Resnet only use contiguous in training mode. (#12977)

* fix contiguous

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* update tol

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* bigger tol

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix tests

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* update tol

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

---------

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

src/diffusers/models/resnet.py

@@ -366,7 +366,12 @@ class ResnetBlock2D(nn.Module):
         hidden_states = self.conv2(hidden_states)
 
         if self.conv_shortcut is not None:
-            input_tensor = self.conv_shortcut(input_tensor.contiguous())
+            # Only use contiguous() during training to avoid DDP gradient stride mismatch warning.
+            # In inference mode (eval or no_grad), skip contiguous() for better performance, especially on CPU.
+            # Issue: https://github.com/huggingface/diffusers/issues/12975
+            if self.training:
+                input_tensor = input_tensor.contiguous()
+            input_tensor = self.conv_shortcut(input_tensor)
 
         output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
 

tests/pipelines/kandinsky2_2/test_kandinsky_inpaint.py

@@ -248,6 +248,9 @@ class KandinskyV22InpaintPipelineFastTests(PipelineTesterMixin, unittest.TestCas
     def test_float16_inference(self):
         super().test_float16_inference(expected_max_diff=5e-1)
 
+    def test_save_load_dduf(self):
+        super().test_save_load_dduf(atol=1e-3, rtol=1e-3)
+
     @is_flaky()
     def test_model_cpu_offload_forward_pass(self):
         super().test_inference_batch_single_identical(expected_max_diff=8e-4)

tests/pipelines/kandinsky3/test_kandinsky3_img2img.py

@@ -191,6 +191,9 @@ class Kandinsky3Img2ImgPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
     def test_inference_batch_single_identical(self):
         super().test_inference_batch_single_identical(expected_max_diff=1e-2)
 
+    def test_save_load_dduf(self):
+        super().test_save_load_dduf(atol=1e-3, rtol=1e-3)
+
 
 @slow
 @require_torch_accelerator