07-Fine-tuning LLM with Argo Workflows - A Kubernetes-native Approach-Shuangkun Tian.pdf

1、Fine-tuning LLM with ArgoWorkflows：A Kubernetes-nativeApproachShuangkun TianArgo Maintainer、Alibaba Cloud Software EngineerThe Challenge of Fine-tuning01Building a TCM Assistant on DeepSeek03Why Argo Workflows for Fine-tuning02ContentBenefits and Future04The Challenge of Fine-tuningPart 01What is Fi

2、ne-tuning？Adapts pre-trained models to specific tasks/domains via targeted training.Base Model:Deepseek R1、GPT-3、BERTFine-tuned Model:DeepSeek-Finance、Claude、SciBERTChallenge of Fine-tuning Substantial Computational Resources Various heterogeneous devices：CPU、GPU、DPU Computa8onal Cost：Single trainin

3、g run over$10k+Complex Workflows Mu8-stages：Preprocessing Training Evalua8on Larger Tasks（1w+）、Mu8-workflows（1000+）Manual workflows=High cost+Low reliabilityWhy Argo Workflows？Part 02What is Argo Workflows？Argo ProjectArgo WorkflowsArgo CDArgo Events、Argo Rollout Third Active Community in CNCF Most

4、Popular Workflow EngineMachine Learning pipelinesData and batch processingInfrastructure automationCI/CDArgo Workflows for ML Pipelines Over 8K companies use Argo or those ML tools based on Argo Argo Workflows has been a core component in the orchestration of AI/ML workloads on Kubernetes.Argo Workf

5、lowsKubeflow PipelinesNuma flowWhy Argo Workflows for AI/Fine-tuning Kubernetes-native Scalability Reproducibility Fault tolerance Visibility Ease of Use YAML/Python Building a Traditional Chinese Medicine Assistant Based on DeepSeekPart 03Workflow DefintionFine-tuning LLM Dataset Prepare HuggingFac

6、e Hub(e.g.,datasets.load_dataset(glue,marc)Raw Data Clean Tokenize Base Model DeepSeek-R1、DeepSeek-R1-Distill 4-bit Quantized Base Model Training LoRA Adapters Full Fine-tuning Evaluation Human evaluationFine-tuning Chinese Medicine Assistant on Deepseek Tradi