AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-10-24 20:59:27 +03:00
Code Activity

Make Stable Diffusion device-map comparisons deterministic (#15010)

Browse Source 
Description:
The balanced device-map tests compare images from single-device and multi-device inference. On XPU, repeated fp16 inference can produce different results even with the same seed, causing the pixel comparison to fail.

Enable PyTorch deterministic algorithms for the two image-comparison tests and restore the previous setting afterward. This keeps other tests unaffected.

Validation: Both affected tests passed on a multi-XPU runner (RUN_NIGHTLY=1 RUN_SLOW=1).

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
This commit is contained in:
Wang, Yi
2026-10-12 12:29:44 +08:00
committed by GitHub
parent b1f6cde2ac
commit 883d51ef02
1 changed files with 12 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
tests/pipelines/stable_diffusion/test_stable_diffusion.py
View File

@@ -1189,6 +1189,16 @@ class TestStableDiffusionPipelineDeviceMap:
gc.collect()
backend_empty_cache(torch_device)
@pytest.fixture
def deterministic_algorithms(self):
enabled = torch.are_deterministic_algorithms_enabled()
warn_only = torch.is_deterministic_algorithms_warn_only_enabled()
torch.use_deterministic_algorithms(True)
try:
yield
finally:
torch.use_deterministic_algorithms(enabled, warn_only=warn_only)
def get_inputs(self, generator_device="cpu", seed=0):
generator = torch.Generator(device=generator_device).manual_seed(seed)
inputs = {
@@ -1212,7 +1222,7 @@ class TestStableDiffusionPipelineDeviceMap:
return no_device_map_image
def test_forward_pass_balanced_device_map(self):
def test_forward_pass_balanced_device_map(self, deterministic_algorithms):
no_device_map_image = self.get_pipeline_output_without_device_map()
sd_pipe_with_device_map = StableDiffusionPipeline.from_pretrained(
@@ -1232,7 +1242,7 @@ class TestStableDiffusionPipelineDeviceMap:
assert len(set(sd_pipe_with_device_map.hf_device_map.values())) >= 2
def test_max_memory(self):
def test_max_memory(self, deterministic_algorithms):
no_device_map_image = self.get_pipeline_output_without_device_map()
sd_pipe_with_device_map = StableDiffusionPipeline.from_pretrained(