🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/134881

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

❌ 2 New Failures, 1 Unrelated Failure

As of commit bac6299 with merge base 83875cd ():

This comment was automatically generated by Dr. CI and updates every 15 minutes.

Benchmark results are not good compared to eager.

[------------------------------------ max_unpool -----------------------------------]
                                                              |  Decomposed  |  Eager
12 threads: -------------------------------------------------------------------------
      {'dim': 0, 'keepdim': False, 'shape': (3000, 3000)}     |     1098     |   1048
      {'dim': 0, 'keepdim': True, 'shape': (3000, 3000)}      |     1070     |   1049
      {'dim': 1, 'keepdim': False, 'shape': (3000, 3000)}     |      689     |    538
      {'dim': 1, 'keepdim': True, 'shape': (3000, 3000)}      |      662     |    542
      {'shape': (3000, 3000)}                                 |     1425     |   1407
      {'dim': 0, 'keepdim': False, 'shape': (200, 200, 200)}  |     1818     |    846
      {'dim': 0, 'keepdim': True, 'shape': (200, 200, 200)}   |     1790     |    846
      {'dim': 1, 'keepdim': False, 'shape': (200, 200, 200)}  |     1768     |    740
      {'dim': 1, 'keepdim': True, 'shape': (200, 200, 200)}   |     1742     |    739
      {'dim': 2, 'keepdim': False, 'shape': (200, 200, 200)}  |      676     |    679
      {'dim': 2, 'keepdim': True, 'shape': (200, 200, 200)}   |      650     |    677
      {'shape': (200, 200, 200)}                              |     1274     |   1263

@isuruf I took a look into the perf, and some observations:

1. it seems the main difference in perf is that for the case where dim is passed in, eager actually has atopk kernel it dispatches to, which is the main reason why eager is faster than decomposed for those cases.
2. There are actually quite a bit of shapes where the decomposition is faster than aten. The issue is that your benchmarking is all with multiples of 100, and not multiples of 128 :P But this seems to mainly come down to whether aten.topk or aten.sort is faster.

For example

torch.Size([24000, 2048])
aten:  3.3983023166656494
compile:  2.7202703952789307

Thanks. I'll look at getting a topk kernel in.

Looks like this PR hasn't been updated in a while so we're going to go ahead and mark this as Stale.
Feel free to remove the Stale label if you feel this was a mistake.
If you are unable to remove the Stale label please contact a maintainer in order to do so.
If you want the bot to never mark this PR stale again, add the no-stale label.
Stale pull requests will automatically be closed after 30 days of inactivity.

I have removed the decompositions for median.dim and only kept it for median.default for now which uses aten.sort and therefore we don't have to get a topk kernel in. I also separated out the OpInfo for median.dim and median.default to two variants to make tests work.