Distributed Training

Distributed Deep Learning Mechanics

Ring All-Reduce

Synchronous data parallel training requires gradients from all GPUs to be averaged before the optimizer step. Instead of a parameter server bottleneck, Ring All-Reduce organizes N GPUs into a logical ring. The gradient tensors are split into N chunks. During the scatter-reduce phase, GPUs pass chunks around the ring, accumulating sums. After N-1 steps, each GPU holds one fully accumulated chunk. In the subsequent all-gather phase, these complete chunks are circulated. Total network transfer per GPU is 2*(N-1)/N * Size, making it bandwidth-optimal and highly scalable.

ZeRO Optimizer (Zero Redundancy Optimizer)

Standard Data Parallelism duplicates the entire model state across all GPUs, which is impossible for LLMs like GPT-3. ZeRO eliminates this memory redundancy by partitioning the state.

• ZeRO Stage 1: Partitions Optimizer States (e.g., Adam momentum/variance). Cuts memory footprint by 4x.
• ZeRO Stage 2: Partitions Gradients. Each GPU only holds gradients for its parameter partition, reducing memory by 8x.
• ZeRO Stage 3: Partitions Parameters. Model weights are scattered. Forward/backward passes use dynamic collective communications (All-Gather) to fetch required parameters just-in-time for a layer's computation, and immediately discard them, allowing training of trillion-parameter models.

flowchart TD
%%{init: {"theme": "default", "flowchart": {"useMaxWidth": true}}}%%
    subgraph CommunicationRingAllReducePhase ["Ring All-Reduce Phase<br><br><br>"]
        GPU_0 -->|"send_chunk()"| GPU_1
        GPU_1 -->|"accumulate()"| GPU_2
        GPU_2 -->|"forward()"| GPU_0
    end
    subgraph ZeROStagesZeROPartitioning ["ZeRO Partitioning<br><br><br>"]
        Stage1 -->|"partition_adam()"| MemorySave
        Stage2 -->|"partition_grads()"| ReduceScatter
        Stage3 -->|"partition_weights()"| AllGather
    end