Lightning-AI / Lightning-AI/lightning-thunder

Need to relax tolerance for a few float64 nvfuser grad tests

Open
#2,363 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

nvfuser testing
Dominant language
Python
Stars
1.5k
Forks
121
PR merge metrics
No merged PRs in 30d

Description

```
=========================== short test summary info ============================
FAILED thunder/tests/test_grad.py::test_vjp_correctness_softsign_nvfuser_cuda_thunder.dtypes.float64 - AssertionError: Scalars are not close!

Expected 0.0015554002273835033 but got 0.001591950186688797.
Absolute difference: 3.654995930529371e-05 (up to 1e-07 allowed)
Relative difference: 0.023498748850498824 (up to 1e-07 allowed)
FAILED thunder/tests/test_grad.py::test_vjp_correctness_normalize_nvfuser_cuda_thunder.dtypes.float64 - AssertionError: Scalars are not close!

Expected 2.9702087605694922 but got 2.970201815694213.
Absolute difference: 6.9448752793022095e-06 (up to 1e-07 allowed)
Relative difference: 2.338177494961881e-06 (up to 1e-07 allowed)
FAILED thunder/tests/test_grad.py::test_vjp_correctness_abs_nvfuser_cuda_thunder.dtypes.float64 - AssertionError: Scalars are not close!

Expected 0.09839208072782858 but got 0.09824138678418542.
Absolute difference: 0.00015069394364315702 (up to 1e-07 allowed)
Relative difference: 0.0015315657777377983 (up to 1e-07 allowed)
= 3 failed, 733 passed, 55 skipped, 23 xfailed, 15 xpassed, 92422 warnings in 839.96s (0:13:59) =
```

CI - https://dev.azure.com/Lightning-AI/lightning/_build/results?buildId=238959&view=logs&j=83d7ca60-c2d2-50c8-1ed9-9f99750fc4f0&t=d88aa08c-a7b7-5504-bdb7-50942829cc54

cc @tfogal @mruberry

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with the three failing cases in thunder/tests/test_grad.py: test_vjp_correctness_softsign_nvfuser_cuda_thunder, test_vjp_correctness_normalize_nvfuser_cuda_thunder, and test_vjp_correctness_abs_nvfuser_cuda_thunder, using the linked CI failure as context. Adjust their float64 tolerance as needed and rerun these tests; done means the three failures pass without causing regressions in the surrounding test file.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
compilers, testing-qa
Issue type
Bug
Difficulty
2/5
Estimated time
1-3 hours
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
42/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.