Bringing Flame Serving Close to Standalone vLLM
How prefix placement, batching, and an asynchronous adapter brought Flame within 1 percent of standalone vLLM median throughput on two B200 GPUs.
Read More →
An open source community to empower AI applications
XFLOPS helps enterprises build AI applications with open tools, frameworks, and best practices. Flame is the community's core project: a distributed engine for secure, cost-effective, and high-performance elastic AI workloads.
What XFLOPS Focuses On
Elastic
Scale AI workloads dynamically based on demand with runtime-aware scheduling and resource optimization.
Security
Use session-based access, isolated runtime environments, and secure component communication for elastic workloads.
Cost Effective
Optimize resource utilization and workload distribution so AI systems can run efficiently at scale.
Heterogeneous
Support varied infrastructure, including CPU, GPU, and accelerator-backed environments.
High Performance
Improve throughput and roundtrip time for task-heavy AI systems through distributed execution.
Cloud Native
Build portable runtime infrastructure for cloud, on-premise, and hybrid deployments.
Core Project: Flame
Flame is the core XFLOPS project for elastic AI workloads. It provides the distributed runtime mechanisms behind sessions, task scheduling, executor reuse, object caching, and secure service integration for agents, reinforcement learning, generated-code execution, and more.
How Flame Works
- Session: A group of related tasks with scheduling, resource, and isolation boundaries.
- Task: A unit of work submitted by a client and executed by a service in an executor.
- Executor: A runtime environment that hosts application services for a session.
- Object cache: A shared data layer used by Runner, common data, and incremental object updates.
Latest from the XFLOPS Blog
The three latest technical walkthroughs from the XFLOPS community and Flame project. Explore Flame serving benchmarks against standalone vLLM, Search-R1 training, and the distributed runtime behind these workloads.
Training a Search-R1 Agent with Flame
How Flame simplifies Search-R1 training with Python services, warm GPU executors, elastic rollouts, gang-scheduled trainers, and incremental weight transfer, demonstrated in a 100-step run.
Read More →
Flame v0.6: A Distributed Engine for Elastic Workloads
Flame v0.6 adds Runner, object cache, and install tooling for elastic jobs such as agents, RL, and crawlers. Batch jobs stay with Volcano; workflow jobs are not covered yet.
Read More →