Implement FSDP from Scratch in PyTorch

Expertgpu-systemsauto-graded

Build Fully Sharded Data Parallel: shard parameters across GPUs, all-gather before forward, reduce-scatter gradients after backward.

Solve it

Check your answer

The grader verifies properties of your implementation, so a correct solution written differently from ours still passes.

pip install torchleet

from torchleet import check
check("fsdp", FakeDistributed, FSDPLinear)

Company tags

How these tags are sourced