Removed CPU randn() from schedulers (#8145)

Fixes performance issued due to extra CPU/GPU sync:
https://nvbugswb.nvidia.com/NvBugs5/SWBug.aspx?bugid=4904446&cmtNo=

---------

Signed-off-by: Boris Fomitchev <bfomitchev@nvidia.com>
Signed-off-by: YunLiu <55491388+KumoLiu@users.noreply.github.com>
Co-authored-by: YunLiu <55491388+KumoLiu@users.noreply.github.com>

monai/networks/schedulers/ddim.py

@@ -220,7 +220,7 @@ def step(
         if eta > 0:
             # randn_like does not support generator https://github.com/pytorch/pytorch/issues/27072
             device: torch.device = torch.device(model_output.device if torch.is_tensor(model_output) else "cpu")
-            noise = torch.randn(model_output.shape, dtype=model_output.dtype, generator=generator).to(device)
+            noise = torch.randn(model_output.shape, dtype=model_output.dtype, generator=generator, device=device)
             variance = self._get_variance(timestep, prev_timestep) ** 0.5 * eta * noise
 
             pred_prev_sample = pred_prev_sample + variance

monai/networks/schedulers/ddpm.py

@@ -241,8 +241,12 @@ def step(
         variance = 0
         if timestep > 0:
             noise = torch.randn(
-                model_output.size(), dtype=model_output.dtype, layout=model_output.layout, generator=generator
-            ).to(model_output.device)
+                model_output.size(),
+                dtype=model_output.dtype,
+                layout=model_output.layout,
+                generator=generator,
+                device=model_output.device,
+            )
             variance = (self._get_variance(timestep, predicted_variance=predicted_variance) ** 0.5) * noise
 
         pred_prev_sample = pred_prev_sample + variance

requirements-dev.txt

@@ -22,7 +22,7 @@ isort>=5.1
 ruff
 pytype>=2020.6.1; platform_system != "Windows"
 types-setuptools
-mypy>=1.5.0
+mypy>=1.5.0, <1.12.0
 ninja
 torchvision
 psutil