chore: bump version to 0.17.2; upgrade torch stack and automl defaults

- Require torch>=2.11, torchaudio>=2.11, torchvision>=0.26, transformers>=5.0
  in core deps to match torchao>=0.17.0 (which requires torch>=2.11); fixes
  LLM fine-tuning crash caused by torchao/torch version mismatch in Docker images
- Update all four Docker images to pin torch==2.12.0, torchvision==0.27.0,
  torchaudio==2.11.0 (latest stable, cu124 for GPU images)
- Drop redundant torch>=2.7 pin from the llm extra; torchao>=0.17.0 replaces old >=0.9.0
- Change AutoML default tabular combiner from tabnet to ft_transformer;
  add ft_transformer and tabtransformer to combiner_defaults with tuned hyperopt configs

Author: w4nderlust

docker/ludwig-gpu/Dockerfile

@@ -36,7 +36,7 @@ RUN if [ -n "${LUDWIG_VERSION}" ]; then \
     else \
       pip install --no-cache-dir '.[full]' --extra-index-url https://download.pytorch.org/whl/cu124; \
     fi
-RUN pip install --no-cache-dir --force-reinstall torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --extra-index-url https://download.pytorch.org/whl/cu124
+RUN pip install --no-cache-dir --force-reinstall torch==2.12.0 torchvision==0.27.0 torchaudio==2.11.0 --extra-index-url https://download.pytorch.org/whl/cu124
 
 WORKDIR /data
 

docker/ludwig-ray-gpu/Dockerfile

@@ -37,4 +37,4 @@ RUN if [ -n "${LUDWIG_VERSION}" ]; then \
     else \
       pip install --no-cache-dir '.[full]' --extra-index-url https://download.pytorch.org/whl/cu124; \
     fi
-RUN pip install --no-cache-dir --force-reinstall torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --extra-index-url https://download.pytorch.org/whl/cu124
+RUN pip install --no-cache-dir --force-reinstall torch==2.12.0 torchvision==0.27.0 torchaudio==2.11.0 --extra-index-url https://download.pytorch.org/whl/cu124

docker/ludwig-ray/Dockerfile

@@ -36,4 +36,4 @@ RUN if [ -n "${LUDWIG_VERSION}" ]; then \
     else \
       pip install --no-cache-dir '.[full]' --extra-index-url https://download.pytorch.org/whl/cpu; \
     fi
-RUN pip install --no-cache-dir --force-reinstall torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --extra-index-url https://download.pytorch.org/whl/cpu
+RUN pip install --no-cache-dir --force-reinstall torch==2.12.0 torchvision==0.27.0 torchaudio==2.11.0 --extra-index-url https://download.pytorch.org/whl/cpu

docker/ludwig/Dockerfile

@@ -32,7 +32,7 @@ RUN if [ -n "${LUDWIG_VERSION}" ]; then \
     else \
       pip install --no-cache-dir '.[full]' --extra-index-url https://download.pytorch.org/whl/cpu; \
     fi
-RUN pip install --no-cache-dir --force-reinstall torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --extra-index-url https://download.pytorch.org/whl/cpu
+RUN pip install --no-cache-dir --force-reinstall torch==2.12.0 torchvision==0.27.0 torchaudio==2.11.0 --extra-index-url https://download.pytorch.org/whl/cpu
 
 WORKDIR /data
 

ludwig/automl/automl.py

@@ -307,7 +307,7 @@ def _model_select(
 ):
     """Performs model selection based on dataset or user specified model.
 
-    Note: Current implementation returns tabnet by default for tabular datasets.
+    Note: Current implementation returns ft_transformer by default for tabular datasets.
     """
     fields = dataset_info.fields
 

ludwig/automl/base_config.py

@@ -59,6 +59,8 @@
     "concat": os.path.join(CONFIG_DIR, "combiner/concat_config.yaml"),
     "tabnet": os.path.join(CONFIG_DIR, "combiner/tabnet_config.yaml"),
     "transformer": os.path.join(CONFIG_DIR, "combiner/transformer_config.yaml"),
+    "ft_transformer": os.path.join(CONFIG_DIR, "combiner/ft_transformer_config.yaml"),
+    "tabtransformer": os.path.join(CONFIG_DIR, "combiner/tabtransformer_config.yaml"),
 }
 
 encoder_defaults = {"text": {"bert": os.path.join(CONFIG_DIR, "text/bert_config.yaml")}}

ludwig/automl/defaults/combiner/ft_transformer_config.yaml

@@ -0,0 +1,39 @@
+combiner:
+  type: ft_transformer
+  num_layers: 3
+  num_heads: 8
+  hidden_size: 64
+  transformer_output_size: 64
+  dropout: 0.1
+  fc_dropout: 0.1
+
+trainer:
+  batch_size: auto
+  learning_rate: auto
+  optimizer:
+    type: adam
+
+hyperopt:
+  parameters:
+    trainer.learning_rate:
+      space: loguniform
+      lower: 0.0001
+      upper: 0.001
+    trainer.batch_size:
+      space: choice
+      categories: [32, 64, 128, 256]
+    combiner.num_layers:
+      space: choice
+      categories: [2, 3, 4]
+    combiner.hidden_size:
+      # Must be divisible by num_heads (8). Valid values: 32, 64, 128, 256.
+      space: choice
+      categories: [32, 64, 128, 256]
+    combiner.dropout:
+      space: uniform
+      lower: 0.0
+      upper: 0.3
+    combiner.fc_dropout:
+      space: uniform
+      lower: 0.0
+      upper: 0.3

ludwig/automl/defaults/combiner/tabtransformer_config.yaml

@@ -0,0 +1,39 @@
+combiner:
+  type: tabtransformer
+  num_layers: 4
+  num_heads: 8
+  hidden_size: 64
+  transformer_output_size: 64
+  dropout: 0.1
+  fc_dropout: 0.1
+
+trainer:
+  batch_size: auto
+  learning_rate: auto
+  optimizer:
+    type: adam
+
+hyperopt:
+  parameters:
+    trainer.learning_rate:
+      space: loguniform
+      lower: 0.0001
+      upper: 0.001
+    trainer.batch_size:
+      space: choice
+      categories: [32, 64, 128, 256]
+    combiner.num_layers:
+      space: choice
+      categories: [2, 4, 6]
+    combiner.hidden_size:
+      # Must be divisible by num_heads (8). Valid values: 32, 64, 128, 256.
+      space: choice
+      categories: [32, 64, 128, 256]
+    combiner.dropout:
+      space: uniform
+      lower: 0.0
+      upper: 0.3
+    combiner.fc_dropout:
+      space: uniform
+      lower: 0.0
+      upper: 0.3

ludwig/constants.py

@@ -254,7 +254,7 @@
 IMBALANCE_DETECTION_RATIO = 0.05
 
 TABULAR = "tabular"
-AUTOML_DEFAULT_TABULAR_MODEL = "tabnet"
+AUTOML_DEFAULT_TABULAR_MODEL = "ft_transformer"
 AUTOML_DEFAULT_TEXT_ENCODER = "bert"
 AUTOML_SMALLER_TEXT_ENCODER = "distilbert"
 AUTOML_TEXT_ENCODER_MAX_TOKEN_LEN = 512

ludwig/globals.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 # ==============================================================================
 
-LUDWIG_VERSION = "0.17.1"
+LUDWIG_VERSION = "0.17.2"
 
 MODEL_FILE_NAME = "model"
 MODEL_WEIGHTS_FILE_NAME = "model_weights"  # legacy pickle format