Bug in AzureML cloud docs

[ncclementi]

In this page https://docs.rapids.ai/deployment/nightly/cloud/azure/azureml/, deployment works, but when running the first import

from azure.ai.ml import MLClient

we get this

File /anaconda/envs/rapids/lib/python3.12/site-packages/azure/ai/ml/_schema/_data_import/data_import.py:9
      1 # ---------------------------------------------------------
      2 # Copyright (c) Microsoft Corporation. All rights reserved.
      3 # ---------------------------------------------------------
      4 
      5 # pylint: disable=unused-argument
      7 from marshmallow import post_load
----> 9 from azure.ai.ml._schema.core.fields import NestedField
     10 from azure.ai.ml._schema.job.input_output_entry import DatabaseSchema, FileSystemSchema
     11 from azure.ai.ml._utils._experimental import experimental

File /anaconda/envs/rapids/lib/python3.12/site-packages/azure/ai/ml/_schema/core/fields.py:19
     17 from marshmallow import RAISE, fields
     18 from marshmallow.exceptions import ValidationError
---> 19 from marshmallow.fields import _T, Field, Nested
     20 from marshmallow.utils import FieldInstanceResolutionError, from_iso_datetime, resolve_field_instance
     22 from ..._utils._arm_id_utils import AMLVersionedArmId, is_ARM_id_for_resource, parse_name_label, parse_name_version

ImportError: cannot import name '_T' from 'marshmallow.fields' (/anaconda/envs/rapids/lib/python3.12/site-packages/marshmallow/fields.py)

This seems to be a known issue Azure/azure-sdk-for-python#39047 and it's technically fixed in Azure/azure-sdk-for-python#39053, but I it doesn't look like there has been another release in conda, last one was in December https://anaconda.org/Microsoft/azure-ai-ml/files

The PR in the conda forge feedstock is broken conda-forge/azure-ai-ml-feedstock#29 not sure who can I ping to take a look at it.

But they are cutting releases in pypi see https://pypi.org/project/azure-ai-ml/

I wonder if we can modify the .sh script to install those from pip? any suggestions?

[jacobtomlinson]

I see no problem switching to a pip install to work around this.