LMCache/LMCache

[RFC]: Non-Python backends

已關閉

#1,362 建立於 2025年8月15日

在 GitHub 查看貢獻者指南
 (11 則留言) (4 個反應) (0 位負責人)Python (1,195 個分叉)user submission
RFCdiscussionenhancementhelp wantedoptimizationstale

倉庫指標

星標
 (8,345 顆星)
PR 合併指標
 (平均合併 6天 13小時) (30 天內合併 204 個 PR)

描述

The GIL can greatly hinder remote backends. For LMCache's vision of being able to connect many serving engines effectively, backends written in lower level languages are critical to maximize core usage and saturate NIC/bus bandwidths.

In the future, maybe the disk/SSD backend can be rewritten in non-python language too

Please leave any comments / thoughts below and we will update this RFC with updates!

One example of a non-performant python backend is redis: https://github.com/LMCache/LMCache/pull/1293

even though store is done asynchronously from vllm, the TTFT for the warm up round (store only) meaning the redis SDK's concurrency model is flawed

貢獻者指南

研究方向
Investigate the performance bottlenecks of Python backends, particularly focusing on the GIL impact as seen in PR #1293. Evaluate candidate languages such as Rust or C for implementing backends, considering performance benchmarks and integration with existing Python codebase. Review comments in the issue for additional requirements and constraints.
技術棧
python
領域
backendperformance
議題類型
功能
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
5
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
超過 1 週
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
活躍
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
清晰
前置要求
無
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
5