workbench 0.8.241__tar.gz → 0.8.242__tar.gz

{workbench-0.8.241 → workbench-0.8.242}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: workbench
-Version: 0.8.241
+Version: 0.8.242
 Summary: Workbench: A Dashboard and Python API for creating and deploying AWS SageMaker Model Pipelines
 Author-email: SuperCowPowers LLC <support@supercowpowers.com>
 License: MIT License

{workbench-0.8.241 → workbench-0.8.242}/model_docker_images/ml_pipelines/Dockerfile

@@ -29,7 +29,7 @@ COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/pytho
 COPY --from=builder /usr/local/bin /usr/local/bin
 
 # Install workbench only
-RUN pip install --no-cache-dir workbench==0.8.240
+RUN pip install --no-cache-dir workbench==0.8.241
 
 # Add the script runner
 COPY ml_pipeline_runner.py /app/ml_pipeline_runner.py

{workbench-0.8.241 → workbench-0.8.242}/src/workbench/core/artifacts/endpoint_core.py

@@ -1138,7 +1138,7 @@ class EndpointCore(Artifact):
             cls.log.error("Error deleting endpoint.")
             raise e
 
-        time.sleep(5)  # Final sleep for AWS to fully register deletions
+        time.sleep(10)  # Final sleep for AWS to fully register deletions
 
     @classmethod
     def delete_endpoint_models(cls, endpoint_name: str):

{workbench-0.8.241 → workbench-0.8.242}/src/workbench/core/transforms/features_to_model/features_to_model.py

@@ -247,10 +247,10 @@ class FeaturesToModel(Transform):
         # Create a Sagemaker Model with our script
         image = ModelImages.get_image_uri(self.sm_session.boto_region_name, self.training_image)
 
-        # Use GPU instance for ChemProp/PyTorch, CPU for others
+        # Use compute-optimized CPU instance for ChemProp/PyTorch (CPU-bound due to data loading)
         if self.model_framework in [ModelFramework.CHEMPROP, ModelFramework.PYTORCH]:
-            train_instance_type = "ml.g6.xlarge"  # NVIDIA L4 GPU, ~$0.80/hr
-            self.log.important(f"Using GPU instance {train_instance_type} for {self.model_framework.value}")
+            train_instance_type = "ml.c7i.2xlarge"  # 8 vCPUs, 16GB RAM, ~$0.40/hr
+            self.log.important(f"Using compute instance {train_instance_type} for {self.model_framework.value}")
         else:
             train_instance_type = "ml.m5.xlarge"
 

{workbench-0.8.241 → workbench-0.8.242}/src/workbench/core/transforms/model_to_endpoint/model_to_endpoint.py

@@ -108,7 +108,6 @@ class ModelToEndpoint(Transform):
 
         # Is this a serverless deployment?
         serverless_config = None
-        instance_type = None
         if self.serverless:
             # For PyTorch or ChemProp we need at least 4GB of memory
             if needs_more_resources and mem_size < 4096:

{workbench-0.8.241 → workbench-0.8.242}/src/workbench/model_scripts/chemprop/chemprop.template

@@ -34,7 +34,7 @@ DEFAULT_HYPERPARAMETERS = {
     "n_folds": 5,
     "max_epochs": 400,
     "patience": 50,
-    "batch_size": 32,
+    "batch_size": 64,
     # Message Passing (ignored when using foundation model)
     "hidden_dim": 700,
     "depth": 6,
@@ -590,8 +590,9 @@ if __name__ == "__main__":
             val_dataset.normalize_targets(target_scaler)
             output_transform = nn.UnscaleTransform.from_standard_scaler(target_scaler)
 
-        train_loader = data.build_dataloader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=3)
-        val_loader = data.build_dataloader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=3)
+        num_workers = min(os.cpu_count() or 4, 8)  # Scale with CPUs, cap at 8
+        train_loader = data.build_dataloader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers, persistent_workers=num_workers > 0)
+        val_loader = data.build_dataloader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers, persistent_workers=num_workers > 0)
 
         # Build model
         pl.seed_everything(hyperparameters["seed"] + fold_idx)

{workbench-0.8.241 → workbench-0.8.242}/src/workbench.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: workbench
-Version: 0.8.241
+Version: 0.8.242
 Summary: Workbench: A Dashboard and Python API for creating and deploying AWS SageMaker Model Pipelines
 Author-email: SuperCowPowers LLC <support@supercowpowers.com>
 License: MIT License