From 15eb0e1b7e85f076cd660e3d3ad30e94949f305d Mon Sep 17 00:00:00 2001 From: Anjiangy Date: Sat, 14 Nov 2026 16:24:42 +0800 Subject: [PATCH] Add PROGRAM-KV entry pointing at ride-kv-controller --- data/control-plane-projects.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/data/control-plane-projects.json b/data/control-plane-projects.json index fb42a45..4d667c2 100644 --- a/data/control-plane-projects.json +++ b/data/control-plane-projects.json @@ -1 +1 @@ -{"schema_version":1,"projects":[{"code":"SLO","name":"SLO-Aware Agent Serving","url":null,"summary_en":"Admission, calibrated prediction, fairness, and resource allocation for mixed agent workloads.","summary_zh":"面向混合智能体负载的准入、校准预测、公平性与资源分配。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Liu Jun","github":"iliujunn"}},{"code":"MODEL","name":"VAMOS","url":null,"owner":"Multi-model control","summary_en":"VRAM-aware multi-model orchestration under serving SLOs.","summary_zh":"面向服务 SLO 的显存感知多模型编排。","status_en":"RESEARCH","status_zh":"研究中"},{"code":"BUDGET","name":"Token-Budget Governor","url":null,"owner":"Admission control","summary_en":"Token-budget admission and overload control before work enters the inference engine.","summary_zh":"在请求进入推理引擎前执行 token 预算准入与过载控制。","status_en":"INCUBATING","status_zh":"孵化中"},{"code":"STATE","name":"Agent State Tiering","url":null,"summary_en":"Workflow-aware placement and prefetch across accelerator, host-memory, and storage tiers.","summary_zh":"跨加速器、主机内存与存储层的工作流感知放置和预取。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Ou Dandan","github":"oddod"}},{"code":"ROUTE","name":"Prefix Cache Routing Reliability","url":null,"summary_en":"Failure-safe distributed prefix-cache routing, stale-view handling, and bounded recovery.","summary_zh":"故障安全的分布式前缀缓存路由、陈旧视图处理与有界恢复。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Wang Jie","github":"WMASTER123"}},{"code":"AGENT-KV","name":"FreshKV","url":null,"summary_en":"Owns agent-aware future-reuse signals, value calibration, physical KV mapping, retention and eviction decisions, and auditable execution receipts.","summary_zh":"统一负责 Agent 感知的未来复用信号、价值校准、物理 KV 映射、保留与淘汰决策，以及可审计的执行回执。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Xincai Zhu","github":"Pygone"}},{"code":"QUALITY","name":"Quality-Bounded Inference","url":null,"summary_en":"Maps task-level quality targets to token budgets, decode policies, and service controls.","summary_zh":"把任务级质量目标映射为 token 预算、解码策略与服务控制动作。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Peng Haoran","github":"Tkhkrnx"}},{"code":"WORKFLOW","name":"Workflow-Aware Serving","url":null,"summary_en":"Uses DAG readiness, critical paths, tool waits, and deadlines to coordinate serving decisions.","summary_zh":"依据 DAG 就绪状态、关键路径、工具等待与截止时间协调服务决策。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Mao Yancan","github":"yancanmao"}},{"code":"E&A","name":"MottoServe","url":null,"summary_en":"Evidence and analysis of serving-policy behavior, planner overhead, and applicability boundaries.","summary_zh":"研究服务策略行为、规划器开销与适用边界的证据分析课题。","status_en":"RESEARCH RECORD","status_zh":"研究记录","lead":{"name":"Li Xuheng","github":"sssarrior"}}]} +{"schema_version":1,"projects":[{"code":"SLO","name":"SLO-Aware Agent Serving","url":null,"summary_en":"Admission, calibrated prediction, fairness, and resource allocation for mixed agent workloads.","summary_zh":"面向混合智能体负载的准入、校准预测、公平性与资源分配。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Liu Jun","github":"iliujunn"}},{"code":"MODEL","name":"VAMOS","url":null,"owner":"Multi-model control","summary_en":"VRAM-aware multi-model orchestration under serving SLOs.","summary_zh":"面向服务 SLO 的显存感知多模型编排。","status_en":"RESEARCH","status_zh":"研究中"},{"code":"BUDGET","name":"Token-Budget Governor","url":null,"owner":"Admission control","summary_en":"Token-budget admission and overload control before work enters the inference engine.","summary_zh":"在请求进入推理引擎前执行 token 预算准入与过载控制。","status_en":"INCUBATING","status_zh":"孵化中"},{"code":"STATE","name":"Agent State Tiering","url":null,"summary_en":"Workflow-aware placement and prefetch across accelerator, host-memory, and storage tiers.","summary_zh":"跨加速器、主机内存与存储层的工作流感知放置和预取。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Ou Dandan","github":"oddod"}},{"code":"ROUTE","name":"Prefix Cache Routing Reliability","url":null,"summary_en":"Failure-safe distributed prefix-cache routing, stale-view handling, and bounded recovery.","summary_zh":"故障安全的分布式前缀缓存路由、陈旧视图处理与有界恢复。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Wang Jie","github":"WMASTER123"}},{"code":"AGENT-KV","name":"FreshKV","url":null,"summary_en":"Owns agent-aware future-reuse signals, value calibration, physical KV mapping, retention and eviction decisions, and auditable execution receipts.","summary_zh":"统一负责 Agent 感知的未来复用信号、价值校准、物理 KV 映射、保留与淘汰决策，以及可审计的执行回执。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Xincai Zhu","github":"Pygone"}},{"code":"QUALITY","name":"Quality-Bounded Inference","url":null,"summary_en":"Maps task-level quality targets to token budgets, decode policies, and service controls.","summary_zh":"把任务级质量目标映射为 token 预算、解码策略与服务控制动作。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Peng Haoran","github":"Tkhkrnx"}},{"code":"WORKFLOW","name":"Workflow-Aware Serving","url":null,"summary_en":"Uses DAG readiness, critical paths, tool waits, and deadlines to coordinate serving decisions.","summary_zh":"依据 DAG 就绪状态、关键路径、工具等待与截止时间协调服务决策。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Mao Yancan","github":"yancanmao"}},{"code":"E&A","name":"MottoServe","url":null,"summary_en":"Evidence and analysis of serving-policy behavior, planner overhead, and applicability boundaries.","summary_zh":"研究服务策略行为、规划器开销与适用边界的证据分析课题。","status_en":"RESEARCH RECORD","status_zh":"研究记录","lead":{"name":"Li Xuheng","github":"sssarrior"}},{"code":"PROGRAM-KV","name":"Program-Aware Request/KV Controller","url":"https://github.com/Anjiangy/ride-kv-controller","summary_en":"Admission and KV-residency control keyed by program progress. The controller packs the resident set across cards by block budget, parks or reloads a program's KV on demand, and switches between fine-grained swapping and batch rotation as oversubscription grows. Measured on 2x Ascend 910B2: at 2.1x per-card oversubscription, 74.4 vs 51.9 step/min against plain round-robin and 5.4 against a request-level router, p50 0.66 s vs 6.33 s.","summary_zh":"以 program 进度为单位的准入与 KV 常驻控制：按块预算把驻留集合装箱到多卡，按需把 program 的 KV 换出到主机内存、再换回（支持跨卡迁移），并随超容比在细粒度换入换出与整批轮转之间切换。2 张昇腾 910B2 实测：每卡 2.1× 超容下，异质负载 74.4 vs 轮询 51.9 step/min、vs 请求级路由器 5.4，p50 0.66 s vs 6.33 s。","status_en":"INCUBATING","status_zh":"孵化中","lead":{"name":"Anjiangy","github":"Anjiangy"}}]} \ No newline at end of file