從全域計算方法論到 AI 原生計算編排
GCM Reference Runtime、計算配置分配演算法、Native Compute Fabric 與 HDUS 介接總體技術白皮書
版本: v0.1日期: 2026-08-25文件性質: 統合技術白皮書/下一階段工程架構文狀態: Canonical Draft上游來源: GCM Series-00、Paper 01–06、TW-01、TW-02、TW-03、GCM Reference Runtime MVP v0.1
摘要
Global Computation Methodology(GCM)已由最初的 24/72 計算配置結構,逐步發展為一套處理異質計算、全域一致性、動態配置路由、觀察與物化分離、有限活動實現以及可追溯歷史的形式方法論,並已進一步形成 Canonical Specification、Reference Runtime Architecture、Conformance Specification 與可執行 Reference Runtime。
因此,下一階段的核心問題已經不再只是:
「什麼是 Global Computation?」
而是:
一個 AI、原生計算機或未來 World-native 作業系統,如何在 GCM 的語義、權限、資源與驗證約束內,自主判斷應採取何種計算、使用何種計算配置、分配多少資源、在哪裡執行、何時切換方法,以及什麼時候結果具有足夠證據可以提交?
本白皮書提出以下長期演進架構:
GCM Semantic Runtime → Compute Orchestrator / Allocator → Native Compute Fabric → HDUS Adapter \boxed{
\text{GCM Semantic Runtime}
\rightarrow
\text{Compute Orchestrator / Allocator}
\rightarrow
\text{Native Compute Fabric}
\rightarrow
\text{HDUS Adapter}
} GCM Semantic Runtime → Compute Orchestrator / Allocator → Native Compute Fabric → HDUS Adapter
其工程順序則固定為:
Reference MVP 100% → Deterministic Compute Allocation → AI-assisted Orchestration → Native Resource Fabric → HDUS Integration . \boxed{
\text{Reference MVP 100\%}
\rightarrow
\text{Deterministic Compute Allocation}
\rightarrow
\text{AI-assisted Orchestration}
\rightarrow
\text{Native Resource Fabric}
\rightarrow
\text{HDUS Integration}.
} Reference MVP 100% → Deterministic Compute Allocation → AI-assisted Orchestration → Native Resource Fabric → HDUS Integration .
第一階段先把現有 GCM Reference Runtime 補足為陌生工程師可以獨立安裝、執行、擴張與驗證的 Developer-Usable Reference Runtime;第二階段建立 deterministic、constraint-first 的計算配置與資源分配演算法;第三階段讓 AI 成為 proposal / planning layer,在 deterministic admissibility、authority、budget 與 verification gates 內進行自主計算編排;第四階段介接 CPU、GPU、NPU、memory、storage、network、cluster 與其他計算載體;最後才透過明確 adapter 對接未來 HDUS。
本文件不是新的第七篇 GCM 核心論文,也不重新定義 00–06。它是從既有形式系統走向 AI-native computational operating layer 的工程交接總圖。
1. GCM 已完成的基礎
目前 GCM 的配置基礎為:
P 24 = B 2 × U 4 × I 3 , \mathfrak P_{24}
=
\mathfrak B_2
\times
\mathfrak U_4
\times
\mathfrak I_3, P 24 = B 2 × U 4 × I 3 ,
以及:
P 72 = P 24 × L 3 . \mathfrak P_{72}
=
\mathfrak P_{24}
\times
\mathfrak L_3. P 72 = P 24 × L 3 .
其中:
P 24 , P 72 = Extensible Computational Configuration Basis \boxed{
\mathfrak P_{24},\mathfrak P_{72}
=
\text{Extensible Computational Configuration Basis}
} P 24 , P 72 = Extensible Computational Configuration Basis
而不是計算世界的封閉窮舉。
GCM 的上層定義仍然是:
Global Computation = Globally Coherent Heterogeneous Computation . \boxed{
\text{Global Computation}
=
\text{Globally Coherent Heterogeneous Computation}.
} Global Computation = Globally Coherent Heterogeneous Computation .
經過 Paper 01–06、TW-01–TW-03 與 Reference Runtime,以下邊界已經成為後續系統不得破壞的基本 invariants:
M G ≠ W \boxed{
\mathcal M_G\neq\mathbf W
} M G = W
Computation ≠ Observation ≠ Materialization \boxed{
\text{Computation}
\neq
\text{Observation}
\neq
\text{Materialization}
} Computation = Observation = Materialization
Can Execute ≠ May Execute \boxed{
\text{Can Execute}
\neq
\text{May Execute}
} Can Execute = May Execute
Executor Success ⇏ Global Commit \boxed{
\text{Executor Success}
\not\Rightarrow
\text{Global Commit}
} Executor Success ⇒ Global Commit
Global Dependency ≠ Full Materialization \boxed{
\text{Global Dependency}
\neq
\text{Full Materialization}
} Global Dependency = Full Materialization
Finite Active Realization + Unbounded Extensibility \boxed{
\text{Finite Active Realization}
+
\text{Unbounded Extensibility}
} Finite Active Realization + Unbounded Extensibility
State Equality ⇏ History Equality . \boxed{
\text{State Equality}
\not\Rightarrow
\text{History Equality}.
} State Equality ⇒ History Equality .
因此,「AI 自主管理計算能力」並不是在 GCM 上加入一個任意 scheduler,而是把 AI 自主性放入這組已經建立的 invariants 內。
2. 為什麼下一階段不只是 AI Scheduler
傳統 scheduler 通常面對的是:
已經存在一組 runnable tasks,如何把它們安排到 processor、node 或 time slice?
但 GCM 下一階段所處理的問題更早。
對一個目標 G G G ,系統首先可能需要判斷:
如何分解 computational domains? 每個 domain 使用何種 computational configuration? 需要哪一種 transition law? 需要哪些 representation bridges? 哪些資料需要 materialize? 哪些區域只需 dormant / coarse realization? 需要多少 compute resolution? 哪些 hardware resources 值得投入? 哪些結果必須增加 verification? \boxed{
\begin{aligned}
&\text{如何分解 computational domains?}\\
&\text{每個 domain 使用何種 computational configuration?}\\
&\text{需要哪一種 transition law?}\\
&\text{需要哪些 representation bridges?}\\
&\text{哪些資料需要 materialize?}\\
&\text{哪些區域只需 dormant / coarse realization?}\\
&\text{需要多少 compute resolution?}\\
&\text{哪些 hardware resources 值得投入?}\\
&\text{哪些結果必須增加 verification?}
\end{aligned}
} 如何分解 computational domains ? 每個 domain 使用何種 computational configuration ? 需要哪一種 transition law ? 需要哪些 representation bridges ? 哪些資料需要 materialize ? 哪些區域只需 dormant / coarse realization ? 需要多少 compute resolution ? 哪些 hardware resources 值得投入? 哪些結果必須增加 verification ?
最後才進入:
CPU、GPU、NPU 或哪一個 node 執行?
因此:
GCM Compute Orchestration ≠ OS Scheduling . \boxed{
\text{GCM Compute Orchestration}
\neq
\text{OS Scheduling}.
} GCM Compute Orchestration = OS Scheduling .
OS scheduler 是未來 GCM 可以驅動或提供 hint 的低層 actuator,但不是 GCM 本身。
同樣:
GCM Compute Orchestration ≠ Compiler Auto-Tuning . \boxed{
\text{GCM Compute Orchestration}
\neq
\text{Compiler Auto-Tuning}.
} GCM Compute Orchestration = Compiler Auto-Tuning .
TVM MetaSchedule 已經可以在 tensor program 的 schedule design space 中,用 cost model、hardware measurement 與 task scheduler 搜尋高效 implementation;Ansor 亦已研究如何自動生成和搜尋大型 tensor-program search spaces。
這些技術未來可以成為 GCM 的 executor 或 optimization backend。
GCM 所處的位置更上層:
Which computation → Which configuration → Which implementation → Which resources . \boxed{
\text{Which computation}
\rightarrow
\text{Which configuration}
\rightarrow
\text{Which implementation}
\rightarrow
\text{Which resources}.
} Which computation → Which configuration → Which implementation → Which resources .
3. 四層總體架構
3.1 Layer 0:GCM Semantic Runtime
這就是目前已完成的 Reference Runtime 核心。
其 canonical World mutation pipeline 為:
Proposal → Reconcile → Verify → Commit / Reject . \boxed{
\text{Proposal}
\rightarrow
\text{Reconcile}
\rightarrow
\text{Verify}
\rightarrow
\text{Commit / Reject}.
} Proposal → Reconcile → Verify → Commit / Reject .
Layer 0 負責:
World / Runtime / Observer / Foundation separation;
typed operation;
reachability;
admissibility;
authority;
representation bridge;
proposal isolation;
reconciliation;
verification;
commit;
rollback / compensation;
materialization;
active support;
typed history。
其最高優先原則仍然是:
Semantic Correctness ≻ Performance Optimization . \boxed{
\text{Semantic Correctness}
\succ
\text{Performance Optimization}.
} Semantic Correctness ≻ Performance Optimization .
也就是後面的 AI、cost model 或 optimization algorithm 都不能反過來修改這一層的合法性定義。
3.2 Layer 1:Compute Orchestrator / Allocator
下一階段真正新增的是這一層。
定義一個 orchestration problem:
Q = ⟨ G , B W , D , P , E , B , R , C , A , H , B g t ⟩ . \mathcal Q
=
\left\langle
G,
B_W,
\mathcal D,
\mathfrak P,
\mathcal E,
\mathcal B,
\mathcal R,
\mathcal C,
\mathcal A,
\mathcal H,
\mathcal Bgt
\right\rangle. Q = ⟨ G , B W , D , P , E , B , R , C , A , H , B g t ⟩ .
其中:
G = Goal / Intent , G
=
\text{Goal / Intent}, G = Goal / Intent ,
B W = World Boundary , B_W
=
\text{World Boundary}, B W = World Boundary ,
D = Candidate Domain Decompositions , \mathcal D
=
\text{Candidate Domain Decompositions}, D = Candidate Domain Decompositions ,
P = Computational Configuration Space , \mathfrak P
=
\text{Computational Configuration Space}, P = Computational Configuration Space ,
E = Executor Registry , \mathcal E
=
\text{Executor Registry}, E = Executor Registry ,
B = Bridge Registry , \mathcal B
=
\text{Bridge Registry}, B = Bridge Registry ,
R = Resource State , \mathcal R
=
\text{Resource State}, R = Resource State ,
C = Constraints / Invariants , \mathcal C
=
\text{Constraints / Invariants}, C = Constraints / Invariants ,
A = Authority Context , \mathcal A
=
\text{Authority Context}, A = Authority Context ,
H = Historical Evidence , \mathcal H
=
\text{Historical Evidence}, H = Historical Evidence ,
B g t = Budget Envelope . \mathcal Bgt
=
\text{Budget Envelope}. B g t = Budget Envelope .
Orchestrator 產生:
Ω = ⟨ Δ D , Γ , E , B r , A R , σ , ρ C , M , V , K ⟩ . \Omega
=
\left\langle
\Delta_D,
\Gamma,
E,
Br,
A_R,
\sigma,
\rho^C,
M,
V,
K
\right\rangle. Ω = ⟨ Δ D , Γ , E , B r , A R , σ , ρ C , M , V , K ⟩ .
其中:
Δ D = Domain Decomposition / Refinement , \Delta_D
=
\text{Domain Decomposition / Refinement}, Δ D = Domain Decomposition / Refinement ,
Γ = Configuration Assignment , \Gamma
=
\text{Configuration Assignment}, Γ = Configuration Assignment ,
E = Executor Binding , E
=
\text{Executor Binding}, E = Executor Binding ,
B r = Representation Bridge Plan , Br
=
\text{Representation Bridge Plan}, B r = Representation Bridge Plan ,
A R = Resource Allocation , A_R
=
\text{Resource Allocation}, A R = Resource Allocation ,
σ = Execution / Scheduling Relation , \sigma
=
\text{Execution / Scheduling Relation}, σ = Execution / Scheduling Relation ,
ρ C = Compute Resolution , \rho^C
=
\text{Compute Resolution}, ρ C = Compute Resolution ,
M = Materialization / Active-Support Plan , M
=
\text{Materialization / Active-Support Plan}, M = Materialization / Active-Support Plan ,
V = Verification Strategy , V
=
\text{Verification Strategy}, V = Verification Strategy ,
K = Recovery / Compensation Contract . K
=
\text{Recovery / Compensation Contract}. K = Recovery / Compensation Contract .
但是:
Ω ≠ Execution Authority . \boxed{
\Omega
\neq
\text{Execution Authority}.
} Ω = Execution Authority .
Orchestrator 的輸出首先是一個 plan proposal。
3.3 Layer 2:Native Compute Fabric
這一層將 GCM 的 logical compute demand 映射到實際計算載體。
定義:
R e s o u r c e A d a p t e r : L o g i c a l D e m a n d ⇀ P h y s i c a l B i n d i n g . \mathsf{ResourceAdapter}:
\mathsf{LogicalDemand}
\rightharpoonup
\mathsf{PhysicalBinding}. ResourceAdapter : LogicalDemand ⇀ PhysicalBinding .
可能包含:
CPU / NUMA , GPU / VRAM , NPU / TPU / accelerator , system memory , unified memory , NVMe / storage , network , remote nodes , cloud executors . \begin{aligned}
&\text{CPU / NUMA},\\
&\text{GPU / VRAM},\\
&\text{NPU / TPU / accelerator},\\
&\text{system memory},\\
&\text{unified memory},\\
&\text{NVMe / storage},\\
&\text{network},\\
&\text{remote nodes},\\
&\text{cloud executors}.
\end{aligned} CPU / NUMA , GPU / VRAM , NPU / TPU / accelerator , system memory , unified memory , NVMe / storage , network , remote nodes , cloud executors .
Ray 已提供 CPU、GPU、memory 與 custom logical resources 的 resource-aware scheduling;其文件也特別指出 logical resource request 並不等於 physical CPU isolation。
所以 GCM 必須明確保持:
Logical Resource Claim ≠ Physical Enforcement . \boxed{
\text{Logical Resource Claim}
\neq
\text{Physical Enforcement}.
} Logical Resource Claim = Physical Enforcement .
Resource Adapter 最低應回報:
R b i n d = ⟨ R e q u e s t e d , G r a n t e d , E n f o r c e d , O b s e r v e d , I s o l a t i o n , T o p o l o g y , F a i l u r e ⟩ . \mathfrak R_{\mathrm{bind}}
=
\left\langle
\mathsf{Requested},
\mathsf{Granted},
\mathsf{Enforced},
\mathsf{Observed},
\mathsf{Isolation},
\mathsf{Topology},
\mathsf{Failure}
\right\rangle. R bind = ⟨ Requested , Granted , Enforced , Observed , Isolation , Topology , Failure ⟩ .
3.4 Layer 3:HDUS Adapter
HDUS 與 GCM 不應互相吞併。
應保持:
HDUS Semantics ≠ GCM Semantics . \boxed{
\text{HDUS Semantics}
\neq
\text{GCM Semantics}.
} HDUS Semantics = GCM Semantics .
兩者之間以 typed adapter 對接:
HDUS ↔ Typed Adapter GCM . \boxed{
\text{HDUS}
\xleftrightarrow{\text{Typed Adapter}}
\text{GCM}.
} HDUS Typed Adapter GCM .
例如:
H D U S I n t e n t → G C M O p e r a t i o n R e q u e s t , \mathsf{HDUSIntent}
\rightarrow
\mathsf{GCMOperationRequest}, HDUSIntent → GCMOperationRequest ,
H D U S W o r l d B o u n d a r y → B W , \mathsf{HDUSWorldBoundary}
\rightarrow
B_W, HDUSWorldBoundary → B W ,
H D U S S u b j e c t A u t h o r i t y → A u t h o r i t y C o n t e x t , \mathsf{HDUSSubjectAuthority}
\rightarrow
\mathsf{AuthorityContext}, HDUSSubjectAuthority → AuthorityContext ,
H D U S R e s o u r c e F a b r i c → R e s o u r c e A d a p t e r R e g i s t r y . \mathsf{HDUSResourceFabric}
\rightarrow
\mathsf{ResourceAdapterRegistry}. HDUSResourceFabric → ResourceAdapterRegistry .
因此不允許簡化為:
HDUS Process Success = GCM World Commit . \boxed{
\text{HDUS Process Success}
=
\text{GCM World Commit}.
} HDUS Process Success = GCM World Commit .
同樣:
HDUS Subject ≠ GCM Executor . \boxed{
\text{HDUS Subject}
\neq
\text{GCM Executor}.
} HDUS Subject = GCM Executor .
HDUS 可以成為未來最深度的 native integration target,但 GCM 必須保持可獨立存在。
4. Reference Runtime MVP 的 100% 定義
目前的 Reference Runtime 已完成 semantic kernel 與 R0-M4 conformance。
但 Developer-Usable Reference MVP 應再完成五個區塊。
4.1 Durable Core
最低要求:
Process Restart ⇏ Semantic Reset . \boxed{
\text{Process Restart}
\not\Rightarrow
\text{Semantic Reset}.
} Process Restart ⇒ Semantic Reset .
需要:
SQLite 或同級 local durable backend;
durable World head;
Foundation version persistence;
append-only provenance / evidence;
canonical serialization;
recoverable transaction marker;
restart recovery;
durable history index。
4.2 Demo UX
最低提供:
gcm demo heterogeneous
gcm demo authority
gcm demo observer
gcm demo bounded-world
gcm demo history
gcm conformance
每個 demo 都必須展示 GCM invariant,而不是只有 happy-path output。
4.3 Plugin SDK
第一版 plugin surface:
ExecutorPlugin
BridgePlugin
ResourceAdapter
CostEstimator
VerificationPlugin
但:
Plugin Capability ⇏ Plugin Authority . \boxed{
\text{Plugin Capability}
\not\Rightarrow
\text{Plugin Authority}.
} Plugin Capability ⇒ Plugin Authority .
Plugin 可以宣告:
我會做什麼。
不能因此宣告:
所以我有權修改 World。
4.4 Recovery Matrix
至少覆蓋:
before execution , after execution / before proposal persist , after proposal / before verification , after verification / before World write , after World write / before receipt , after receipt / before acknowledgement , during archive / restore , during compensation , during Foundation revision . \begin{aligned}
&\text{before execution},\\
&\text{after execution / before proposal persist},\\
&\text{after proposal / before verification},\\
&\text{after verification / before World write},\\
&\text{after World write / before receipt},\\
&\text{after receipt / before acknowledgement},\\
&\text{during archive / restore},\\
&\text{during compensation},\\
&\text{during Foundation revision}.
\end{aligned} before execution , after execution / before proposal persist , after proposal / before verification , after verification / before World write , after World write / before receipt , after receipt / before acknowledgement , during archive / restore , during compensation , during Foundation revision .
目標不是所有 fault 都能「神奇恢復」,而是:
Fault ⇏ Unknown Silent State . \boxed{
\text{Fault}
\not\Rightarrow
\text{Unknown Silent State}.
} Fault ⇒ Unknown Silent State .
4.5 External Developer Packaging
一個沒有讀過 GCM 論文的工程師應可:
Install → Run → Inspect → Extend → Conform . \boxed{
\text{Install}
\rightarrow
\text{Run}
\rightarrow
\text{Inspect}
\rightarrow
\text{Extend}
\rightarrow
\text{Conform}.
} Install → Run → Inspect → Extend → Conform .
具體而言:
安裝 package;
跑五組 demos;
執行 conformance;
新增一個簡單 executor;
查看 proposal / receipt / history;
理解 route 為什麼被 reject;
不需要先讀完 00–06。
完成以上五項後,才正式封板:
GCM Reference Runtime MVP v0.1 = 100 % . \boxed{
\text{GCM Reference Runtime MVP v0.1}
=
100\%.
} GCM Reference Runtime MVP v0.1 = 100%.
5. 第一版 Compute Allocation Algorithm
Reference MVP 完成後,下一個正式工程不是先加入 LLM,而是:
G C M A l l o c a t o r 0 = Deterministic Constraint-First Allocator . \boxed{
\mathsf{GCMAllocator}_{0}
=
\text{Deterministic Constraint-First Allocator}.
} GCMAllocator 0 = Deterministic Constraint-First Allocator .
5.1 Safe Set
所有候選 orchestration plans 為:
Ω . \mathbb \Omega. Ω.
先建立:
Ω s a f e = { Ω ∈ Ω ∣ R e a c h a b l e ∧ A d m i s s i b l e ∧ A u t h o r i z e d ∧ B r i d g e V a l i d ∧ B u d g e t F e a s i b l e ∧ I n v a r i a n t P r e s e r v i n g } . \mathbb \Omega_{\mathrm{safe}}
=
\left\{
\Omega\in\mathbb \Omega
\mid
\mathsf{Reachable}
\land
\mathsf{Admissible}
\land
\mathsf{Authorized}
\land
\mathsf{BridgeValid}
\land
\mathsf{BudgetFeasible}
\land
\mathsf{InvariantPreserving}
\right\}. Ω safe = { Ω ∈ Ω ∣ Reachable ∧ Admissible ∧ Authorized ∧ BridgeValid ∧ BudgetFeasible ∧ InvariantPreserving } .
只有:
Ω ∈ Ω s a f e \Omega\in\mathbb \Omega_{\mathrm{safe}} Ω ∈ Ω safe
才能進入 optimization。
所以:
Admissibility First → Optimization Second . \boxed{
\text{Admissibility First}
\rightarrow
\text{Optimization Second}.
} Admissibility First → Optimization Second .
5.2 Multi-objective Resource Evaluation
對 plan Ω \Omega Ω 定義:
J ( Ω ) = [ T , C , E , M , D , R , ϵ , H , X ] . \mathbf J(\Omega)
=
\left[
T,
C,
E,
M,
D,
R,
\epsilon,
H,
X
\right]. J ( Ω ) = [ T , C , E , M , D , R , ϵ , H , X ] .
可分別表示:
T = latency , T=\text{latency}, T = latency ,
C = compute / financial cost , C=\text{compute / financial cost}, C = compute / financial cost ,
E = energy , E=\text{energy}, E = energy ,
M = memory pressure , M=\text{memory pressure}, M = memory pressure ,
D = data movement , D=\text{data movement}, D = data movement ,
R = operational risk , R=\text{operational risk}, R = operational risk ,
ϵ = approximation error , \epsilon=\text{approximation error}, ϵ = approximation error ,
H = history / verification burden , H=\text{history / verification burden}, H = history / verification burden ,
X = configuration switching / bridge cost . X=\text{configuration switching / bridge cost}. X = configuration switching / bridge cost .
預設不要求全部壓成:
J ( Ω ) ∈ R . J(\Omega)\in\mathbb R. J ( Ω ) ∈ R .
可先使用:
Pareto-safe Selection \boxed{
\text{Pareto-safe Selection}
} Pareto-safe Selection
再由 policy 指定 ranking。
因此:
Mathematics ≠ Verification ≠ Optimization . \boxed{
\text{Mathematics}
\neq
\text{Verification}
\neq
\text{Optimization}.
} Mathematics = Verification = Optimization .
6. AI 的正確位置
第一版 AI 不應直接取代 deterministic allocator。
AI 應產生:
Ω 1 A I , Ω 2 A I , … , Ω n A I . \Omega^{AI}_1,
\Omega^{AI}_2,\ldots,\Omega^{AI}_n. Ω 1 A I , Ω 2 A I , … , Ω n A I .
而 deterministic core 執行:
V a l i d a t e ( Ω k A I ) . \mathsf{Validate}
\left(
\Omega^{AI}_k
\right). Validate ( Ω k A I ) .
只有:
Ω k A I ∈ Ω s a f e \Omega^{AI}_k
\in
\mathbb\Omega_{\mathrm{safe}} Ω k A I ∈ Ω safe
的計畫才可以執行。
因此:
AI Planning Authority ≠ Execution Authority ≠ Commit Authority . \boxed{
\text{AI Planning Authority}
\neq
\text{Execution Authority}
\neq
\text{Commit Authority}.
} AI Planning Authority = Execution Authority = Commit Authority .
7. AI 自主計算編排階梯
Level 0 — Deterministic
使用:
capability matching;
fixed constraints;
deterministic ranking。
Level 1 — Heuristic
加入:
known hardware preference;
locality;
cache;
estimated execution cost。
Level 2 — Learned Cost Model
從歷史 receipt 與 metrics 學習:
J ^ ( Ω ) . \widehat{\mathbf J}(\Omega). J ( Ω ) .
但 learned model 只能:
排 safe candidates。
不能:
把 unsafe candidate 變成 safe。
TVM MetaSchedule 已展示以 cost model、實際 hardware measurement 與 tuning database 驅動 search 的成熟工程方式,可作為此層重要參照。
Level 3 — AI Proposal Planner
AI 可以提出:
domain decomposition;
configuration;
executor;
bridge;
resource;
materialization;
verification strategy。
Level 4 — Bounded Autonomous Orchestrator
AI 在:
⟨ B W , A , B g t , C ⟩ \left\langle
B_W,
\mathcal A,
\mathcal Bgt,
\mathcal C
\right\rangle ⟨ B W , A , B g t , C ⟩
約束內自動重規劃。
Level 5 — Continual Self-Tuning
根據 history 自動調整:
cost model;
routing preference;
resource policy;
exploration policy。
即使到了 Level 5:
Foundation Auto-Revision = Forbidden by Default . \boxed{
\text{Foundation Auto-Revision}
=
\text{Forbidden by Default}.
} Foundation Auto-Revision = Forbidden by Default .
8. Closed-loop Compute Orchestration
真正的 AI-native compute runtime 不應只做一次 allocation。
其控制迴路為:
Observe → Plan → Validate → Allocate → Execute → Measure → Replan . \boxed{
\text{Observe}
\rightarrow
\text{Plan}
\rightarrow
\text{Validate}
\rightarrow
\text{Allocate}
\rightarrow
\text{Execute}
\rightarrow
\text{Measure}
\rightarrow
\text{Replan}.
} Observe → Plan → Validate → Allocate → Execute → Measure → Replan .
但 World mutation 仍然是另一條:
Proposal → Reconcile → Verify → Commit . \boxed{
\text{Proposal}
\rightarrow
\text{Reconcile}
\rightarrow
\text{Verify}
\rightarrow
\text{Commit}.
} Proposal → Reconcile → Verify → Commit .
所以:
Adaptive Resource Reallocation ⇏ Unverified World Mutation . \boxed{
\text{Adaptive Resource Reallocation}
\not\Rightarrow
\text{Unverified World Mutation}.
} Adaptive Resource Reallocation ⇒ Unverified World Mutation .
9. History 從 audit 變成 allocator evidence
Paper-06 所建立的 typed history 未來可以作為 allocator 的 empirical evidence base。
定義:
E v i d e n c e ( Ω ) = ⟨ R u n t i m e , C o s t , E r r o r , F a i l u r e , R e c o v e r y , H a r d w a r e , W o r l d C o n t e x t ⟩ . \mathsf{Evidence}(\Omega)
=
\left\langle
\mathsf{Runtime},
\mathsf{Cost},
\mathsf{Error},
\mathsf{Failure},
\mathsf{Recovery},
\mathsf{Hardware},
\mathsf{WorldContext}
\right\rangle. Evidence ( Ω ) = ⟨ Runtime , Cost , Error , Failure , Recovery , Hardware , WorldContext ⟩ .
Allocator 可以學習:
某類 domain 適合哪種 configuration;
某 executor 在某硬體上的實際性能;
某 bridge 的 loss;
哪種 resolution 已經足夠;
哪些任務值得 GPU;
哪些任務 CPU 反而更快;
哪些 route 雖然快但 verification burden 過大。
但:
Historical Performance ⇏ Semantic Admissibility . \boxed{
\text{Historical Performance}
\not\Rightarrow
\text{Semantic Admissibility}.
} Historical Performance ⇒ Semantic Admissibility .
10. 與既有技術的公平定位
StarPU 已經是一套成熟的 heterogeneous runtime,可以處理 task dependencies、CPU/GPU execution、data movement、asynchronous execution,並使用 performance models 進行 processor-aware scheduling。
Legion 已經透過 logical regions、privileges 與 coherence 建立 correctness constraints,也明確讓 mapper 負責硬體 mapping,而 mapping decision 原則上不應改變程式 correctness。
Ray 已處理 distributed CPU/GPU/custom logical resource scheduling。
TVM MetaSchedule 與 Ansor 已處理 automatic search space exploration、cost model、measurement 與 tuning budget。
MLIR Transform Dialect 已經提供 compiler transformation orchestration surface。
Linux sched_ext 則已允許使用 BPF 實作 extensible scheduler,並在 scheduler failure 時退回 kernel default scheduling behavior。
截至 2026 年,Kubernetes Dynamic Resource Allocation 已經成為 stable resource-allocation mechanism,並持續擴展 accelerator、CPU、memory 與更多 device/resource classes。
因此 GCM 不宣稱發明上述技術。
GCM 下一階段真正研究的是:
World-relative computational planning + extensible computational configuration selection + authority-preserving orchestration + verification-bound adaptation + observer/materialization/history-aware resource use . \boxed{
\begin{aligned}
&\text{World-relative computational planning}\\
+&\text{extensible computational configuration selection}\\
+&\text{authority-preserving orchestration}\\
+&\text{verification-bound adaptation}\\
+&\text{observer/materialization/history-aware resource use}.
\end{aligned}
} + + + + World-relative computational planning extensible computational configuration selection authority-preserving orchestration verification-bound adaptation observer/materialization/history-aware resource use .
11. Native Compute Fabric
Native Compute Fabric 應作為 GCM 和底層系統之間的 adapter layer。
第一版本可只支援:
CPU + Memory + Local GPU + Storage + Executor Concurrency . \boxed{
\text{CPU}
+
\text{Memory}
+
\text{Local GPU}
+
\text{Storage}
+
\text{Executor Concurrency}.
} CPU + Memory + Local GPU + Storage + Executor Concurrency .
後續再擴張:
NUMA , NPU , unified memory , remote GPU , cluster , network-aware allocation , cloud , energy-aware execution . \begin{aligned}
&\text{NUMA},\\
&\text{NPU},\\
&\text{unified memory},\\
&\text{remote GPU},\\
&\text{cluster},\\
&\text{network-aware allocation},\\
&\text{cloud},\\
&\text{energy-aware execution}.
\end{aligned} NUMA , NPU , unified memory , remote GPU , cluster , network-aware allocation , cloud , energy-aware execution .
Linux 上未來可考慮:
cgroups;
affinity;
NUMA APIs;
accelerator APIs;
sched_ext。
但:
GCM Plan ≠ Kernel Scheduling Decision . \boxed{
\text{GCM Plan}
\neq
\text{Kernel Scheduling Decision}.
} GCM Plan = Kernel Scheduling Decision .
GCM 可以給低層 scheduler policy / hint / resource contract,但不必取代 kernel scheduler。
12. HDUS 的長期位置
HDUS 未來可以成為:
World-native OS / Runtime \boxed{
\text{World-native OS / Runtime}
} World-native OS / Runtime
而 GCM 成為:
Computational Semantics / Orchestration Layer . \boxed{
\text{Computational Semantics / Orchestration Layer}.
} Computational Semantics / Orchestration Layer .
所以更合理的關係是:
Human / AI / HDUS Subject
↓
Intent / World Operation
↓
GCM Compute Orchestrator
↓
GCM Semantic Validation
↓
Native Compute Fabric
↓
HDUS / Linux / Hardware Runtime
而不是:
HDUS = GCM
也不是:
GCM = HDUS Scheduler
兩套系統應各自保留 canonical semantics,再透過 adapter 合成。
13. 正式工程 Roadmap
Phase A|Reference Runtime MVP 100%
A1 — Durable Core
完成 local persistence、restart recovery 與 durable provenance。
A2 — Demo UX
完成:
heterogeneous
authority
observer
bounded-world
history
conformance
六條 developer-facing path。
A3 — Plugin SDK
完成第一版:
Executor
Bridge
ResourceAdapter
Estimator
Verifier
plugin surface。
A4 — Recovery Matrix
將 Reference Runtime 的 crash/recovery coverage 從單一重點 case 擴張到關鍵 fault boundaries。
A5 — Developer Release
完成 package、installation、tutorial、extension example、evidence inspection 與 release validation。
達成:
GCM Reference Runtime MVP v0.1 = 100 % . \boxed{
\text{GCM Reference Runtime MVP v0.1}
=
100\%.
} GCM Reference Runtime MVP v0.1 = 100%.
Phase B|Deterministic Compute Allocator
建立:
G C M A l l o c a t o r 0 . \boxed{
\mathsf{GCMAllocator}_{0}.
} GCMAllocator 0 .
流程:
Goal → Candidate Decomposition → Safe Set → Configuration Selection → Resource Allocation → Pareto Ranking → Plan Proposal . \boxed{
\text{Goal}
\rightarrow
\text{Candidate Decomposition}
\rightarrow
\text{Safe Set}
\rightarrow
\text{Configuration Selection}
\rightarrow
\text{Resource Allocation}
\rightarrow
\text{Pareto Ranking}
\rightarrow
\text{Plan Proposal}.
} Goal → Candidate Decomposition → Safe Set → Configuration Selection → Resource Allocation → Pareto Ranking → Plan Proposal .
Phase C|AI-assisted Compute Orchestrator
加入:
AI domain decomposition;
route generation;
counterfactual plans;
learned cost model;
adaptive replanning;
evidence-driven self-tuning。
但 deterministic gates 保留。
Phase D|Native Compute Fabric
先 local machine,再 heterogeneous hardware,再 distributed systems。
Phase E|HDUS Integration
完成:
H D U S G C M A d a p t e r . \boxed{
\mathsf{HDUSGCMAdapter}.
} HDUSGCMAdapter .
讓 HDUS 可以原生請求 GCM 計算編排,但不重新定義 GCM。
14. 下一階段十大不可破壞 Invariants
AO-01
AI Proposal ≠ Execution Authority . \boxed{
\text{AI Proposal}
\neq
\text{Execution Authority}.
} AI Proposal = Execution Authority .
AO-02
Execution Authority ≠ Commit Authority . \boxed{
\text{Execution Authority}
\neq
\text{Commit Authority}.
} Execution Authority = Commit Authority .
AO-03
Resource Optimization ⇏ Constraint Relaxation . \boxed{
\text{Resource Optimization}
\not\Rightarrow
\text{Constraint Relaxation}.
} Resource Optimization ⇒ Constraint Relaxation .
AO-04
Historical Performance ⇏ Semantic Admissibility . \boxed{
\text{Historical Performance}
\not\Rightarrow
\text{Semantic Admissibility}.
} Historical Performance ⇒ Semantic Admissibility .
AO-05
Logical Resource Claim ≠ Physical Enforcement . \boxed{
\text{Logical Resource Claim}
\neq
\text{Physical Enforcement}.
} Logical Resource Claim = Physical Enforcement .
AO-06
GCM Orchestration ≠ Kernel Scheduling . \boxed{
\text{GCM Orchestration}
\neq
\text{Kernel Scheduling}.
} GCM Orchestration = Kernel Scheduling .
AO-07
Native Compute Fabric ≠ GCM World . \boxed{
\text{Native Compute Fabric}
\neq
\text{GCM World}.
} Native Compute Fabric = GCM World .
AO-08
HDUS Integration ≠ GCM Redefinition . \boxed{
\text{HDUS Integration}
\neq
\text{GCM Redefinition}.
} HDUS Integration = GCM Redefinition .
AO-09
Adaptive Routing ⇏ Foundation Auto-Revision . \boxed{
\text{Adaptive Routing}
\not\Rightarrow
\text{Foundation Auto-Revision}.
} Adaptive Routing ⇒ Foundation Auto-Revision .
AO-10
Optimization ≠ Mathematics ≠ Verification . \boxed{
\text{Optimization}
\neq
\text{Mathematics}
\neq
\text{Verification}.
} Optimization = Mathematics = Verification .
15. 結論
GCM 已經完成:
P 24 / P 72 → Formal Methodology → Runtime Architecture → Conformance → Executable Reference Runtime . \boxed{
\mathfrak P_{24}/\mathfrak P_{72}
\rightarrow
\text{Formal Methodology}
\rightarrow
\text{Runtime Architecture}
\rightarrow
\text{Conformance}
\rightarrow
\text{Executable Reference Runtime}.
} P 24 / P 72 → Formal Methodology → Runtime Architecture → Conformance → Executable Reference Runtime .
下一步應該先把 Reference Runtime 完成到真正的 Developer-Usable 100%,再建立 deterministic Compute Allocator。
之後才逐步加入 AI autonomy。
最終演進為:
GCM Reference Runtime → Deterministic Compute Allocator → AI Compute Orchestrator → Native Compute Fabric → HDUS Integration . \boxed{
\begin{aligned}
&\text{GCM Reference Runtime}\\
\rightarrow\;&\text{Deterministic Compute Allocator}\\
\rightarrow\;&\text{AI Compute Orchestrator}\\
\rightarrow\;&\text{Native Compute Fabric}\\
\rightarrow\;&\text{HDUS Integration}.
\end{aligned}
} → → → → GCM Reference Runtime Deterministic Compute Allocator AI Compute Orchestrator Native Compute Fabric HDUS Integration .
到那時,AI 所做的不只是:
「把這個程式丟到 GPU。」
而是:
判斷這個目標應採何種計算、如何分解、在哪裡執行、使用多少解析度、哪些資料應物化、哪些部分應保持休眠、需要投入多少 CPU/GPU/NPU/memory/storage、何時切換配置,以及哪些結果具備足夠證據可以 commit。
這才是 GCM 從一套 Global Computation Methodology 逐步成為:
AI-native Computational Operating Layer \boxed{
\text{AI-native Computational Operating Layer}
} AI-native Computational Operating Layer
的合理下一階段。
Canonical Handoff
本白皮書之後的立即工程工作固定為:
先將 GCM Reference Runtime MVP v0.1 補至 Developer-Usable 100%。 \boxed{
\textbf{先將 GCM Reference Runtime MVP v0.1 補至 Developer-Usable 100\%。}
} 先將 GCM Reference Runtime MVP v0.1 補至 Developer-Usable 100% 。
完成後才正式啟動:
GCM Deterministic Compute Allocation Algorithm v0.1。 \boxed{
\textbf{GCM Deterministic Compute Allocation Algorithm v0.1。}
} GCM Deterministic Compute Allocation Algorithm v0.1 。
AI Orchestration、Native Compute Fabric 與 HDUS Adapter 依序建立於其上。
除非出現新的 explicit canonical revision,否則不得跳過 MVP hardening,直接讓 AI autonomy 控制尚未完成 durability、plugin、recovery 與 external-developer contract 的 Reference Runtime。