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，
到 Model Gateway、AI 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 / HIGH-BANDWIDTH PUBLIC EDGE

25Gbps 公网入口，按流量横向扩展。

25Gbps-class ingress.Scale out with traffic.

每个 Edge 作为精简、可替换的公网数据平面；
随着流量增长直接增加 Edge 实例，而 Controller 保持业务权威。
连接通过 QUIC v1、TLS 1.3 与固定 Ed25519 身份建立。

Each Edge is a thin, replaceable public data plane. Add Edge instances as traffic grows while Controller retains business authority, connected over QUIC v1, TLS 1.3 and fixed Ed25519 identity.

25Gbps-class unitreplaceable edgecontroller authority
25GEdge / 01TLS ingress · byte streamactive
25GEdge / 02TLS ingress · byte streamactive
25GEdge / 03TLS ingress · byte streamscale
25GEdge / Nreplaceable data plane+ edge
+Edge = + ingress capacity
AUTHORITY ENDPOINT

Controller

Identity
fixed Ed25519
Transport
QUIC v1 · TLS 1.3
Durable state
not stored at Edge
Control
Controller initiated
Failure domain · Edge keeps no Service / DNS / ACME / persistent TLS authority
02 / INDEPENDENT SCALE DOMAINS

每一种资源，都有自己的扩展边界。

Scale by role,not by cluster boundary.

Controller、Execution Node、Model Gateway、Data Gateway 与 Storage
按照各自压力单独增加；机器规模增长不要求重定义整体系统拓扑。

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

NO PRE-FIXED CLUSTER SCALE BOUNDARY只扩展出现压力的职责域；示意容量，不构成 benchmark。Only the pressured role expands. Capacity is illustrative, not a benchmark.
AUTHORITATIVE STATEController / Product Statereplaceable replicas
explicit ownership
Controllercontrol pressure
steady
Execution NodeYaoBox / Task
+3NODE
Model Gatewaymodel traffic
+2GATEWAY
Data Gatewayblob / asset traffic
steady
Storagebackend capacity
+1BACKEND
Architecture goal · horizontal scaling by responsibility, not one global scheduler boundary
03 / PURPOSE-BUILT AI RUNTIME

更少的层，更明确的执行边界。

Fewer layers.Explicit execution boundaries.

YaoBox 不以 Pod、Service Mesh 或通用 Scheduler 作为核心抽象，
而直接组合 Linux 的进程、身份、网络、文件系统、资源和配额边界，
围绕 AI Task 定义执行生命周期与故障语义。

YaoBox is built directly from Linux process, identity, network, filesystem, resource and quota boundaries around the lifecycle and failure semantics of an AI Task.

fewer layersexplicit ownershipno inbound node rpc
PROCESS BOUNDARYPID namespace · crun
IDENTITY / RESOURCEuser namespace · cgroup v2
NETWORK BOUNDARYnetwork namespace · veth
FILESYSTEM / QUOTAEROFS · OverlayFS · XFS project quota
YaoBoxTask execution envelope
Task
NodeSession scopedpull-based commands
ProcessPID namespace · crun
Identityuser namespace · cgroup v2
Networknetwork namespace · veth
FilesystemEROFS · OverlayFS · quota
04 / INTELLIGENCE PLANE

执行计算与模型智能，可以独立演进。

Compute and intelligencescale independently.

Execution Node 只提交绑定当前 Task 的标准 Model Request；
Model Gateway 负责 staging、routing、Provider Credential、Provider effect
与流式响应标准化，形成统一的 Intelligence Plane。

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

PROVIDER CREDENTIAL TERMINATES AT GATEWAYNode 不持有真实 Provider Endpoint 或 Credential；模型容量与执行容量独立扩展。Nodes never hold real provider endpoints or credentials; model and execution 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
Credential Boundary
Stage / Routemodel selection · request staging
Provider Effectauthorized external execution
Normalizestandard response semantics
Streamnormalized streaming output
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。

Failure is explicit.Recovery is deterministic.

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；
一旦请求可能到达 Provider，就禁止第二 Gateway 自动重放。

External effects are not hidden behind automatic retry. The system establishes unique ownership first, performs the effect, then resolves it explicitly; once a request may have reached a provider, automatic replay by a second gateway is forbidden.

failure explicitbounded authoritydeterministic recovery
AUTHORITATIVE
STATE
POSSIBLE请求已持久化 · 唯一 identitypersisted request · unique identity
GatewayInstance绑定唯一 product ownersingle product owner
RESOLVED确定状态已提交deterministic state committed
UNCERTAIN EFFECT WINDOW
EXTERNAL
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
AT-MOST-ONCE PROVIDER EFFECT
×SECOND GATEWAY
AUTO-REPLAY FORBIDDEN
Controller outage仅在既有授权与单调时限内继续；不产生新的 runtime extension 或 model grant。continue only within existing grants and monotonic deadlines; no new runtime extension or model grant.
06 / MODEL-SYSTEM CO-DESIGN

模型与系统，作为同一个工程问题。

Model and system,engineered as one.

模型结构、数值体系、GPU 通信、长上下文执行与训练 / rollout policy identity
被共同定义，让 Research 直接连接最终 Runtime 与 Serving。

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

ONE ENGINEERING SURFACE
Architecture ↔ Numerics ↔ Runtime ↔ Serving
模型架构Architecturestructure / attention / state
Granular MoEGated DeltaNetHierarchical Sparse Attention
数值与训练Numericsprecision / optimization
NVFP4-firstMXFP8 AttentionMuonGPU Communication
Runtimelong context / rollout
StateFlowLong ContextAgent / RLVR
Servingroute / stream / models
Model GatewayEmberNovaprovider abstraction
07 / SYSTEMS FOUNDATION

让技术真正进入大规模商业产品。

A systems foundationfor real products.

Infrastructure、Commerce、Distribution 与产品形态连接到同一技术底座。
Yaoledge、Yaoedge、Ember、Nova 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 the same systems stack.

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
FoundationNetwork · 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
INFRA OVH / ISIF / AÉZACOMMERCE ALIPAYDISTRIBUTION GDT / PANGLE / KUAISHOU