Lightning-AI / Lightning-AI/lightning-thunder
[FSDP] Support gradient clipping by norm
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 1.5k
- Forks
- 121
- PR merge metrics
- No merged PRs in 30d
Description
🚀 Feature
Pitch
Port https://github.com/pytorch/pytorch/blob/c4a157086482899f0640d03292e5d2c9a6a3db68/torch/distributed/fsdp/fully_sharded_data_parallel.py#L1069-L1194 to work with Thunder's FSDP.
This could be importable through from thunder.distributed.utils import clip_grad_norm_.
We could also move FSDP into thunder.distributed.fsdp and put this alongside it (from thunder.distributed.fsdp import clip_grad_norm_). Bikeshedding welcome.
PyTorch docs: https://pytorch.org/docs/stable/generated/torch.nn.utils.clip_grad_norm_.html, https://pytorch.org/docs/stable/fsdp.html#torch.distributed.fsdp.FullyShardedDataParallel.clip_grad_norm_
cc @carmocca @awaelchli @crcrpar
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reading the referenced PyTorch implementation in torch/distributed/fsdp/fully_sharded_data_parallel.py, lines 1069-1194, alongside the linked clip_grad_norm_ and FSDP documentation. Compare the available Thunder FSDP entry points and decide whether the function belongs in thunder.distributed.utils or thunder.distributed.fsdp. Done means Thunder's FSDP supports gradient clipping by norm through the chosen import path.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- distributed-systems, machine-learning
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 42/100