intel / intel/torch-xpu-ops

[upstream_ut] test_gather_backward_one_dim_xpu accuracy mismatch

Open
#5,307 0 comments 0 reactions 0 assignees View on GitHub
test: ut
Dominant language
Python
Stars
113
Forks
128
Avg merge
5d 13h
Merged PRs (30d)
107

Description

## Description
`test_gather_backward_one_dim_xpu` fails with incorrect gradient values from the 1-D gather backward on XPU.

## Error log
```
File "test/test_torch.py", line 2023, in _test_gather_backward_one_dim
self.assertEqual(grad, expected, atol=0, rtol=0)
AssertionError: Tensor-likes are not equal!
Mismatched elements: 1741 / 2989 (58.2%)
Greatest absolute difference: 4.0 at index (423,)
Greatest relative difference: 3.5682600696418376e-07 at index (423,)
```

## Reproduce
Using pytorch/pytorch#196421 on a PVC machine:
```
python test/test_torch.py TestTorchDeviceTypeXPU.test_gather_backward_one_dim_xpu
```

## Expected
gather backward accumulates gradients correctly (bit-exact per the test's atol=0/rtol=0), matching CPU/CUDA.

_Filed with AI assistance; reproduced per the details above._

Contributor guide

Open the contributing guide

Research direction

Start with test/test_torch.py and the _test_gather_backward_one_dim test around line 2023. Run TestTorchDeviceTypeXPU.test_gather_backward_one_dim_xpu on a PVC machine and compare the XPU gradient with the CPU/CUDA results and the exact assertions. Done means the XPU gather backward gradients match the expected values with atol=0 and rtol=0.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
backend, testing
Issue type
Bug
Difficulty
3/5
Estimated time
1-2 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
68/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.