Fix Pixart Slow Tests #6962
Fix Pixart Slow Tests #6962
Conversation
|
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. |
| @@ -332,7 +338,7 @@ def tearDown(self): | |||
| torch.cuda.empty_cache() | |||
|
|
|||
| def test_pixart_1024(self): | |||
| generator = torch.manual_seed(0) | |||
There was a problem hiding this comment.
Shouldn't it by default create it on CPU?
There was a problem hiding this comment.
Yes. I just changed it to match how the generator is created in the other tests.
| @@ -341,14 +347,13 @@ def test_pixart_1024(self): | |||
| image = pipe(prompt, generator=generator, output_type="np").images | |||
|
|
|||
| image_slice = image[0, -3:, -3:, -1] | |||
| expected_slice = np.array([0.2891, 0.2749, 0.2595, 0.3020, 0.2698, 0.2671, 0.3169, 0.2993, 0.3179]) | |||
There was a problem hiding this comment.
That's quite big a change. Why is this coming?
There was a problem hiding this comment.
The difference produced currently between the output and expected output is quite large
https://github.com/huggingface/diffusers/actions/runs/7876381153/job/21491555064#step:7:485
I just ran print_tensor_test on the current outputs in our new containers (with torch 2.2) and updated. If this seems like a deeper issue, let me know how you want to tackle it.
There was a problem hiding this comment.
Can we pintpoint a version that passes with the current assertion values? Maybe with Torch 2.1, etc.?
There was a problem hiding this comment.
Actually, the values will have to change here. This test is running 20 inference steps, which is too many.
There was a problem hiding this comment.
But we're not changing the num_inference_steps yet no? It's okay if the values are changed because of a reduced number of steps. But at 20 steps, the assertion values shouldn't change.
There was a problem hiding this comment.
The tests are failing with torch 2.1 as well. It could also be the CUDA version in the runners. I've run the pipeline example in the docs and the generated image seems fine.
It feels like the test is most likely failing due to changes in torch version or CUDA version. But if you would like to work on an alternative solution, feel free to take a look.
What does this PR do?
Fixes precision related issues in the Pixart Slow tests.
Fixes # (issue)
Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.