Lightning-AI / Lightning-AI/lightning-thunder

Guard nvFuser-based tests with minimum required compute capability or BF16 capability

Open
#755 2 comments 0 reactions 0 assignees View on GitHub
ci / tests
Dominant language
Python
Stars
1.5k
Forks
121
PR merge metrics
No merged PRs in 30d

Description

Running `thunder/tests/test_torch_compile_executor.py::test_torch_compile_cat_nvfuser_phi2_tanh` and likely other tests would fail on pre-Ampere GPUs with

```py
RuntimeError: Target compute capability is 7.0 but this fusion requires at least 8.0. Reason: Fusion contains BFloat16 values which was introduced in Ampere (8.0)
```
PyTorch has a utility function to query whether the current device supports BF16 `torch.cuda.is_bf16_supported`.

cc @borda @tfogal

Contributor guide

No contributing guide indexed for this repository

Research direction

Start in thunder/tests/test_torch_compile_executor.py with test_torch_compile_cat_nvfuser_phi2_tanh, then review nearby nvFuser-based tests. Check torch.cuda.is_bf16_supported and the existing test behavior on pre-Ampere GPUs. Done means unsupported devices no longer fail these tests while supported devices still run them.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
testing-qa
Issue type
Bug
Difficulty
2/5
Estimated time
1-3 hours
Activity status
Stale
Clarity
Clearly specified
Newbie friendliness
55/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.