v1.7.3: disabled RealMLP lightning log files, removed pynvml dependency

Author: dholzmueller

README.md

@@ -200,6 +200,10 @@ and https://docs.ray.io/en/latest/cluster/vms/user-guides/community/slurm.html
 
 ## Releases (see git tags)
 
+- v1.7.3:
+    - disabled RealMLP lightning log file creation that was accidentally introduced 
+      in predict() in >=v1.7.0.
+    - removed pynvml dependency.
 - v1.7.2: 
     - Added scikit-learn 1.8 compatibility.
     - Removed debug print in RealMLP.

docs/requirements.txt

@@ -15,7 +15,6 @@ pandas>=2.0
 patool>=1.0
 probmetrics>=0.0.1
 psutil>=5.0
-pynvml>=11.0
 pytest-cov>=4.0
 pytest>=7.0
 pytorch_lightning>=2.0

pyproject.toml

@@ -90,7 +90,6 @@ hpo = [
 bench = [
     "fire", # argparse utilities
     "ray>=2.8", # parallelization
-    "pynvml>=11.0", # NVIDIA GPU utilization
     "openml>=0.14", # OpenML data download
     # ----- UCI import ------
     "requests>=2.0",

pytabkit/__about__.py

@@ -2,4 +2,4 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 
-__version__ = "1.7.2"
+__version__ = "1.7.3"

pytabkit/bench/scheduling/execution.py

@@ -62,21 +62,6 @@ def measure_node_resources(node_id: int) -> Tuple[NodeResources, NodeResources]:
         # alloc dummy tensors to know how much memory PyTorch uses for its runtime
         dummy_tensors = [torch.ones(1).to(f'cuda:{i}') for i in range(n_gpus)]
         gpu_rams_gb, gpu_rams_fixed_gb = get_gpu_rams_gb()
-        # import pynvml
-        # pynvml.nvmlInit()
-        #
-        # gpu_rams_gb = []
-        # gpu_rams_fixed_gb = []
-        #
-        # for i in range(n_gpus):
-        #     # adapted torch.cuda.list_gpu_processes(gpu)
-        #     h = pynvml.nvmlDeviceGetHandleByIndex(i)
-        #     info = pynvml.nvmlDeviceGetMemoryInfo(h)
-        #     total = info.total
-        #     # print(f'free     : {info.free}')
-        #     used = info.used
-        #     gpu_rams_gb.append(total / (1024. ** 3))
-        #     gpu_rams_fixed_gb.append(used / (1024. ** 3))
     else:
         gpu_rams_gb = []
         gpu_rams_fixed_gb = []

pytabkit/models/alg_interfaces/nn_interfaces.py

@@ -166,7 +166,8 @@ def predict(self, ds: DictDataset) -> torch.Tensor:
 
         # create new trainer so we don't have to pickle the full trainer that references the dataset somehow
         # update devices since the model device may have been moved since
-        trainer = pl.Trainer(**(self.min_trainer_kwargs | dict(accelerator=pl_accelerator, devices=pl_devices)))
+        trainer = pl.Trainer(**(self.min_trainer_kwargs | dict(accelerator=pl_accelerator, devices=pl_devices,
+                                                               logger=pl.loggers.logger.DummyLogger())))
         y_pred = trainer.predict(model=self.model, dataloaders=self.model.get_predict_dataloader(ds_x))
         y_pred = cat_if_necessary(y_pred, dim=-2).to('cpu')  # concat along batch dimension
         y_pred = postprocess_multiquantile(y_pred, **self.config)  # postprocessing in case of multiquantile loss