Updates for Pytorch 2.7

.github/workflows/pythonapp-min.yml

@@ -124,7 +124,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        pytorch-version: ['2.4.1', '2.5.1', '2.6.0']  # FIXME: add 'latest' back once PyTorch 2.7 issues are resolved
+        pytorch-version: ['2.4.1', '2.5.1', '2.6.0', 'latest']
     timeout-minutes: 40
     steps:
     - uses: actions/checkout@v4

docs/requirements.txt

@@ -1,5 +1,5 @@
 -f https://download.pytorch.org/whl/cpu/torch-2.3.0%2Bcpu-cp39-cp39-linux_x86_64.whl
-torch>=2.4.1, <2.7.0
+torch>=2.4.1
 pytorch-ignite==0.4.11
 numpy>=1.20
 itk>=5.2

monai/networks/layers/spatial_transforms.py

@@ -11,6 +11,7 @@
 
 from __future__ import annotations
 
+import sys
 from collections.abc import Sequence
 
 import torch
@@ -526,6 +527,11 @@ def forward(
             ValueError: When affine and image batch dimension differ.
 
         """
+
+        # In some cases it's necessary to convert inputs to grid_sample from float64 to float32 to work around known
+        # issues with PyTorch, see https://github.com/Project-MONAI/MONAI/pull/8429
+        convert_f32 = sys.platform == "win32" and src.dtype == torch.float64 and src.device == torch.device("cpu")
+
         # validate `theta`
         if not isinstance(theta, torch.Tensor):
             raise TypeError(f"theta must be torch.Tensor but is {type(theta).__name__}.")
@@ -582,11 +588,17 @@ def forward(
             )
 
         grid = nn.functional.affine_grid(theta=theta[:, :sr], size=list(dst_size), align_corners=self.align_corners)
+
+        _input = src.contiguous()
+        if convert_f32:
+            _input = _input.to(torch.float32)
+            grid = grid.to(torch.float32)
+
         dst = nn.functional.grid_sample(
-            input=src.contiguous(),
-            grid=grid,
-            mode=self.mode,
-            padding_mode=self.padding_mode,
-            align_corners=self.align_corners,
+            input=_input, grid=grid, mode=self.mode, padding_mode=self.padding_mode, align_corners=self.align_corners
         )
+
+        if convert_f32:
+            dst = dst.to(torch.float64)
+
         return dst

monai/transforms/spatial/array.py

@@ -14,6 +14,7 @@
 
 from __future__ import annotations
 
+import sys
 import warnings
 from collections.abc import Callable, Sequence
 from copy import deepcopy
@@ -2106,13 +2107,30 @@ def __call__(
             if self.norm_coords:
                 for i, dim in enumerate(img_t.shape[sr + 1 : 0 : -1]):
                     grid_t[0, ..., i] *= 2.0 / max(2, dim)
+
+            # In some cases it's necessary to convert inputs to grid_sample from float64 to float32 to work around known
+            # issues with PyTorch, see https://github.com/Project-MONAI/MONAI/pull/8429
+            convert_f32 = (
+                sys.platform == "win32" and img_t.dtype == torch.float64 and img_t.device == torch.device("cpu")
+            )
+
+            _img_t = img_t.unsqueeze(0)
+
+            if convert_f32:
+                _img_t = _img_t.to(torch.float32)
+                grid_t = grid_t.to(torch.float32)
+
             out = torch.nn.functional.grid_sample(
-                img_t.unsqueeze(0),
+                _img_t,
                 grid_t,
                 mode=_interp_mode,
                 padding_mode=_padding_mode,
                 align_corners=None if _align_corners == TraceKeys.NONE else _align_corners,  # type: ignore
             )[0]
+
+            if convert_f32:
+                out = out.to(torch.float64)
+
         out_val, *_ = convert_to_dst_type(out, dst=img, dtype=np.float32)
         return out_val
 

pyproject.toml

@@ -2,7 +2,7 @@
 requires = [
   "wheel",
   "setuptools",
-  "torch>=2.4.1, <2.7.0",
+  "torch>=2.4.1",
   "ninja",
   "packaging"
 ]

requirements-dev.txt

@@ -20,7 +20,7 @@ pyflakes
 black>=25.1.0
 isort>=5.1, <6.0
 ruff
-pytype>=2020.6.1; platform_system != "Windows"
+pytype>=2020.6.1, <=2024.4.11; platform_system != "Windows"
 types-setuptools
 mypy>=1.5.0, <1.12.0
 ninja

requirements.txt

@@ -1,2 +1,2 @@
-torch>=2.4.1, <2.7.0
+torch>=2.4.1
 numpy>=1.24,<3.0

setup.cfg

@@ -42,7 +42,7 @@ setup_requires =
     ninja
     packaging
 install_requires =
-    torch>=2.4.1, <2.7.0
+    torch>=2.4.1
     numpy>=1.24,<3.0
 
 [options.extras_require]

tests/integration/test_pad_collation.py

@@ -11,8 +11,10 @@
 
 from __future__ import annotations
 
+import os
 import random
 import unittest
+from contextlib import redirect_stderr
 from functools import wraps
 
 import numpy as np
@@ -35,7 +37,7 @@
     RandZoomd,
     ToTensor,
 )
-from monai.utils import set_determinism
+from monai.utils import first, set_determinism
 
 
 @wraps(pad_list_data_collate)
@@ -97,8 +99,9 @@ def test_pad_collation(self, t_type, collate_method, transform):
         # Default collation should raise an error
         loader_fail = DataLoader(dataset, batch_size=10)
         with self.assertRaises(RuntimeError):
-            for _ in loader_fail:
-                pass
+            # stifle PyTorch error reporting, we expect failure so don't need to look at it
+            with open(os.devnull) as f, redirect_stderr(f):
+                _ = first(loader_fail)
 
         # Padded collation shouldn't
         loader = DataLoader(dataset, batch_size=10, collate_fn=collate_method)

tests/lazy_transforms_utils.py

@@ -11,6 +11,7 @@
 
 from __future__ import annotations
 
+import sys
 from copy import deepcopy
 
 from monai.data import MetaTensor, set_track_meta
@@ -62,6 +63,13 @@ def test_resampler_lazy(
         resampler.set_random_state(seed=seed)
     set_track_meta(True)
     resampler.lazy = True
+
+    # FIXME: this is a fix for https://github.com/Project-MONAI/MONAI/pull/8429, remove when PyTorch has
+    # fixed the underlying issue
+    if sys.platform == "win32":
+        atol = 1e-4
+        rtol = 1e-4
+
     pending_output = resampler(**deepcopy(call_param))
     if output_idx is not None:
         expected_output, pending_output = (expected_output[output_idx], pending_output[output_idx])