00 / EGGYBYTE TECHNOLOGY

为 AI 执行而生的基础设施。为 AI 执行而生的基础设施。

Infrastructurepurpose-builtfor AI.Infrastructurepurpose-builtfor AI.

从公网边缘、分布式控制与隔离执行,
Model GatewayAI Runtime 与模型工程,
我们构建面向大规模 AI 产品的纵向系统栈。

From high-bandwidth edge ingress and distributed control to isolated execution, Model Gateway, AI Runtime and model engineering — one integrated systems stack for AI products at scale.

向下进入系统Explore the system
Shanghai · Systems / AI
Public EdgeQUIC v1 · TLS 1.3
AuthorityController · Product State
ExecutionNodeSession · YaoBox
IntelligenceModel Gateway · Models
ONE AI EXECUTION FABRIC / 统一执行织层ONE AI EXECUTION FABRICNetwork → Authority → Execution → Intelligence
01 / 25Gbps 公网入口01 / 25Gbps PUBLIC EDGE

25Gbps 公网入口,按流量横向扩展。25Gbps 公网入口,按流量横向扩展。

25Gbps public ingress.Scale out by traffic.25Gbps public ingress.Scale out by traffic.

25Gbps 级公网 Edge 作为可替换数据平面单元;
公网流量增长时直接增加 Edge 实例,Controller 主动建立 QUIC v1 / TLS 1.3 连接,
通过固定 Ed25519 身份保持认证边界,业务权威不随边缘扩张。

Use 25Gbps-class public Edge instances as replaceable data-plane units. As ingress grows, add Edge capacity horizontally; Controller initiates QUIC v1 / TLS 1.3 connections with fixed Ed25519 identity while business authority stays in the core.

25Gbps单元级公网入口PUBLIC EDGE UNIT
+ Edge = + 公网容量+ EDGE = + INGRESS CAPACITY
Internet → Edge Pool → ControllerQUIC v1 · TLS 1.3 · fixed Ed25519 identity
AUTHORITY / 权威平面AUTHORITY PLANE
Controller
主动连接 Edge · 无状态副本initiates Edge links · stateless replicas
Product StateFoundationDB / Yaostate
执行平面Execution Planepull-based · no inbound Node RPC
NodeSessionYaoBoxTask
需求 ↑ → + NodeDEMAND ↑ → + NODE
数据平面Data Planedata traffic / storage
Data GatewayStorage BoxBlob
需求 ↑ → + GatewayDEMAND ↑ → + GATEWAY
智能平面Intelligence Planerouting decoupled from nodes
Model GatewayEmberNova
需求 ↑ → + Model GatewayDEMAND ↑ → + MODEL GATEWAY
Cross-unit: Protobuf / ConnectController 可替换 · Product State 权威replaceable Controller · authoritative Product State
02 / SCALE BY ROLE

每一种资源,都有自己的扩展边界。每一种资源,都有自己的扩展边界。

Scale by role,not by clusterboundary.Scale by role,not by clusterboundary.

ControllerExecution NodeData Gateway
Model GatewayStorage 按各自需求横向增加;
机器规模增长不要求重定义整体系统拓扑。

Controller, execution nodes, Data Gateway, Model Gateway and storage expand according to their own demand without redefining the topology as machine count grows.

需求增长只扩展对应职责DEMAND GROWTH EXPANDS ONE ROLE
角色Role独立容量域Capacity domainΔ
ControllerAuthority
Steady
Execution NodeCompute / YaoBox
+4
Model GatewayIntelligence
+2
Data GatewayData Traffic
Steady
StorageBlob / Asset
+1
架构目标DESIGN GOAL · no pre-fixed cluster scale boundary示意模型 · 非 benchmarkillustrative model · not a benchmark
03 / PURPOSE-BUILT AI RUNTIME

更少的层,更明确的执行边界。更少的层,更明确的执行边界。

Fewer layers.Explicit executionboundaries.Fewer layers.Explicit executionboundaries.

YaoBox 不以 PodService Mesh 或通用 Scheduler
作为核心抽象,
而直接围绕 TaskNodeSessionGateway
与 Linux 资源边界组织 Runtime

YaoBox is defined around Task, NodeSession, gateways and Linux resource boundaries instead of a generic Pod / Service Mesh / scheduler abstraction.

fewer layersexplicit ownershipno inbound node rpc
进程边界ProcessPID namespace · crun
身份边界Identityuser namespace
资源边界Resourcecgroup v2
YaoBoxTask execution envelope
Task
vethoverlay
cgroup v2
文件系统FilesystemEROFS · OverlayFS
网络边界Networknetwork namespace · veth
配额边界QuotaXFS project quota
Node 主动拉取命令NODE PULLS COMMANDS · 执行环境拥有独立 PID / network / user / cgroup / writable filesystem 边界isolated PID / network / user / cgroup / writable filesystem boundaries
04 / INTELLIGENCE PLANE

执行计算与模型智能,可以独立演进。执行计算与模型智能,可以独立演进。

Compute andintelligence scaleindependently.Compute andintelligence scaleindependently.

Execution Node 只提交绑定当前
Task 的标准 Model Request
Model Gateway 负责 staging、routing、
Provider CredentialProvider effect
与流式响应标准化。

Execution nodes submit a standard Model Request bound to the current Task. Model Gateway owns staging, routing, provider credentials, provider effects and stream normalization.

Provider Credential 终止于 GatewayPROVIDER CREDENTIAL TERMINATES AT GATEWAYExecution Node 不持有真实 Provider Endpoint 或 Credential;计算容量与模型容量独立扩展。Execution Nodes do not hold real provider endpoints or credentials; compute and model capacity scale independently.
Task-bound RequestNode 只知道内部系统语义internal semantics only
Grant / OwnershipController 定义 authority 与计费语义authority + billing semantics
MODEL GATEWAY
Routemodel selection / staging
Provider Effectauthorized external execution
Credential Boundary
Normalizestandard response semantics
Streamnormalized streaming output
staging · effect · stream normalization
compute capacity ≠ model capacity
Provider A
Provider B
Ember
Nova
Normalized response stream
05 / RELIABILITY & CORRECTNESS

故障显式发生,恢复结果可确定。故障显式发生,恢复结果可确定。

Failure is explicit.Recovery isdeterministic.Failure is explicit.Recovery isdeterministic.

外部 effect 不被隐藏在自动 retry 中。
系统先建立唯一 ownership,再执行外部 effect
最后通过显式 readback 或确定结果完成状态提交。

External effects are not hidden behind automatic retry. The system establishes unique ownership first, performs the external effect, then resolves it through explicit readback or a deterministic outcome.

故障显式failure explicitAuthority 有界bounded authority恢复确定deterministic recovery
权威状态AUTHORITATIVE STATE
POSSIBLE请求已持久化 · 唯一 identitypersisted request · unique identity
GatewayInstance唯一 product ownerbound as the single product owner
RESOLVED确定状态已提交deterministic state committed
外部 effect 不确定窗口UNCERTAIN EFFECT WINDOW
外部 EFFECTEXTERNAL EFFECT
Provider Effect首个外部请求字节first external request byte
Readback显式确定外部结果resolve external result
Provider effect 至多一次AT-MOST-ONCE PROVIDER EFFECT
×禁止第二 Gateway 自动重放SECOND GATEWAY AUTO-REPLAY FORBIDDEN
06 / MODEL-SYSTEM CO-DESIGN

模型与系统,作为同一个工程问题。模型与系统,作为同一个工程问题。

Model and system,engineered as one.Model and system,engineeredas one.

模型结构、数值体系、GPU 通信、长上下文执行和训练 /
rollout policy identity 被共同定义,
Research 直接连接最终 RuntimeServing

Architecture, numerics, GPU communication, long-context execution and training / rollout policy identity are co-defined so research connects directly to runtime and serving.

同一工程表面ONE ENGINEERING SURFACE
Architecture ↔ Numerics ↔ Runtime ↔ Serving
模型架构Model Architecturestructure / attention / state
Granular MoEGated DeltaNetHierarchical Sparse Attention
训练系统Training Systemnumerics / optimization / communication
NVFP4-firstMXFP8 AttentionGPU CommunicationMuon
Runtime & Servinglong context / agent / rollout
StateFlowLong ContextAgent / RLVRServing
07 / REAL-WORLD SYSTEMS

让技术真正进入大规模商业产品。让技术真正进入大规模商业产品。

A systemsfoundation forreal products.A systemsfoundation forreal products.

InfrastructureCommerceDistribution 与产品形态
连接到同一技术底座。
YaoledgeYaoedgeEmberNova
是这套系统在不同层面的表达。

Infrastructure, commerce, distribution and product surfaces connect to one technical foundation. Yaoledge, Yaoedge, Ember and Nova are different expressions of that system.

EggyByte Technology上海旦享计算机科技有限责任公司
Shanghai Danxiang Computer Technology Co., Ltd.
基础设施InfrastructureOVH · ISIF · Aéza ·
25Gbps Edge
商业基础设施CommerceAlipay
EggyByte Systems
Foundation
Network · Authority
Execution · Intelligence
商业分发Distribution优量汇 · 穿山甲 · 快手
产品与模型Products & ModelsYaoledge · Yaoedge ·
Ember · Nova