ローカルAIクラスター

単体端末では収まらないモデルや複数利用者の推論を、既存のPC/Mac/GPU資産を束ねて成立させられる可能性がある。一方、総RAM/VRAMの単純加算は性能・可用性・安全性を保証せず、network topology、runtime、scheduler、heterogeneous device、failure domainまで含む運用設計が必要。

Enterprise AI Core / implementation / non-LLM component / infrastructure / interoperability / benchmark / operability

Local AI Clusterはどの層で成立し、単一node推論と比べて何が増え、どの条件で「複数台を束ねる」ことが実益になるか。consumer/office hardwareのheterogeneous clusterをproduction-grade shared inferenceへ外挿できるか。

• source class: author-maintained OSS / implementation exact title: exo: Run frontier AI locally. author/org: exo labs / exo-explore canonical URL: https://github.com/exo-explore/exo publication/update/access: accessed 2026-09-16 version/release/tag/commit: exact current tag/commit not pinned in this run supporting locator: README Features / Quick Start; automatic discovery, topology-aware auto parallel, tensor parallelism, MLX distributed, offline mode, read-only pre-downloaded model directories evidence label: OSS_PRACTICE / PRODUCT_VOLATILE relation: support/context

• source class: author-maintained OSS / implementation exact title: llama.cpp RPC backend README author/org: ggml-org canonical URL: https://github.com/ggml-org/llama.cpp/blob/master/tools/rpc/README.md publication/update/access: accessed 2026-09-16 version/release/tag/commit: master; exact commit not pinned supporting locator: Overview / Important; remote ggml devices, one-or-several RPC servers, proof-of-concept, fragile and insecure, do not expose to open network or sensitive environment evidence label: OSS_PRACTICE relation: support/limit

• source class: author issue / failure report exact title: Bug: RPC inference is drastically slower even on localhost #8832 author/org: ggml-org/llama.cpp issue reporter canonical URL: https://github.com/ggml-org/llama.cpp/issues/8832 publication/update/access: opened 2024-08-02; accessed 2026-09-16 version/release/tag/commit: issue-specific environment supporting locator: reported llama3.1-8B CUDA RPC 1.9 tok/s vs local 25 tok/s on same host evidence label: SOURCE_REPORT relation: counter/limit

• source class: author issue / failure report exact title: Eval bug: Crash when offloading large models via RPC if model size exceeds ~75% of server RAM #15055 author/org: ggml-org/llama.cpp issue reporter canonical URL: https://github.com/ggml-org/llama.cpp/issues/15055 publication/update/access: accessed 2026-09-16 version/release/tag/commit: issue-specific Apple M3 Ultra environment supporting locator: Problem description; large-model remote offload failure under high memory pressure evidence label: SOURCE_REPORT relation: counter/limit

• source class: official scheduler documentation exact title: Slurm Workload Manager — Generic Resource (GRES) Scheduling author/org: SchedMD canonical URL: https://slurm.schedmd.com/gres.html publication/update/access: accessed 2026-09-16 version/release/tag/commit: current docs; AutoDetect=nvidia noted as added in Slurm 24.11 supporting locator: Overview / AutoDetect / Accounting; GPU GRES, explicit requests, GPU memory/utilization accounting, heterogeneous GPU types evidence label: VERIFIED_SPEC relation: support/context

• source class: author-maintained framework documentation exact title: Ray Cluster Key Concepts author/org: Ray Project / Anyscale community canonical URL: https://github.com/ray-project/ray/blob/master/doc/source/cluster/key-concepts.rst publication/update/access: accessed 2026-09-16 version/release/tag/commit: master; exact commit not pinned supporting locator: Head Node / Worker Node; head carries GCS/autoscaler/driver management, workers execute tasks/actors evidence label: OSS_PRACTICE relation: context

• 「束ねる」には少なくとも二種類ある。(A) 1 inferenceを複数device/nodeへshardするmodel-parallel/distributed inference、(B) independent replicas/jobsを複数nodeへ配置するthroughput/queue cluster。両者を同じcapacity claimにしない。
• exoはconsumer/Apple Siliconを含むlocal devicesを自動発見し、topology-aware parallelism、MLX distributed、複数API互換、offline modeを実装している。Local AI Clusterが研究玩具だけでなく実装可能なclassであることは確立。
• Slurm GRESはGPUを明示的resourceとしてschedule/accountできる。したがって小規模clusterでも「空いているGPUを何となく使う」から、allocation/accountingを伴うshared resourceへ移行できる。
• llama.cpp RPCはremote device offloadを実装するが、README自身がproof-of-concept・fragile・insecureと明記。local clusterという概念の存在証拠にはなるが、production securityの根拠にはしない。
• network/interconnectがcritical pathになるmodel parallelismでは、総memory増加とlatency/throughput改善は別。heterogeneous slow nodeを足すとmodel fitは改善してもsingle-request latencyを悪化させ得る。

• aggregate RAM/VRAM >= model size は必要条件になり得るが十分条件ではない。runtimeがそのdevice combinationをsupportすること、tensor placement、network bandwidth/latency、workspace/KV headroomが必要。
• llama.cpp RPC issue #8832の13x slowdownは単一issue/environmentであり全version一般化禁止。ただし「remote offload overheadを無視できない」反証として保持。
• issue #15055も全deployment一般化禁止。高memory pressure時のremote offload failure envelopeを示すSOURCE_REPORTとしてのみ使用。
• consumer auto-discoveryは便利だが、production networkでのidentity/authentication/authorizationと同義ではない。
• node loss時にmodel-parallel request全体が失敗する構成と、replica clusterで別nodeへretry可能な構成はfailure domainが異なる。

OS/architecture; accelerator/runtime compatibility; model format; RAM/VRAM per node; interconnect bandwidth/latency; switch/NIC/Thunderbolt/RDMA availability; scheduler/runtime version; model license; LAN segmentation; authentication; storage/NFS/model distribution; power/thermal envelope.

OBSERVED externally only through primary docs/OSS; no tenant live test performed. Practical patterns: exo/MLX distributed for Apple Silicon-oriented cluster; llama.cpp RPC for experimental remote ggml offload; Slurm for explicit GPU job scheduling/accounting; Ray for task/actor cluster orchestration. These are not interchangeable layers.

Future TEST_PROFILE: fixed model artifact and prompt set; compare 1 node vs 2/4 nodes; model-parallel vs replica/queue mode; homogeneous vs heterogeneous nodes; Ethernet/Wi-Fi/Thunderbolt/RDMA where applicable. Measure model-load time, TTFT, TPOT, tokens/s, requests/s, p50/p95/p99, inter-node traffic, peak memory/node, power, node-loss recovery, cold/warm restart, scheduler wait, output equivalence. Never report only aggregate memory.

Security: RPC/device-discovery plane must not be presumed trusted; isolate management/data planes and authenticate endpoints. Privacy: local execution reduces external data transfer but does not by itself enforce user/tenant isolation. Audit: node/model/runtime/topology/job/request mapping required. Cost: reuse of owned hardware can reduce acquisition cost but power, networking, admin time, low utilization and heterogeneous tail latency must be included. Operability: patch/driver/runtime skew and failed node replacement are primary burdens.

「複数のPC/Mac/GPUを束ねるLocal AI Clusterは、単体では収まらないモデルの分散推論または複数requestの共有処理を可能にし得る。ただし総RAM/VRAMの加算だけでは性能を予測できず、runtime対応、network topology、scheduler、failure domainを実測する必要がある。consumer向けdistributed runtimeにはexperimental/security制約を持つ実装もあるため、production採用は別途評価する。」

Core候補 LOCAL_AI_CLUSTER_CONTRACT: cluster_id; node_id; hardware/accelerator; OS; driver/runtime; model artifact/hash; partition/replica strategy; topology/interconnect; scheduler/runtime version; per-node memory; request placement; benchmark fixture; TTFT/TPOT/throughput; network traffic; failure/recovery; authn/authz; energy/cost; deployment/rollback. EA-088 ACCELERATOR CAPACITY CONTRACTと接続する。

exo exact current release/tag/commit and production hardening state; llama.cpp RPC current exact commit/security evolution; cross-platform heterogeneous runtime matrix; small-office RDMA/Thunderbolt operational evidence; independent long-duration production case studies. Stop condition for this run: architecture/implementation/counterevidence/test profileまでboundedに確立。

---

---