enable pipeline test cases on xpu #11527
Conversation
Signed-off-by: Yao Matrix <matrix.yao@intel.com>
src/diffusers/utils/testing_utils.py
Outdated
| response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) | ||
| response.raise_for_status() | ||
| arry = torch.load(BytesIO(response.content), map_location=map_location) | ||
| arry = torch.load(BytesIO(response.content), map_location=map_location, weights_only=False) |
There was a problem hiding this comment.
weights_onlydefault value changed fromFalsetoTruefrom PyTorch 2.6, so set explicitly toFalsehere, otherwisetests/pipelines/text_to_video_synthesis/test_text_to_video_zero.py::TextToVideoZeroPipelineSlowTests::test_full_modelwill raise error- set
map_locationdefault as None which aligns w.torch.load, otherwisetests/pipelines/text_to_video_synthesis/test_text_to_video_zero.py::TextToVideoZeroPipelineSlowTests::test_full_modelwill raise error
There was a problem hiding this comment.
Could we add weights_only=True as an arg to load_pt and pass in to torch load. In the test itself we can set it to false so that it's clear what's happening in the test. I would avoid doing this under the hood because it is a potential security hole.
There was a problem hiding this comment.
@DN6 , pls help review, thx.
| pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config) | ||
| generator = torch.Generator(device="cuda").manual_seed(0) | ||
| generator = torch.Generator(device="cpu").manual_seed(0) |
There was a problem hiding this comment.
PRNG will behave different across accelerators, set back to "cpu" for cross-device reproducibility, as other cases in diffusers
| @@ -220,7 +220,7 @@ def test_dict_tuple_outputs_equivalent(self, expected_max_difference=1e-4): | |||
| self.assertLess(max_diff, expected_max_difference) | |||
|
|
|||
| @unittest.skipIf(torch_device not in ["cuda", "xpu"], reason="float16 requires CUDA or XPU") | |||
| @require_accelerator | |||
| @require_torch_accelerator | |||
There was a problem hiding this comment.
actually it's torch case, use require_torch_accelerator to reflect fact
|
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. |
|
@DN6, could you pls review and comment? Thx very much. |
Signed-off-by: Matrix Yao <matrix.yao@intel.com>
incl.
stable_diffusion_k_diffusion
stable_diffusion_sag
stable_diffusion_ldm3d
text_to_video_synthesis
unclip
stable_unclip
@a-r-r-o-w @DN6 , pls help review, thx very much.