autogluon.tabular 1.3.2b20250716__tar.gz → 1.3.2b20250717__tar.gz

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: autogluon.tabular
-Version: 1.3.2b20250716
+Version: 1.3.2b20250717
 Summary: Fast and Accurate ML in 3 Lines of Code
 Home-page: https://github.com/autogluon/autogluon
 Author: AutoGluon Community

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/mitra/_internal/config/config_run.py

@@ -29,4 +29,4 @@ class ConfigRun(ConfigSaveLoadMixin):
             seed=seed,
             model_name=model_name,
             hyperparams=hyperparams
-        )
+        )

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/mitra/_internal/config/enums.py

@@ -159,4 +159,4 @@ class MetricName(StrEnum):
 class LossName(StrEnum):
     CROSS_ENTROPY = "cross_entropy"
     MSE = "mse"
-    MAE = "mae"
+    MAE = "mae"

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/mitra/_internal/core/get_loss.py

@@ -51,4 +51,4 @@ def get_loss_pretrain(cfg: ConfigPretrain):
     elif cfg.data.task == Task.CLASSIFICATION:
         return CrossEntropyLossExtraBatch(cfg.optim.label_smoothing)
     else:
-        raise ValueError(f"Unsupported task {cfg.data.task} and (regression) loss {cfg.optim.regression_loss}")
+        raise ValueError(f"Unsupported task {cfg.data.task} and (regression) loss {cfg.optim.regression_loss}")

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/mitra/_internal/core/prediction_metrics.py

@@ -129,4 +129,4 @@ class PredictionMetricsTracker():
         y_pred = np.concatenate(self.ys_pred, axis=0)
         y_true = np.concatenate(self.ys_true, axis=0)
 
-        return PredictionMetrics.from_prediction(y_pred, y_true, self.task)
+        return PredictionMetrics.from_prediction(y_pred, y_true, self.task)

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/mitra/_internal/core/trainer_finetune.py

@@ -358,4 +358,4 @@ class TrainerFinetune(BaseEstimator):
                     f"| Train acc: {metrics_train.metrics[MetricName.ACCURACY]:.4f} "
                     f"| Val CE: {metrics_valid.metrics[MetricName.LOG_LOSS]:.4f} "
                     f"| Val acc: {metrics_valid.metrics[MetricName.ACCURACY]:.4f}"
-                ))
+                ))

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/mitra/_internal/data/preprocessor.py

@@ -417,4 +417,4 @@ class Preprocessor():
 
         x = x[:, self.new_feature_order]
 
-        return x
+        return x

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/mitra/_internal/models/tab2d.py

@@ -661,4 +661,4 @@ class MultiheadAttention(torch.nn.Module):
 
         output = self.o(output)
 
-        return output
+        return output

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/mitra/mitra_model.py

@@ -1,8 +1,14 @@
+# TODO: To ensure deterministic operations we need to set torch.use_deterministic_algorithms(True) 
+# and os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8'. The CUBLAS environment variable configures
+# the workspace size for certain CUBLAS operations to ensure reproducibility when using CUDA >= 10.2.
+# Both settings are required to ensure deterministic behavior in operations such as matrix multiplications.
+import os
+os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8'
+
 import os
 from typing import List, Optional
 
 import pandas as pd
-import torch
 
 from autogluon.common.utils.resource_utils import ResourceManager
 from autogluon.core.models import AbstractModel
@@ -41,6 +47,17 @@ class MitraModel(AbstractModel):
         num_cpus: int = 1,
         **kwargs,
     ):
+        
+        # TODO: Reset the number of threads based on the specified num_cpus
+        need_to_reset_torch_threads = False
+        torch_threads_og = None
+        if num_cpus is not None and isinstance(num_cpus, (int, float)):
+            torch_threads_og = torch.get_num_threads()
+            if torch_threads_og != num_cpus:
+                # reset torch threads back to original value after fit
+                torch.set_num_threads(num_cpus)
+                need_to_reset_torch_threads = True
+
         model_cls = self.get_model_cls()
 
         hyp = self._get_model_params()
@@ -69,6 +86,9 @@ class MitraModel(AbstractModel):
             time_limit=time_limit,
         )
 
+        if need_to_reset_torch_threads:
+            torch.set_num_threads(torch_threads_og)
+
     def _set_default_params(self):
         default_params = {
             "device": "cpu",
@@ -142,13 +162,9 @@ class MitraModel(AbstractModel):
     def _get_default_resources(self) -> tuple[int, int]:
         # Use only physical cores for better performance based on benchmarks
         num_cpus = ResourceManager.get_cpu_count(only_physical_cores=True)
-        
-        # Only request GPU if CUDA is available
-        if torch.cuda.is_available():
-            num_gpus = 1
-        else:
-            num_gpus = 0
-            
+
+        num_gpus = min(1, ResourceManager.get_gpu_count_torch(cuda_only=True))
+
         return num_cpus, num_gpus
 
     def _estimate_memory_usage(self, X: pd.DataFrame, **kwargs) -> int:
@@ -233,7 +249,7 @@ class MitraModel(AbstractModel):
         **kwargs,
     ) -> int:
         rows, features = X.shape[0], X.shape[1]
-        
+
         # For very small datasets, use a more conservative estimate
         if rows * features < 100:  # Small dataset threshold
             # Use a simpler linear formula for small datasets

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/mitra/sklearn_interface.py

@@ -1,5 +1,6 @@
 import time
 from pathlib import Path
+import contextlib
 
 import numpy as np
 import pandas as pd
@@ -311,23 +312,25 @@ class MitraClassifier(MitraBase, ClassifierMixin):
             Returns self
         """
 
-        if isinstance(X, pd.DataFrame):
-            X = X.values
-        if isinstance(y, pd.Series):
-            y = y.values
+        with mitra_deterministic_context():
 
-        self.X, self.y = X, y
+            if isinstance(X, pd.DataFrame):
+                X = X.values
+            if isinstance(y, pd.Series):
+                y = y.values
 
-        if X_val is not None and y_val is not None:
-            if isinstance(X_val, pd.DataFrame):
-                X_val = X_val.values
-            if isinstance(y_val, pd.Series):
-                y_val = y_val.values
-            X_train, X_valid, y_train, y_valid = X, X_val, y, y_val
-        else:
-            X_train, X_valid, y_train, y_valid = self._split_data(X, y)
+            self.X, self.y = X, y
 
-        return self._train_ensemble(X_train, y_train, X_valid, y_valid, self.task, DEFAULT_CLASSES, n_classes=DEFAULT_CLASSES, time_limit=time_limit)
+            if X_val is not None and y_val is not None:
+                if isinstance(X_val, pd.DataFrame):
+                    X_val = X_val.values
+                if isinstance(y_val, pd.Series):
+                    y_val = y_val.values
+                X_train, X_valid, y_train, y_valid = X, X_val, y, y_val
+            else:
+                X_train, X_valid, y_train, y_valid = self._split_data(X, y)
+
+            return self._train_ensemble(X_train, y_train, X_valid, y_valid, self.task, DEFAULT_CLASSES, n_classes=DEFAULT_CLASSES, time_limit=time_limit)
 
     def predict(self, X):
         """
@@ -363,15 +366,19 @@ class MitraClassifier(MitraBase, ClassifierMixin):
         p : ndarray of shape (n_samples, n_classes)
             The class probabilities of the input samples
         """
-        if isinstance(X, pd.DataFrame):
-            X = X.values
 
-        preds = []
-        for trainer in self.trainers:
-            logits = trainer.predict(self.X, self.y, X)[...,:len(np.unique(self.y))] # Remove extra classes
-            preds.append(np.exp(logits) / np.exp(logits).sum(axis=1, keepdims=True)) # Softmax
-        preds = sum(preds) / len(preds)  # Averaging ensemble predictions
-        return preds
+        with mitra_deterministic_context():
+
+            if isinstance(X, pd.DataFrame):
+                X = X.values
+
+            preds = []
+            for trainer in self.trainers:
+                logits = trainer.predict(self.X, self.y, X)[...,:len(np.unique(self.y))] # Remove extra classes
+                preds.append(np.exp(logits) / np.exp(logits).sum(axis=1, keepdims=True)) # Softmax
+            preds = sum(preds) / len(preds)  # Averaging ensemble predictions
+
+            return preds
 
 
 class MitraRegressor(MitraBase, RegressorMixin):
@@ -433,23 +440,25 @@ class MitraRegressor(MitraBase, RegressorMixin):
             Returns self
         """
 
-        if isinstance(X, pd.DataFrame):
-            X = X.values
-        if isinstance(y, pd.Series):
-            y = y.values
+        with mitra_deterministic_context():
 
-        self.X, self.y = X, y
+            if isinstance(X, pd.DataFrame):
+                X = X.values
+            if isinstance(y, pd.Series):
+                y = y.values
 
-        if X_val is not None and y_val is not None:
-            if isinstance(X_val, pd.DataFrame):
-                X_val = X_val.values
-            if isinstance(y_val, pd.Series):
-                y_val = y_val.values
-            X_train, X_valid, y_train, y_valid = X, X_val, y, y_val
-        else:
-            X_train, X_valid, y_train, y_valid = self._split_data(X, y)
+            self.X, self.y = X, y
+
+            if X_val is not None and y_val is not None:
+                if isinstance(X_val, pd.DataFrame):
+                    X_val = X_val.values
+                if isinstance(y_val, pd.Series):
+                    y_val = y_val.values
+                X_train, X_valid, y_train, y_valid = X, X_val, y, y_val
+            else:
+                X_train, X_valid, y_train, y_valid = self._split_data(X, y)
 
-        return self._train_ensemble(X_train, y_train, X_valid, y_valid, self.task, 1, time_limit=time_limit)
+            return self._train_ensemble(X_train, y_train, X_valid, y_valid, self.task, 1, time_limit=time_limit)
 
     def predict(self, X):
         """
@@ -465,8 +474,30 @@ class MitraRegressor(MitraBase, RegressorMixin):
         y : ndarray of shape (n_samples,)
             The predicted values
         """
-        if isinstance(X, pd.DataFrame):
-            X = X.values
 
-        preds = [trainer.predict(self.X, self.y, X) for trainer in self.trainers]
-        return sum(preds) / len(preds)  # Averaging ensemble predictions
+        with mitra_deterministic_context():
+
+            if isinstance(X, pd.DataFrame):
+                X = X.values
+            
+            preds = []
+            for trainer in self.trainers:
+                preds.append(trainer.predict(self.X, self.y, X))
+        
+            return sum(preds) / len(preds)  # Averaging ensemble predictions
+    
+
+@contextlib.contextmanager
+def mitra_deterministic_context():
+    """Context manager to set deterministic settings only for Mitra operations."""
+    
+    original_deterministic_algorithms_set = False
+
+    try:
+        torch.use_deterministic_algorithms(True)
+        original_deterministic_algorithms_set = True
+        yield
+        
+    finally:
+        if original_deterministic_algorithms_set:
+            torch.use_deterministic_algorithms(False)

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/realmlp/realmlp_model.py

@@ -274,9 +274,9 @@ class RealMLPModel(AbstractModel):
     def _get_default_resources(self) -> tuple[int, int]:
         # Use only physical cores for better performance based on benchmarks
         num_cpus = ResourceManager.get_cpu_count(only_physical_cores=True)
-        # Only request GPU if CUDA is available (RealMLP doesn't support MPS)
-        import torch
-        num_gpus = 1 if torch.cuda.is_available() else 0
+
+        num_gpus = min(1, ResourceManager.get_gpu_count_torch(cuda_only=True))
+
         return num_cpus, num_gpus
 
     def _estimate_memory_usage(self, X: pd.DataFrame, **kwargs) -> int:

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/tabicl/tabicl_model.py

@@ -111,9 +111,8 @@ class TabICLModel(AbstractModel):
     def _get_default_resources(self) -> tuple[int, int]:
         # Use only physical cores for better performance based on benchmarks
         num_cpus = ResourceManager.get_cpu_count(only_physical_cores=True)
-        # Only request GPU if CUDA is available (TabICL doesn't support MPS)
-        import torch
-        num_gpus = 1 if torch.cuda.is_available() else 0
+
+        num_gpus = min(1, ResourceManager.get_gpu_count_torch(cuda_only=True))
         return num_cpus, num_gpus
 
     def _estimate_memory_usage(self, X: pd.DataFrame, **kwargs) -> int:

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/tabm/tabm_model.py

@@ -150,9 +150,8 @@ class TabMModel(AbstractModel):
     def _get_default_resources(self) -> tuple[int, int]:
         # Use only physical cores for better performance based on benchmarks
         num_cpus = ResourceManager.get_cpu_count(only_physical_cores=True)
-        # Only request GPU if CUDA is available (TabM doesn't support other accelerators such as MPS)
-        import torch
-        num_gpus = 1 if torch.cuda.is_available() else 0
+
+        num_gpus = min(1, ResourceManager.get_gpu_count_torch(cuda_only=True))
         return num_cpus, num_gpus
 
     def _estimate_memory_usage(self, X: pd.DataFrame, **kwargs) -> int:

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/models/tabpfnv2/tabpfnv2_model.py

@@ -282,9 +282,9 @@ class TabPFNV2Model(AbstractModel):
     def _get_default_resources(self) -> tuple[int, int]:
         # Use only physical cores for better performance based on benchmarks
         num_cpus = ResourceManager.get_cpu_count(only_physical_cores=True)
-        # Only request GPU if CUDA is available (TabPFNV2 doesn't support other accelerators such as MPS)
-        import torch
-        num_gpus = 1 if torch.cuda.is_available() else 0
+
+        num_gpus = min(1, ResourceManager.get_gpu_count_torch(cuda_only=True))
+
         return num_cpus, num_gpus
 
     def _set_default_params(self):

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon/tabular/version.py

@@ -1,4 +1,4 @@
 """This is the autogluon version file."""
 
-__version__ = "1.3.2b20250716"
+__version__ = "1.3.2b20250717"
 __lite__ = False

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon.tabular.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: autogluon.tabular
-Version: 1.3.2b20250716
+Version: 1.3.2b20250717
 Summary: Fast and Accurate ML in 3 Lines of Code
 Home-page: https://github.com/autogluon/autogluon
 Author: AutoGluon Community

{autogluon.tabular-1.3.2b20250716 → autogluon.tabular-1.3.2b20250717}/src/autogluon.tabular.egg-info/requires.txt

@@ -3,21 +3,21 @@ scipy<1.17,>=1.5.4
 pandas<2.4.0,>=2.0.0
 scikit-learn<1.8.0,>=1.4.0
 networkx<4,>=3.0
-autogluon.core==1.3.2b20250716
-autogluon.features==1.3.2b20250716
+autogluon.core==1.3.2b20250717
+autogluon.features==1.3.2b20250717
 
 [all]
-lightgbm<4.7,>=4.0
-autogluon.core[all]==1.3.2b20250716
-pytabkit<1.6,>=1.5
+einops<0.9,>=0.7
+spacy<3.9
 huggingface_hub[torch]
+autogluon.core[all]==1.3.2b20250717
+lightgbm<4.7,>=4.0
 torch<2.8,>=2.2
-spacy<3.9
-numpy<2.3.0,>=1.25
+pytabkit<1.6,>=1.5
+fastai<2.9,>=2.3.1
 xgboost<3.1,>=2.0
+numpy<2.3.0,>=1.25
 catboost<1.3,>=1.2
-einops<0.9,>=0.7
-fastai<2.9,>=2.3.1
 
 [catboost]
 numpy<2.3.0,>=1.25
@@ -41,7 +41,7 @@ omegaconf
 transformers
 
 [ray]
-autogluon.core[all]==1.3.2b20250716
+autogluon.core[all]==1.3.2b20250717
 
 [realmlp]
 pytabkit<1.6,>=1.5