graph TD
subgraph Hardware_Layer [Hardware Layer]
GPU[NVIDIA GPU - CUDA 12.4]
end
subgraph MLOps_and_Telemetry [MLOps & Telemetry]
WandB[Weights and Biases Cloud]
Redis[Redis In-Memory Datastore]
ReactUI[React + Vite Web UI]
Redis -->|WebSockets| ReactUI
end
subgraph IPC_Communication_Backbone [IPC & Storage Backbone]
ZMQ_Socket((ZeroMQ PUSH/PULL Socket))
JIT_File[(model_jit.pt Checkpoint)]
end
subgraph Rust_Self_Play_Engine[Rust Self-Play Process - The Muscle]
direction TB
subgraph Rust_Workers[Lightweight Threads N=100+]
W1[Game Thread 1: u128 Bitboard + Gumbel MCTS]
W2[Game Thread 2: u128 Bitboard + Gumbel MCTS]
WN[Game Thread N: u128 Bitboard + Gumbel MCTS]
end
CB_Req((Crossbeam Lock-Free Queue: EvalRequests))
CB_Resp((Crossbeam Oneshot: EvalResponses))
LibTorch[Dedicated LibTorch Inference Thread]
FileWatcher[Notify FS Watcher: Hot Reload]
EpAggregator[Episode Aggregator]
W1 -->|Sends State| CB_Req
W2 -->|Sends State| CB_Req
WN -->|Sends State| CB_Req
CB_Req -->|Batches up to 256| LibTorch
LibTorch == Recurrent Inference ==> GPU
LibTorch -->|Unbatches| CB_Resp
CB_Resp -->|Returns Value/Policy| W1
CB_Resp -->|Returns Value/Policy| W2
CB_Resp -->|Returns Value/Policy| WN
W1 -->|Completed Game| EpAggregator
W2 -->|Completed Game| EpAggregator
WN -->|Completed Game| EpAggregator
EpAggregator -->|Serializes EpisodeMeta| ZMQ_Socket
FileWatcher -->|Detects Update| JIT_File
FileWatcher -->|Swaps CModule Pointer| LibTorch
W1 -.->|Spectator State| Redis
end
subgraph Python_Training_Engine [Python Training Process - The Brain]
direction TB
ZMQ_Receiver[ZeroMQ PULL Thread]
PyBuffer[(PyTorch Replay Buffer)]
DataLoader[PyTorch DataLoader workers=0]
Trainer[MuZero Trainer Loop]
Model[MuZeroNet: Rep, Dyn, Pred]
Opt[AdamW Optimizer + LR Scheduler]
Reanalyze[Reanalyze Daemon]
ZMQ_Socket -->|Streams Bytes| ZMQ_Receiver
ZMQ_Receiver -->|Appends| PyBuffer
PyBuffer -->|Samples| DataLoader
DataLoader -->|Yields Batches| Trainer
Trainer -->|Forward/Backward| Model
Model == Gradient Descent ==> GPU
Trainer -->|Updates| Opt
Trainer -->|Saves every N steps| JIT_File
Trainer -->|Logs Loss/LR| WandB
Trainer -.->|Logs Status| Redis
Reanalyze <-->|Updates Stale Targets| PyBuffer
end
classDef rust fill:#b7410e,stroke:#000,stroke-width:2px,color:#fff;
classDef python fill:#2b5b84,stroke:#3776ab,stroke-width:2px,color:#fff;
classDef hardware fill:#76b900,stroke:#000,stroke-width:2px,color:#000;
classDef ipc fill:#4a4a4a,stroke:#fff,stroke-width:2px,color:#fff;
classDef mlops fill:#eeb422,stroke:#000,stroke-width:2px,color:#000;
class Rust_Self_Play_Engine,Rust_Workers,W1,W2,WN,CB_Req,CB_Resp,LibTorch,FileWatcher,EpAggregator rust;
class Python_Training_Engine,ZMQ_Receiver,PyBuffer,DataLoader,Trainer,Model,Opt,Reanalyze python;
class GPU hardware;
class IPC_Communication_Backbone,ZMQ_Socket,JIT_File ipc;
class MLOps_and_Telemetry,WandB,Redis,ReactUI mlops;
graph TD
subgraph Hardware_Layer [Hardware Layer]
GPU[NVIDIA GPU - CUDA 12.4]
end
subgraph MLOps_and_Telemetry [MLOps & Telemetry]
WandB[Weights and Biases Cloud]
Redis[Redis In-Memory Datastore]
ReactUI[React + Vite Web UI]
Redis -->|WebSockets| ReactUI
end
subgraph IPC_Communication_Backbone [IPC & Storage Backbone]
ZMQ_Socket((ZeroMQ PUSH/PULL Socket))
JIT_File[(model_jit.pt Checkpoint)]
end
subgraph Rust_Self_Play_Engine[Rust Self-Play Process - The Muscle]
direction TB
subgraph Rust_Workers[Lightweight Threads N=100+]
W1[Game Thread 1: u128 Bitboard + Gumbel MCTS]
W2[Game Thread 2: u128 Bitboard + Gumbel MCTS]
WN[Game Thread N: u128 Bitboard + Gumbel MCTS]
end
CB_Req((Crossbeam Lock-Free Queue: EvalRequests))
CB_Resp((Crossbeam Oneshot: EvalResponses))
LibTorch[Dedicated LibTorch Inference Thread]
FileWatcher[Notify FS Watcher: Hot Reload]
EpAggregator[Episode Aggregator]
W1 -->|Sends State| CB_Req
W2 -->|Sends State| CB_Req
WN -->|Sends State| CB_Req
CB_Req -->|Batches up to 256| LibTorch
LibTorch == Recurrent Inference ==> GPU
LibTorch -->|Unbatches| CB_Resp
CB_Resp -->|Returns Value/Policy| W1
CB_Resp -->|Returns Value/Policy| W2
CB_Resp -->|Returns Value/Policy| WN
W1 -->|Completed Game| EpAggregator
W2 -->|Completed Game| EpAggregator
WN -->|Completed Game| EpAggregator
EpAggregator -->|Serializes EpisodeMeta| ZMQ_Socket
FileWatcher -->|Detects Update| JIT_File
FileWatcher -->|Swaps CModule Pointer| LibTorch
W1 -.->|Spectator State| Redis
end
subgraph Python_Training_Engine [Python Training Process - The Brain]
direction TB
ZMQ_Receiver[ZeroMQ PULL Thread]
PyBuffer[(PyTorch Replay Buffer)]
DataLoader[PyTorch DataLoader workers=0]
Trainer[MuZero Trainer Loop]
Model[MuZeroNet: Rep, Dyn, Pred]
Opt[AdamW Optimizer + LR Scheduler]
Reanalyze[Reanalyze Daemon]
ZMQ_Socket -->|Streams Bytes| ZMQ_Receiver
ZMQ_Receiver -->|Appends| PyBuffer
PyBuffer -->|Samples| DataLoader
DataLoader -->|Yields Batches| Trainer
Trainer -->|Forward/Backward| Model
Model == Gradient Descent ==> GPU
Trainer -->|Updates| Opt
Trainer -->|Saves every N steps| JIT_File
Trainer -->|Logs Loss/LR| WandB
Trainer -.->|Logs Status| Redis
Reanalyze <-->|Updates Stale Targets| PyBuffer
end
classDef rust fill:#b7410e,stroke:#000,stroke-width:2px,color:#fff;
classDef python fill:#2b5b84,stroke:#3776ab,stroke-width:2px,color:#fff;
classDef hardware fill:#76b900,stroke:#000,stroke-width:2px,color:#000;
classDef ipc fill:#4a4a4a,stroke:#fff,stroke-width:2px,color:#fff;
classDef mlops fill:#eeb422,stroke:#000,stroke-width:2px,color:#000;
class Rust_Self_Play_Engine,Rust_Workers,W1,W2,WN,CB_Req,CB_Resp,LibTorch,FileWatcher,EpAggregator rust;
class Python_Training_Engine,ZMQ_Receiver,PyBuffer,DataLoader,Trainer,Model,Opt,Reanalyze python;
class GPU hardware;
class IPC_Communication_Backbone,ZMQ_Socket,JIT_File ipc;
class MLOps_and_Telemetry,WandB,Redis,ReactUI mlops;