Can not compile an XGBoost 1.0 model

[NukaCody]

Is your feature request related to a problem? Please describe.
Can not compile an XGBoost model with version 1.0. I've been able to do it with previous versions but it breaks once you updgrade to a stable 1.0 branch

Describe the solution you'd like

compiled_model = xgb_estimator.compile_model(
    target_instance_family = "ml_m5",
    role = sage_exec,
    input_shape = {'data': [1, 64]},
    output_path = xgb_estimator.output_path,
    framework = "xgboost",
    framework_version = "1.0"
)

Should compile a 1.0 XGboost model when the model was trained with

container = get_image_uri(region, 'xgboost', repo_version='1.0-1')

xgb_estimator = Estimator(
    container,
    role = sage_exec,
    sagemaker_session = session,
    input_mode = 'File',
    output_path = f"s3://{bucket}{prefix}/output",
    
    train_instance_count = 2,
    train_instance_type = 'ml.m5.2xlarge',
    
    train_use_spot_instances = True,
    train_max_run = 300,
    train_max_wait = 600,
)

xgb_estimator.fit(s3_data)

Describe alternatives you've considered
None

Additional context
Trying to compile with what's above returns:

ClientError: InputConfiguration: Invalid XGBoost model. Please make sure the framework you select is correct. local variable 'model_objects' referenced before assignment

[ajaykarpur]

Hi @codypenta, thanks for bringing this to our attention. Can you please provide the following information?

• The version of the SageMaker Python SDK you're using. (You can check this using sagemaker.__version__.)
• The previous version of XGBoost that you used to successfully compile a model.
• Any other information from the bug report issue template that you think may be relevant.

[NukaCody]

Hope this helps:
Sagemaker Version: 1.60.0
Python Version: 3

What's worked previously

container = get_image_uri(region, 'xgboost', repo_version='0.90-2')

compiled_model = xgb.compile_model(
    target_instance_family = "ml_m5",
    role = sagemaker.get_execution_role(),
    input_shape = {'data': [1, 64]},
    output_path = f"{xgb.output_path}/compiled",
    framework = "xgboost",
    framework_version = "0.90-2"
)

I've tried 1.0 and 1.0-1 for framework version but they both returned the same error

[NukaCody]

Update:

I tried taking the model artifiact produced from:

container = get_image_uri(region, 'xgboost', repo_version='1.0-1')

xgb_estimator = Estimator(
    container,
    role = sage_exec,
    sagemaker_session = session,
    input_mode = 'File',
    output_path = f"s3://{bucket}{prefix}/output",
    
    train_instance_count = 2,
    train_instance_type = 'ml.m5.2xlarge',
    
    train_use_spot_instances = True,
    train_max_run = 300,
    train_max_wait = 600,
)

xgb_estimator.fit(s3_data)

And manually compiling it through the console. The same error exist. ClientError: InputConfiguration: Invalid XGBoost model. Please make sure the framework you select is correct. local variable 'model_objects' referenced before assignment

[akozd]

Same issue for me here. Error I'm seeing for 1.0-1:

ClientError: InputConfiguration: Treelite failed to convert XGBoost model. Please make sure the framework you select is correct. std::bad_alloc

boto3 call:

response = sm.create_compilation_job(
    CompilationJobName=job_name,
    RoleArn=get_execution_role(),
    InputConfig={
        'S3Uri': model_path,
        'DataInputConfig': '{"data":[1,10]}',
        'Framework': 'XGBOOST'
    },
    OutputConfig={
        'S3OutputLocation': output_path,
        'TargetDevice': 'ml_m5'
    },
    StoppingCondition={
        'MaxRuntimeInSeconds':3600,
        'MaxWaitTimeInSeconds': 3600
    }
)

[coverton-usgs]

I'm not sure if this relates but,...

I'm encountered the same error with sagemaker 2.26.0 when I accidentally loaded:
retrieve(framework="xgboost", region=region, version='latest')
which is discouraged according to the AWS documentation

when loading a specific container:
retrieve(framework="xgboost", region=region, version='1.2-1')

it worked just fine