XNNPACK test_fp16_gelu is producing a bug

[mergennachin]

🐛 Describe the bug

 buck test @fbcode//mode/dev fbcode//executorch/backends/xnnpack/test:test_xnnpack_ops -- --exact 'fbcode//executorch/backends/xnnpack/test:test_xnnpack_ops - test_fp16_gelu (executorch.backends.xnnpack.test.ops.test_gelu.TestGelu)' --run-disabled

The XNNPACK execution is producing all zeros for fp16 GELU while the reference produces correct values. Looking at the output:

• Model (XNNPACK): median=0, mean=0, max=0, min=0 (all zeros)
• Reference: median=-0.071, mean=0.275, max=1.466, min=-0.169 (correct values)

Versions

main branch

[mergennachin]

Internal test X link:

https://www.internalfb.com/intern/test/562950220305931?ref_report_id=0

[GregoryComer]

Looks like there is a bug in older versions of XNNPACK where it's silently doing nothing on fp16 GeLU due to no kernels and a logic bug / oversight that causes it to silently fail in this case. There are now upstream fp16 GeLU kernels, so it will get fixed when we update the dependency. For now, I'll just disable partitioning for fp16 gelu.

Note that this bug only exists internally, as we're on an old version of XNNPACK.

[mergennachin]

@GregoryComer great, we can remove 1.1 release blocker, right?

[GregoryComer]

Yeah, for the OSS release, it should be fine as is. I'll confirm that the test passes in OSS. I also just put up a PR to disable partitioning of fp16 gelu until we update the XNNPACK dep internally in a few weeks.

[GregoryComer]

I validated that the test passes on the latest master in OSS:

python -m unittest backends.xnnpack.test.ops.test_gelu.TestGelu.test_fp16_gelu
...
Ran 1 test in 1.488s

OK