[tests] make test_from_pretrained_low_cpu_mem_usage_equal less flaky

[gante]

What does this PR do?

test_from_pretrained_low_cpu_mem_usage_equal has been failing frequently:

• list of flaky tests
• example

Increasing the tolerance a little bit should do it 🤗

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ydshieh]

Thank you @gante I was also looking at this, but interrupted by the lunch . It looks like it fails in the first run (where there is still no local cache for this model), and I am going to push something and let's see.

[ydshieh]

works for me, but let's ping @muellerzr / @SunMarc for a second eyes as they might know if there is any recent change in accelerate that cause this flakyness

Feel free to merge (to avoid failing on main) however: we could revise if the above experts have different opinions

[ydshieh]

For the record: Still get 3 failures in a 100 runs.

[SunMarc]

I spent a bit of time on this test and the flakyness comes from the cpu memory report which is a bit inconsistant. Also, due to this inconsistency, we are not testing the loading correctly since the model is way too small (tiny bert is only 500kb and we are allowing a delta of 2MB) ;). With the modification I proposed, it should be better. If you think that the model is still too big, maybe we can make this a slow test.

[gante]

@SunMarc the test indeed becomes slow (~10 secs) -- added the @slow decorator 🤗

[SunMarc]

Thanks !