flame
flame copied to clipboard
xflops
A distributed system for Agentic AI
Currently, the resource will only reshuffle when task completed; it's better to provide preemption feature to speed it up.
Currently, we only build Rust client; it's better to support more client, e.g. python, R
It's better to provide more machinary for Session/Task error handling, e.g. retry limits.
Currently, we only support log & stdio shim; it's better to support more shim, e.g. gRPC & Restful.
Currently, the instance will be start/stop for each task in stdio-shim. It's better to start the instance in `on_session_enter` and stop the instance in `on_session_leave`; and communicate with the instance...
**Motivation** Currently, we already support rust-sdk; so it's easy to integrate with [tch-rs](https://github.com/LaurentMazare/tch-rs) as an AI example. **Function Specification** Provide distributed training/inference example of [tch-rs](https://github.com/LaurentMazare/tch-rs).
**Motivation** Currently, we already support grpc shim in Flame. It's better to enhance python SDK for grpc shim to build `FlameService` callback in Python. **Function Specification** Introduce a FlameService python...
**Motivation** Currently, we already move the application configuration into FSM, but it sill missed the following gRPCs: `register_applicatin`, `unregister_application`, `update_application` and `get_application`. The purpose of this issue is to enhance...