Fully Sharded Data Parallel

April 5, 2023 ยท View on GitHub

FSDP: Fully Sharded Data Parallel: faster AI training with fewer GPUs

Training Step

bash alpaca/scripts/fsdp/run_train.sh

bash alpaca/scripts/fsdp/run_train_cpu_offload.sh

Inference Step

  • (Batch-Level) Please prepare the test file.

    bash alpaca/scripts/fsdp/inference/run_inf.sh

  • (Instance-Level) Using alpaca/src/inference.py line 17 to set prompts.

    bash alpaca/scripts/fsdp/inference/run_inf_hub.sh

Contents

  1. 1Training Step
  2. 2Inference Step