Latest news, updates, and insights from the XFLOPS team and community.
Use hermes-flamepy to execute agent-generated code on a Flame cluster: preserve working files across tool calls, submit independent snippets together, and inspect individual failures.
Two controlled evaluations of Flame serving: a 0.55–0.86% median throughput difference with matched prefix placement, and a 16.8% DAS gain over round-robin vLLM under cache pressure.
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.
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.
A walkthrough of the Candle Based Rust example in Flame, showing that the main code change is wrapping local inference in a Flame service macro, while flmctl deploy handles packaging and registration.
A technical blog on the TorchRL DQN example in Flame, covering distributed collectors, Flame-backed TorchRL replay storage, sharded sampling, and an Apple M4 Podman test run.
A technical blog on Flame's replay-buffer example, covering patch_object delta updates, parallel sampling, and the handler-plus-data programming model.
