Bump Numpy version limit to include 2.x only + Pytorch 2.6

.github/workflows/pythonapp-min.yml

@@ -124,7 +124,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        pytorch-version: ['1.13.1', '2.0.1', '2.2.2', '2.3.1', '2.4.1', 'latest']
+        pytorch-version: ['1.13.1', '2.0.1', '2.2.2', '2.3.1', '2.4.1', '2.5.1', 'latest']
     timeout-minutes: 40
     steps:
     - uses: actions/checkout@v4

environment-dev.yml

@@ -5,7 +5,7 @@ channels:
   - nvidia
   - conda-forge
 dependencies:
-  - numpy>=1.24,<2.0
+  - numpy>=1.24,<3.0
   - pytorch>=1.13.1
   - torchio
   - torchvision

monai/apps/deepedit/interaction.py

@@ -72,7 +72,7 @@ def __call__(self, engine: SupervisedTrainer | SupervisedEvaluator, batchdata: d
 
                 with torch.no_grad():
                     if engine.amp:
-                        with torch.cuda.amp.autocast():
+                        with torch.autocast("cuda"):
                             predictions = engine.inferer(inputs, engine.network)
                     else:
                         predictions = engine.inferer(inputs, engine.network)

monai/apps/deepgrow/interaction.py

@@ -67,7 +67,7 @@ def __call__(self, engine: SupervisedTrainer | SupervisedEvaluator, batchdata: d
             engine.network.eval()
             with torch.no_grad():
                 if engine.amp:
-                    with torch.cuda.amp.autocast():
+                    with torch.autocast("cuda"):
                         predictions = engine.inferer(inputs, engine.network)
                 else:
                     predictions = engine.inferer(inputs, engine.network)

monai/apps/detection/networks/retinanet_detector.py

@@ -180,7 +180,7 @@ def forward(self, images: torch.Tensor):
                 nesterov=True,
             )
             torch.save(detector.network.state_dict(), 'model.pt')  # save model
-            detector.network.load_state_dict(torch.load('model.pt'))  # load model
+            detector.network.load_state_dict(torch.load('model.pt', weights_only=True))  # load model
     """
 
     def __init__(

monai/apps/mmars/mmars.py

@@ -241,7 +241,7 @@ def load_from_mmar(
         return torch.jit.load(_model_file, map_location=map_location)
 
     # loading with `torch.load`
-    model_dict = torch.load(_model_file, map_location=map_location)
+    model_dict = torch.load(_model_file, map_location=map_location, weights_only=True)
     if weights_only:
         return model_dict.get(model_key, model_dict)  # model_dict[model_key] or model_dict directly
 

monai/bundle/scripts.py

@@ -760,7 +760,7 @@ def load(
     if load_ts_module is True:
         return load_net_with_metadata(full_path, map_location=torch.device(device), more_extra_files=config_files)
     # loading with `torch.load`
-    model_dict = torch.load(full_path, map_location=torch.device(device))
+    model_dict = torch.load(full_path, map_location=torch.device(device), weights_only=True)
 
     if not isinstance(model_dict, Mapping):
         warnings.warn(f"the state dictionary from {full_path} should be a dictionary but got {type(model_dict)}.")
@@ -1279,9 +1279,8 @@ def verify_net_in_out(
         if input_dtype == torch.float16:
             # fp16 can only be executed in gpu mode
             net.to("cuda")
-            from torch.cuda.amp import autocast
 
-            with autocast():
+            with torch.autocast("cuda"):
                 output = net(test_data.cuda(), **extra_forward_args_)
             net.to(device_)
         else:
@@ -1330,7 +1329,7 @@ def _export(
         # here we use ignite Checkpoint to support nested weights and be compatible with MONAI CheckpointSaver
         Checkpoint.load_objects(to_load={key_in_ckpt: net}, checkpoint=ckpt_file)
     else:
-        ckpt = torch.load(ckpt_file)
+        ckpt = torch.load(ckpt_file, weights_only=True)
         copy_model_state(dst=net, src=ckpt if key_in_ckpt == "" else ckpt[key_in_ckpt])
 
     # Use the given converter to convert a model and save with metadata, config content

monai/data/dataset.py

@@ -22,7 +22,6 @@
 import warnings
 from collections.abc import Callable, Sequence
 from copy import copy, deepcopy
-from inspect import signature
 from multiprocessing.managers import ListProxy
 from multiprocessing.pool import ThreadPool
 from pathlib import Path
@@ -372,10 +371,7 @@ def _cachecheck(self, item_transformed):
 
         if hashfile is not None and hashfile.is_file():  # cache hit
             try:
-                if "weights_only" in signature(torch.load).parameters:
-                    return torch.load(hashfile, weights_only=False)
-                else:
-                    return torch.load(hashfile)
+                return torch.load(hashfile, weights_only=False)
             except PermissionError as e:
                 if sys.platform != "win32":
                     raise e
@@ -1674,7 +1670,4 @@ def _load_meta_cache(self, meta_hash_file_name):
         if meta_hash_file_name in self._meta_cache:
             return self._meta_cache[meta_hash_file_name]
         else:
-            if "weights_only" in signature(torch.load).parameters:
-                return torch.load(self.cache_dir / meta_hash_file_name, weights_only=False)
-            else:
-                return torch.load(self.cache_dir / meta_hash_file_name)
+            return torch.load(self.cache_dir / meta_hash_file_name, weights_only=False)

monai/engines/evaluator.py

@@ -82,8 +82,8 @@ class Evaluator(Workflow):
             default to `True`.
         to_kwargs: dict of other args for `prepare_batch` API when converting the input data, except for
             `device`, `non_blocking`.
-        amp_kwargs: dict of the args for `torch.cuda.amp.autocast()` API, for more details:
-            https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.autocast.
+        amp_kwargs: dict of the args for `torch.autocast("cuda")` API, for more details:
+            https://pytorch.org/docs/stable/amp.html#torch.autocast.
 
     """
 
@@ -214,8 +214,8 @@ class SupervisedEvaluator(Evaluator):
             default to `True`.
         to_kwargs: dict of other args for `prepare_batch` API when converting the input data, except for
             `device`, `non_blocking`.
-        amp_kwargs: dict of the args for `torch.cuda.amp.autocast()` API, for more details:
-            https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.autocast.
+        amp_kwargs: dict of the args for `torch.autocast("cuda")` API, for more details:
+            https://pytorch.org/docs/stable/amp.html#torch.autocast.
         compile: whether to use `torch.compile`, default is False. If True, MetaTensor inputs will be converted to
             `torch.Tensor` before forward pass,  then converted back afterward with copied meta information.
         compile_kwargs: dict of the args for `torch.compile()` API, for more details:
@@ -329,7 +329,7 @@ def _iteration(self, engine: SupervisedEvaluator, batchdata: dict[str, torch.Ten
         # execute forward computation
         with engine.mode(engine.network):
             if engine.amp:
-                with torch.cuda.amp.autocast(**engine.amp_kwargs):
+                with torch.autocast("cuda", **engine.amp_kwargs):
                     engine.state.output[Keys.PRED] = engine.inferer(inputs, engine.network, *args, **kwargs)
             else:
                 engine.state.output[Keys.PRED] = engine.inferer(inputs, engine.network, *args, **kwargs)
@@ -399,8 +399,8 @@ class EnsembleEvaluator(Evaluator):
             default to `True`.
         to_kwargs: dict of other args for `prepare_batch` API when converting the input data, except for
             `device`, `non_blocking`.
-        amp_kwargs: dict of the args for `torch.cuda.amp.autocast()` API, for more details:
-            https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.autocast.
+        amp_kwargs: dict of the args for `torch.autocast("cuda")` API, for more details:
+            https://pytorch.org/docs/stable/amp.html#torch.autocast.
 
     """
 
@@ -492,7 +492,7 @@ def _iteration(self, engine: EnsembleEvaluator, batchdata: dict[str, torch.Tenso
         for idx, network in enumerate(engine.networks):
             with engine.mode(network):
                 if engine.amp:
-                    with torch.cuda.amp.autocast(**engine.amp_kwargs):
+                    with torch.autocast("cuda", **engine.amp_kwargs):
                         if isinstance(engine.state.output, dict):
                             engine.state.output.update(
                                 {engine.pred_keys[idx]: engine.inferer(inputs, network, *args, **kwargs)}

monai/engines/trainer.py

@@ -126,8 +126,8 @@ class SupervisedTrainer(Trainer):
             more details: https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html.
         to_kwargs: dict of other args for `prepare_batch` API when converting the input data, except for
             `device`, `non_blocking`.
-        amp_kwargs: dict of the args for `torch.cuda.amp.autocast()` API, for more details:
-            https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.autocast.
+        amp_kwargs: dict of the args for `torch.autocast("cuda")` API, for more details:
+            https://pytorch.org/docs/stable/amp.html#torch.autocast.
         compile: whether to use `torch.compile`, default is False. If True, MetaTensor inputs will be converted to
             `torch.Tensor` before forward pass,  then converted back afterward with copied meta information.
         compile_kwargs: dict of the args for `torch.compile()` API, for more details:
@@ -255,7 +255,7 @@ def _compute_pred_loss():
         engine.optimizer.zero_grad(set_to_none=engine.optim_set_to_none)
 
         if engine.amp and engine.scaler is not None:
-            with torch.cuda.amp.autocast(**engine.amp_kwargs):
+            with torch.autocast("cuda", **engine.amp_kwargs):
                 _compute_pred_loss()
             engine.scaler.scale(engine.state.output[Keys.LOSS]).backward()
             engine.fire_event(IterationEvents.BACKWARD_COMPLETED)
@@ -341,8 +341,8 @@ class GanTrainer(Trainer):
             more details: https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html.
         to_kwargs: dict of other args for `prepare_batch` API when converting the input data, except for
             `device`, `non_blocking`.
-        amp_kwargs: dict of the args for `torch.cuda.amp.autocast()` API, for more details:
-            https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.autocast.
+        amp_kwargs: dict of the args for `torch.autocast("cuda")` API, for more details:
+            https://pytorch.org/docs/stable/amp.html#torch.autocast.
 
     """
 
@@ -518,8 +518,8 @@ class AdversarialTrainer(Trainer):
             more details: https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html.
         to_kwargs: dict of other args for `prepare_batch` API when converting the input data, except for
             `device`, `non_blocking`.
-        amp_kwargs: dict of the args for `torch.cuda.amp.autocast()` API, for more details:
-            https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.autocast.
+        amp_kwargs: dict of the args for `torch.autocast("cuda")` API, for more details:
+            https://pytorch.org/docs/stable/amp.html#torch.autocast.
     """
 
     def __init__(
@@ -689,7 +689,7 @@ def _compute_generator_loss() -> None:
         engine.state.g_optimizer.zero_grad(set_to_none=engine.optim_set_to_none)
 
         if engine.amp and engine.state.g_scaler is not None:
-            with torch.cuda.amp.autocast(**engine.amp_kwargs):
+            with torch.autocast("cuda", **engine.amp_kwargs):
                 _compute_generator_loss()
 
             engine.state.output[Keys.LOSS] = (
@@ -737,7 +737,7 @@ def _compute_discriminator_loss() -> None:
         engine.state.d_network.zero_grad(set_to_none=engine.optim_set_to_none)
 
         if engine.amp and engine.state.d_scaler is not None:
-            with torch.cuda.amp.autocast(**engine.amp_kwargs):
+            with torch.autocast("cuda", **engine.amp_kwargs):
                 _compute_discriminator_loss()
 
             engine.state.d_scaler.scale(engine.state.output[AdversarialKeys.DISCRIMINATOR_LOSS]).backward()

monai/engines/workflow.py

@@ -90,8 +90,8 @@ class Workflow(Engine):
             default to `True`.
         to_kwargs: dict of other args for `prepare_batch` API when converting the input data, except for
             `device`, `non_blocking`.
-        amp_kwargs: dict of the args for `torch.cuda.amp.autocast()` API, for more details:
-            https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.autocast.
+        amp_kwargs: dict of the args for `torch.autocast("cuda")` API, for more details:
+            https://pytorch.org/docs/stable/amp.html#torch.autocast.
 
     Raises:
         TypeError: When ``data_loader`` is not a ``torch.utils.data.DataLoader``.

monai/fl/client/monai_algo.py

@@ -574,7 +574,7 @@ def get_weights(self, extra=None):
                 model_path = os.path.join(self.bundle_root, cast(str, self.model_filepaths[model_type]))
                 if not os.path.isfile(model_path):
                     raise ValueError(f"No best model checkpoint exists at {model_path}")
-                weights = torch.load(model_path, map_location="cpu")
+                weights = torch.load(model_path, map_location="cpu", weights_only=True)
                 # if weights contain several state dicts, use the one defined by `save_dict_key`
                 if isinstance(weights, dict) and self.save_dict_key in weights:
                     weights = weights.get(self.save_dict_key)

monai/handlers/checkpoint_loader.py

@@ -122,7 +122,7 @@ def __call__(self, engine: Engine) -> None:
         Args:
             engine: Ignite Engine, it can be a trainer, validator or evaluator.
         """
-        checkpoint = torch.load(self.load_path, map_location=self.map_location)
+        checkpoint = torch.load(self.load_path, map_location=self.map_location, weights_only=False)
 
         k, _ = list(self.load_dict.items())[0]
         # single object and checkpoint is directly a state_dict

monai/losses/perceptual.py

@@ -374,7 +374,7 @@ def __init__(
         else:
             network = torchvision.models.resnet50(weights=None)
             if pretrained is True:
-                state_dict = torch.load(pretrained_path)
+                state_dict = torch.load(pretrained_path, weights_only=True)
                 if pretrained_state_dict_key is not None:
                     state_dict = state_dict[pretrained_state_dict_key]
                 network.load_state_dict(state_dict)

monai/networks/layers/vector_quantizer.py

@@ -100,7 +100,7 @@ def quantize(self, inputs: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, to
             torch.Tensor: Quantization indices of shape [B,H,W,D,1]
 
         """
-        with torch.cuda.amp.autocast(enabled=False):
+        with torch.autocast("cuda", enabled=False):
             encoding_indices_view = list(inputs.shape)
             del encoding_indices_view[1]
 
@@ -138,7 +138,7 @@ def embed(self, embedding_indices: torch.Tensor) -> torch.Tensor:
         Returns:
             torch.Tensor: Quantize space representation of encoding_indices in channel first format.
         """
-        with torch.cuda.amp.autocast(enabled=False):
+        with torch.autocast("cuda", enabled=False):
             embedding: torch.Tensor = (
                 self.embedding(embedding_indices).permute(self.quantization_permutation).contiguous()
             )

monai/networks/nets/hovernet.py

@@ -633,9 +633,9 @@ def _remap_preact_resnet_model(model_url: str):
     # download the pretrained weights into torch hub's default dir
     weights_dir = os.path.join(torch.hub.get_dir(), "preact-resnet50.pth")
     download_url(model_url, fuzzy=True, filepath=weights_dir, progress=False)
-    state_dict = torch.load(weights_dir, map_location=None if torch.cuda.is_available() else torch.device("cpu"))[
-        "desc"
-    ]
+    map_location = None if torch.cuda.is_available() else torch.device("cpu")
+    state_dict = torch.load(weights_dir, map_location=map_location, weights_only=True)["desc"]
+
     for key in list(state_dict.keys()):
         new_key = None
         if pattern_conv0.match(key):
@@ -668,7 +668,8 @@ def _remap_standard_resnet_model(model_url: str, state_dict_key: str | None = No
     # download the pretrained weights into torch hub's default dir
     weights_dir = os.path.join(torch.hub.get_dir(), "resnet50.pth")
     download_url(model_url, fuzzy=True, filepath=weights_dir, progress=False)
-    state_dict = torch.load(weights_dir, map_location=None if torch.cuda.is_available() else torch.device("cpu"))
+    map_location = None if torch.cuda.is_available() else torch.device("cpu")
+    state_dict = torch.load(weights_dir, map_location=map_location, weights_only=True)
     if state_dict_key is not None:
         state_dict = state_dict[state_dict_key]
 

monai/networks/nets/resnet.py

@@ -493,7 +493,7 @@ def _resnet(
         if isinstance(pretrained, str):
             if Path(pretrained).exists():
                 logger.info(f"Loading weights from {pretrained}...")
-                model_state_dict = torch.load(pretrained, map_location=device)
+                model_state_dict = torch.load(pretrained, map_location=device, weights_only=True)
             else:
                 # Throw error
                 raise FileNotFoundError("The pretrained checkpoint file is not found")
@@ -665,7 +665,7 @@ def get_pretrained_resnet_medicalnet(resnet_depth: int, device: str = "cpu", dat
                 raise EntryNotFoundError(
                     f"{filename} not found on {medicalnet_huggingface_repo_basename}{resnet_depth}"
                 ) from None
-        checkpoint = torch.load(pretrained_path, map_location=torch.device(device))
+        checkpoint = torch.load(pretrained_path, map_location=torch.device(device), weights_only=True)
     else:
         raise NotImplementedError("Supported resnet_depth are: [10, 18, 34, 50, 101, 152, 200]")
     logger.info(f"{filename} downloaded")

monai/networks/nets/senet.py

@@ -302,7 +302,7 @@ def _load_state_dict(model: nn.Module, arch: str, progress: bool):
 
     if isinstance(model_url, dict):
         download_url(model_url["url"], filepath=model_url["filename"])
-        state_dict = torch.load(model_url["filename"], map_location=None)
+        state_dict = torch.load(model_url["filename"], map_location=None, weights_only=True)
     else:
         state_dict = load_state_dict_from_url(model_url, progress=progress)
     for key in list(state_dict.keys()):

monai/networks/nets/swin_unetr.py

@@ -1118,7 +1118,7 @@ def filter_swinunetr(key, value):
         )
         ssl_weights_path = "./ssl_pretrained_weights.pth"
         download_url(resource, ssl_weights_path)
-        ssl_weights = torch.load(ssl_weights_path)["model"]
+        ssl_weights = torch.load(ssl_weights_path, weights_only=True)["model"]
 
         dst_dict, loaded, not_loaded = copy_model_state(model, ssl_weights, filter_func=filter_swinunetr)
 

monai/networks/nets/transchex.py

@@ -68,7 +68,8 @@ def from_pretrained(
         weights_path = cached_file(path_or_repo_id, filename, cache_dir=cache_dir)
         model = cls(num_language_layers, num_vision_layers, num_mixed_layers, bert_config, *inputs, **kwargs)
         if state_dict is None and not from_tf:
-            state_dict = torch.load(weights_path, map_location="cpu" if not torch.cuda.is_available() else None)
+            map_location = "cpu" if not torch.cuda.is_available() else None
+            state_dict = torch.load(weights_path, map_location=map_location, weights_only=True)
         if from_tf:
             return load_tf_weights_in_bert(model, weights_path)
         old_keys = []

monai/networks/utils.py

@@ -1238,7 +1238,7 @@ def __init__(self, mod):
 
     def forward(self, x):
         dtype = x.dtype
-        with torch.amp.autocast("cuda", enabled=False):
+        with torch.autocast("cuda", enabled=False):
             ret = self.mod.forward(x.to(torch.float32)).to(dtype)
         return ret
 
@@ -1255,7 +1255,7 @@ def __init__(self, mod):
 
     def forward(self, *args):
         from_dtype = args[0].dtype
-        with torch.amp.autocast("cuda", enabled=False):
+        with torch.autocast("cuda", enabled=False):
             ret = self.mod.forward(*cast_all(args, from_dtype=from_dtype, to_dtype=torch.float32))
         return cast_all(ret, from_dtype=torch.float32, to_dtype=from_dtype)