → ModelArguments (ukukhetha imodeli, quantization)
    → DataArguments (isethi yedatha, isifanekiso, preprocessing)
    → FinetuningArguments (i-LoRA rank/target, indlela yokuqeqesha)
    → TrainingArguments (izuza ku-HF, ukusatshalaliswa/amalogi/checkpoint)

10.3 Ukumiswa okujwayelekile kwepayipi ledatha

Leli thuluzi lisekela amafomethi edatha amabili ajwayelekile, i-Alpaca ne-ShareGPT, lingaguqula ngokuzenzakalela amasethi edatha angaphezu kwama-50 ahlelwe kusengaphambili, futhi lihambisana namafomethi amafayela e-JSON, JSONL, CSV, Parquet ne-Arrow.

10.4 Ukwesekwa kwe-quantization okuqhubekayo

Lisekela amazinga ayisithupha okunemba, kusukela ekunembeni okuphelele kwamabhithi angu-16 kuya ku-QLoRA yamabhithi angu-2, futhi lingasetshenziswa ezindaweni zehadiwe ezihlukene, kusukela kumaqoqo amaseva e-A100 kuya kumalaptop ajwayelekile.

11. I-matrix yokuqhathanisa nezimbangi

• Isici｜LlamaFactory｜HuggingFace TRL｜Unsloth｜Axolotl

• Stars｜72K｜12K｜20K+｜8K

• Ukwesekwa kwamamodeli｜Uchungechunge olungu-100+｜Wonke ama-HF｜LLaMA/Qwen/Mistral｜Uchungechunge olujwayelekile

• GUI｜✅ LLaMA Board｜❌｜❌｜❌

• Ngaphandle kwekhodi｜✅ CLI + YAML｜❌ Kudinga ukubhala ikhodi｜❌ Kudinga ukubhala ikhodi｜I-YAML kuphela

• Izindlela zokuqeqesha｜8｜4 (SFT/DPO/PPO/GRPO)｜3｜5

• Ukujula kwe-quantization｜2-16bit｜4-16bit｜4-16bit｜4-16bit

• Multimodal｜✅ Isithombe/ividiyo/umsindo｜❌｜❌｜❌

• Imvelo yesiShayina｜✅ ModelScope + 魔搭｜⚠️ Kudinga i-VPN｜⚠️ Kudinga i-VPN｜❌

• Iphepha lezemfundo｜ACL 2024｜-｜-｜-

• Injini yokwenza inference｜vLLM/SGLang/HF｜HF kuphela｜HF kuphela｜vLLM/HF

• Ijika lokufunda｜Liphansi kakhulu｜Elimaphakathi｜Liphansi｜Eliphakathi kuya phezulu

12. Amaphuzu abalulekile okusebenzisa ekukhiqizeni

12.1 Ukuqalisa ngokushesha (izinyathelo ezi-3)

# Isinyathelo 1: Faka
git clone --depth 1 github.com/hiyouga/LlamaFactory.git
cd LlamaFactory
pip install -e ".[torch,metrics]"

# Isinyathelo 2: Qalisa i-Web UI
llamafactory-cli webui

# Isinyathelo 3: Vula i-localhost:7860 esipheqululini, khetha imodeli → isethi yedatha → qeqesha!

12.2 Izincomo zendawo yokukhiqiza

# Ukusetshenziswa kwe-Docker (kunconyiwe)
docker run -it --rm --gpus=all \
    -v ./hf_cache:/root/.cache/huggingface \
    -v ./data:/app/shared_data \
    -v ./output:/app/output \
    -p 7860:7860 -p 8000:8000 \
    hiyouga/llamafactory:latest

# Ukusetshenziswa kwe-API (okuhambisana ne-OpenAI)
API_PORT=8000 llamafactory-cli api \
    examples/inference/qwen3.yaml \
    infer_backend=vllm

12.3 Okumele kuqashelwe abasebenzisi be-Windows

• Kudingeka ufake mathupha inguqulo ye-PyTorch GPU: pip install torch torchvision torchaudio --index-url download.pytorch.org/whl/cu126

• I-QLoRA idinga ukufakwa kwenguqulo ye-bitsandbytes ehlanganiswe kusengaphambili, ekhethekile ku-Windows

• I-FlashAttention-2 idinga ukuhlanganiswa mathupha ku-Windows

• Setha dataloader_num_workers: 0 ukugwema amaphutha e-serialization

13. Umbono wesikhathi esizayo

1. Ukuzinziswa kwesakhiwo se-v1: Igatsha lokuhlola kulindeleke ukuba libe yisakhiwo esiyinhloko ngo-2026-2027, ngezinga eliphezulu lokumodulara

2. Ukujuliswa kokulungiswa kwe-Agent: Ukulungiswa kwe-Tool Use / Function Calling kuzoba umsebenzi ojwayelekile

3. Ukulungiswa kwamamodeli e-inference (Reasoning): I-DeepSeek-R1 distillation + ukulungiswa kwe-GRPO (iphrojekthi ye-EasyR1)

4. Ukusingathwa kwe-cloud-native: Ngemva kokuvalwa kwe-LLaMA Factory Online, kungavela izixazululo zokusingathwa zezinkampani zangaphandle

5. Ukuhlanganiswa kokuqeqeshwa kwe-multimodal: Umbhalo/izithombe/ividiyo/umsindo kulungiswe ngaphakathi kohlaka olulodwa

14. Ukukala kanye nesifinyezo

• Isici｜Amaphuzu｜Ukuhlaziya

• Isakhiwo sobuchwepheshe｜9.5/10｜Ukwakhiwa okumodular, ukuvela kwezakhiwo ezimbili, i-matrix yezinjini ebanzi

• Ukuphelela kwemisebenzi｜10/10｜Izindlela zokuqeqesha eziyi-8 × izindlela zokulungisa eziyisi-6 × amamodeli angu-100+

• Ukusebenziseka kalula｜10/10｜Ukulungisa ngomyalo owodwa + i-Web UI engenakhodi, ehamba phambili embonini

• Ikhwalithi yemibhalo｜7.5/10｜I-README inemininingwane kodwa yinde kakhulu, ihlakazeke kakhulu

• Umsebenzi womphakathi｜9.5/10｜Ama-Stars angu-72K, ukuthandwa okumangalisayo emphakathini wase-China

• Ukulungela ukukhiqiza｜9.0/10｜I-Docker/API/ukusatshalaliswa kuphelele, kuqinisekiswe kabanzi ngamabhizinisi

• Ukusimama｜6.5/10｜Bus Factor=1, izici zephrojekthi yezemfundo, ikusasa lesikhathi eside liyangabazeka

• Ukusungula｜8.5/10｜Iphepha le-ACL, indlela yokubhalisa izifanekiso, uhlelo lwamapharamitha olunezingqimba ezine

Amaphuzu esewonke: 9.

Isifinyezo ngomusho owodwa

I-LLaMA Factory iwuhlaka oluvulekile lokulungisa amamodeli amakhulu, olwakhiwe ngokusekelwe ephepheni le-ACL 2024 futhi oluthole ukuqashelwa komphakathi ngama-Stars acishe abe ngu-72 000 ku-GitHub. Uhlaka lulandela umqondo wokuklama “wokuhlukanisa i-algorithm, imodeli nedatha”, lunikeza i-command-line interface (CLI) ne-Web UI engenakhodi, futhi lusekela ukulungiswa kwamamodeli angaphezu kwe-100, ngaleyo ndlela lwehlise kakhulu umkhawulo wobuchwepheshe wokulungisa amamodeli amakhulu. Lungasetshenziswa njengengqalasizinda yokusebenzisa izinhlelo zamamodeli amakhulu.