A GPipe implementation in PyTorch
-
Updated
Jul 25, 2024 - Python
A GPipe implementation in PyTorch
Model parallelism for NN architectures with skip connections (eg. ResNets, UNets)
Development of Project HPGO | Hybrid Parallelism Global Orchestration
A fully distributed hyperparameter optimization tool for PyTorch DNNs
Simulated Multi-GPU inference engine implementing Megatron-style Tensor Parallelism, GPipe Pipeline Parallelism, and KV Cache Sharding from scratch. Features bit-identical fidelity validation on Qwen2-0.5B weights and analytical communication cost modeling.
To associate your repository with the gpipe topic, visit your repo's landing page and select "manage topics."