```mermaid
graph TD
User((User / App)) -->|HTTP Request| LB[Load Balancer]
LB -->|gRPC / REST| API_Server[推論APIサーバー<br>FastAPI / Triton / vLLM]
subgraph Physical_Server [物理サーバー / ノード]
CPU[<b>CPU & System RAM</b><br>・リクエスト受付<br>・前処理 / Tokenizer<br>・GPUへの命令出し]
API_Server -->|Input Tokens| CPU
subgraph GPU_Cluster [GPU VRAM空間]
direction LR
subgraph GPU_0 [GPU #0 <br>NVIDIA H100]
VRAM_0[<b>VRAM 80GB</b>]
Model_Part_A[モデルの重み<br>Part A]
KV_Cache_A[KV Cache<br>短期記憶]
end
subgraph GPU_1 [GPU #1 <br>NVIDIA H100]
VRAM_1[<b>VRAM 80GB</b>]
Model_Part_B[モデルの重み<br>Part B]
KV_Cache_B[KV Cache<br>短期記憶]
end
%% NVLinkによる高速通信
GPU_0 <==>|<b>NVLink / Switch</b><br>900GB/sの帯域で同期| GPU_1
end
CPU -->|PCIe Gen5| GPU_0
end
GPU_0 & GPU_1 -->|Output Logits| CPU
CPU -->|Response| User
classDef gpu fill:#e1f5fe,stroke:#01579b,stroke-width:2px;
classDef vram fill:#fff9c4,stroke:#fbc02d,stroke-width:1px;
class GPU_0,GPU_1 gpu;
class VRAM_0,VRAM_1,Model_Part_A,Model_Part_B,KV_Cache_A,KV_Cache_B vram;
```
```mermaid
graph TB
subgraph Storage_System [分散ストレージ]
Lustre[(<b>Lustre / GPFS</b><br>数PBの学習データ)]
end
subgraph Compute_Cluster [計算クラスター <br>数千ノード]
Node1[Node 1<br>8x H100]
Node2[Node 2<br>8x H100]
Node3[Node 3<br>8x H100]
NodeN[Node N...]
end
%% 高速ネットワークファブリック
Switch_Spine{<b>Spine Switch</b><br>InfiniBand / RoCE}
Switch_Leaf1{Leaf Switch}
Switch_Leaf2{Leaf Switch}
Lustre ==>|1TB/s~ Read| Switch_Spine
Switch_Spine ==> Switch_Leaf1
Switch_Spine ==> Switch_Leaf2
Switch_Leaf1 ==> Node1
Switch_Leaf1 ==> Node2
Switch_Leaf2 ==> Node3
Switch_Leaf2 ==> NodeN
%% パラメータ同期
Node1 <.->|All-Reduce通信<br>勾配の同期| Node2
Node2 <.->|All-Reduce通信| Node3
classDef storage fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px;
classDef net fill:#e0f2f1,stroke:#00695c,stroke-width:2px;
class Lustre storage;
class Switch_Spine,Switch_Leaf1,Switch_Leaf2 net;
```