Merge branch 'main' into release-docker-img

setup.py

@@ -52,7 +52,7 @@
 ]
 
 install_requires = [
-    'mosaicml[libcloud,wandb,oci,gcs,mlflow]>=0.24.1,<0.25',
+    'mosaicml[libcloud,wandb,oci,gcs,mlflow]>=0.25.0,<0.26',
     'mlflow>=2.14.1,<2.17',
     'accelerate>=0.25,<0.34',  # for HF inference `device_map`
     'transformers>=4.43.2,<4.44',
@@ -91,15 +91,15 @@
 ]
 
 extra_deps['databricks'] = [
-    'mosaicml[databricks]>=0.24.1,<0.25',
+    'mosaicml[databricks]>=0.25.0,<0.26',
     'numpy<2',
     'databricks-sql-connector>=3,<4',
     'databricks-connect==14.1.0',
     'lz4>=4,<5',
 ]
 
 extra_deps['tensorboard'] = [
-    'mosaicml[tensorboard]>=0.24.1,<0.25',
+    'mosaicml[tensorboard]>=0.25.0,<0.26',
 ]
 
 # Flash 2 group kept for backwards compatibility
@@ -110,7 +110,7 @@
 extra_deps['gpu'] = copy.deepcopy(extra_deps['gpu-flash2'])
 
 extra_deps['peft'] = [
-    'mosaicml[peft]>=0.24.1,<0.25',
+    'mosaicml[peft]>=0.25.0,<0.26',
 ]
 
 extra_deps['openai'] = [

tests/a_scripts/inference/test_convert_composer_to_hf.py

@@ -1563,6 +1563,8 @@ def test_mptmoe_huggingface_conversion_callback(
 
             # Check output equivalence
             loaded_model = loaded_model.cuda().bfloat16()  # type: ignore
+            for k, v in batch.items():
+                batch[k] = v.cuda()
             loaded_model_logits = loaded_model(
                 input_ids=batch.get('input_ids', None),
                 attention_mask=batch.get('attention_mask', None),