Lightning-AI / Lightning-AI/lightning-thunder
Extend catching empty autocast regions to catch other empty contexts
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 1.5k
- Forks
- 121
- PR merge metrics
- No merged PRs in 30d
Description
Originally posted by @kshitij12345 in https://github.com/Lightning-AI/lightning-thunder/pull/1400#discussion_r1833018302
Besides {_enter/_exit}_autocast, from trace_rules.py, I could find
{_enter/_exit}_dual_level for forward mode auto-diff, {_enter/_exit}_inference_mode for inference mode.
Besides these, looking at ctx_manager.py, there could also be Stream, Grad, Device related context managers which are mapped to something like the following
old_stream = torch.cuda.get_current_stream()
new_stream = torch.cuda.Stream()
set_stream(new_stream)
# Some operations
set_stream(old_stream)
Refs:
https://github.com/pytorch/pytorch/blob/781c68c86549aa2fd155f92afb0406b7577ab31c/torch/_dynamo/trace_rules.py#L550-L552
https://github.com/pytorch/pytorch/blob/781c68c86549aa2fd155f92afb0406b7577ab31c/torch/_dynamo/trace_rules.py#L2393-L2394
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 trace_rules.py and ctx_manager.py entries for dual-level, inference-mode, Stream, Grad, and Device contexts. Trace how lightning-thunder currently handles empty autocast regions, then define coverage for the other empty contexts and verify that the relevant empty-context cases are handled without changing non-empty behavior.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- compilers
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100