# Flyte An open-source runtime for durable AI workflows, model training, serving, and agents. - Findings24: - ウェブサイト: - 料金: Freemium - カテゴリー: AIエージェント基盤, AI推論プラットフォーム, AIワークフロー自動化 ## Flyteとは？ Flyte is an open-source orchestration platform whose Flyte 2 runtime runs AI workloads across training, inference, and agent execution. It coordinates tasks and compute resources while recording execution inputs and outputs so interrupted workflows can resume from failed steps. Developers write workflows as Python functions with ordinary branching and loops. Task environments specify container images, resources, secrets, and retry policies. The platform supports deployment on your own infrastructure, with an enterprise offering available through Union.ai. ## Flyteでできること ### Infrastructure-aware recovery Recover from out-of-memory kills, node preemption, and lost GPUs. Configure exponential retry backoff, per-attempt runtime limits, and execution deadlines; infrastructure-triggered retries are handled separately from the task retry budget. ### Dynamic Python workflows Use async functions, conditionals, loops, and exception handling to compose tasks. Typed inputs and outputs support files, DataFrames, and dataclasses without a separate workflow configuration language. ### Parallel execution and batching Run parallel tasks and distributed jobs with map and fan-out patterns. DynamicBatcher groups inference requests using cost budgets, batch size limits, dispatch timeouts, and queue backpressure. ### Model and application serving Deploy long-running apps through FastAPI, Streamlit, vLLM, SGLang, and Ollama integrations. Configure resources, replicas, and scale-to-zero behavior alongside workflows. ### Execution records and caching Executions are logged and versioned with their inputs and recovery events visible in the UI. Task caching can reuse results for identical inputs. ### Sandboxes for generated code Run LLM-generated orchestration code in a Monty-based Python sandbox that dispatches work to container tasks, or execute scripts and shell commands in disposable containers. ## 利用シーン ### Distributed model training Coordinate fine-tuning or pretraining across GPU resources, with integrations such as Ray and PyTorch Elastic for distributed training jobs. ### Large-scale batch inference Process datasets in parallel shards while batching requests to keep GPUs occupied. Retry an unsuccessful shard without rerunning completed work. ### Agent execution loops Build agents that plan, invoke tools, and revise actions based on results. Wrap tools as tasks and trace LLM calls across the decision loop. ### Model endpoints and dashboards Expose models through OpenAI-compatible endpoints or FastAPI services, and provide interactive Streamlit dashboards alongside processing workflows. ## よくある質問 ### How do I start a local development environment? Install the SDK with \`pip install flyte\`, then run \`flyte start devbox\`. The platform page shows the local Flyte UI at http://localhost:30080. ### Can Flyte run on my own infrastructure? Yes. Flyte supports open-source deployment on your infrastructure. Union.ai provides the enterprise Flyte platform with managed infrastructure. ### Can tasks fail immediately when retries would not help? Yes. Raise \`NonRecoverableError\` for invalid inputs or malformed configuration to stop a task without exhausting its retry budget. ### How does Flyte connect to existing data and ML tools? Its integration catalog includes Spark, BigQuery, Snowflake, Databricks, Weights & Biases, MLflow, and Dask. Some integrations require backend setup, and developers can create additional integrations. ## クイック情報 - **License**: Apache 2.0 - **Project affiliation**: Linux Foundation AI & Data - **Current runtime**: Flyte 2 - **Workflow authoring**: Python SDK - **Website**: https://flyte.org