Distributed Training & Training Infrastructure

Parent: LLM Models and APIs · researched 2026-05-31T21:35:36.545Z· 35 sources · 13 concepts · skill distributed-training

> Hub reference under ai-agent-engineering (hub-and-spoke). Owns the LLM training-infrastructure layer: how you split a model + optimizer + activations across many GPUs to TRAIN it. Loaded on demand w

Distributed Training & Training Infrastructure

Concepts covered (MECE, 12)

Boundaries (defers to siblings — no duplication)

Sources

Children

Frontier under this node: 3D/ND parallelism composition (TP×PP×CP×EP×DP placement) + expert parallelism, Collective communication (NCCL all-reduce/all-gather/reduce-scatter/all-to-all, ring vs tree, overlap), Context/sequence parallelism for long context (Ring Attention, DeepSpeed-Ulysses, USP), Data parallelism & PyTorch DDP (gradient bucketing, backward/comm overlap), Distributed checkpointing (PyTorch DCP sharded + async, resharding), FSDP & FSDP2 (FlatParameter→per-parameter DTensor, HSDP), Frameworks (torchtitan, Megatron-Core, DeepSpeed, NeMo, Composer) & MFU/HFU scaling efficiency, Gradient checkpointing / selective activation recomputation & gradient accumulation, Mixed precision (FP16 vs BF16 vs FP8/Transformer-Engine, loss scaling), Pipeline parallelism (GPipe, 1F1B, interleaved, the bubble, Seq1F1B/DualPipe), Tensor parallelism (Megatron column/row split) + sequence parallelism, Training stability (loss spikes, z-loss, QK-norm, LR warmup, init, grad clip/ZClip), ZeRO optimizer stages 1/2/3 + ZeRO-Offload/ZeRO-Infinity

← the whole tree · 3D view· how to read this page