fix: ModelBuilder resolves private hub artifacts correctly by lhnealreilly · Pull Request #5985 · aws/sagemaker-python-sdk

lhnealreilly · 2026-07-01T23:06:41Z

Fix two defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket.

Defect 1: from_jumpstart_config sets hub_name AFTER init has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: call _initialize_jumpstart_config() again after setting hub_name so hub_arn is correctly derived.

Defect 2: _build_for_jumpstart does not forward hub_arn to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn=getattr(self, 'hub_arn', None) to all get_init_kwargs call sites in the build path (_build_for_jumpstart, _detect_jumpstart_image).

Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket.

Testing:

8 unit tests covering both defects and end-to-end flow
Integration test for private hub deployment (requires env config)

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Fix three defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket. Defect 1: from_jumpstart_config sets hub_name AFTER __init__ has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: derive hub_arn inline after setting hub_name. Defect 2: _build_for_jumpstart does not forward hub_arn or sagemaker_session to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn (when set) and sagemaker_session to all get_init_kwargs call sites in the build path. Defect 3: The v3 Session class (sagemaker.core.helper.session_helper.Session) is missing hub API methods (describe_hub_content, list_hub_content_versions, list_hub_contents) that the JumpStart cache calls during hub content resolution. Fix: add these methods as thin wrappers around the boto3 sagemaker_client calls. Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket. Testing: - 8 new unit tests covering hub_arn derivation, forwarding, and e2e flow - 2 existing unit tests updated to match new call signatures

…6039) * fix(serve): support aliased hub content names in private hub deploys CreateHubContentReference accepts a custom HubContentName, but the SDK always looks up hub content by model_id, so a reference named differently from the public model id fails with ResourceNotFound. Add optional hub_content_name to JumpStartConfig and thread it through from_jumpstart_config; when set, _build_for_jumpstart resolves hub content by the reference's actual name. Verified E2E against a live private hub with an execution role that has zero S3 permissions: an aliased ModelReference now deploys successfully with SageMaker brokering artifact access through the hub content reference. * test(serve): private hub regression coverage for HubAccessConfig and aliased names Unit tests (extends the #5985 file, shared setup hoisted to a module-level _build_jumpstart_builder helper per package convention): - model_reference_arn propagation regression test guarding the #6036 fix (supersedes the subset assertions added there) - CreateModel payload chain test: drives the real unmocked path _build_for_jumpstart -> _prepare_container_def_base and asserts the container definition carries S3DataSource.HubAccessConfig.HubContentArn for hub builds and omits it for public builds -- the assertion that would have caught the v3.15.1-v3.16.0 regression - hub_content_name lookup and config threading tests Integration tests (slow_test, self-provisioning with full teardown): - test_deploy_with_no_s3_execution_role: zero-S3 execution role deploy, asserts HubAccessConfig on the created Model (service-side contract) - test_deploy_with_aliased_hub_content_name: aliased reference deploy Verified red->green at every layer: 4 unit tests and both integ tests fail on v3.16.0 baseline; all pass with the fixes (integ verified live in us-east-1 with a real hub, zero-S3 role, 2 passed in 38s). --------- Co-authored-by: Eli Davidson <elleedee@amazon.com>

lhnealreilly · 2026-07-01T23:06:41Z

Fix two defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket.

Defect 1: from_jumpstart_config sets hub_name AFTER init has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: call _initialize_jumpstart_config() again after setting hub_name so hub_arn is correctly derived.

Defect 2: _build_for_jumpstart does not forward hub_arn to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn=getattr(self, 'hub_arn', None) to all get_init_kwargs call sites in the build path (_build_for_jumpstart, _detect_jumpstart_image).

Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket.

Testing:

8 unit tests covering both defects and end-to-end flow
Integration test for private hub deployment (requires env config)

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Fix three defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket. Defect 1: from_jumpstart_config sets hub_name AFTER __init__ has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: derive hub_arn inline after setting hub_name. Defect 2: _build_for_jumpstart does not forward hub_arn or sagemaker_session to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn (when set) and sagemaker_session to all get_init_kwargs call sites in the build path. Defect 3: The v3 Session class (sagemaker.core.helper.session_helper.Session) is missing hub API methods (describe_hub_content, list_hub_content_versions, list_hub_contents) that the JumpStart cache calls during hub content resolution. Fix: add these methods as thin wrappers around the boto3 sagemaker_client calls. Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket. Testing: - 8 new unit tests covering hub_arn derivation, forwarding, and e2e flow - 2 existing unit tests updated to match new call signatures

…6039) * fix(serve): support aliased hub content names in private hub deploys CreateHubContentReference accepts a custom HubContentName, but the SDK always looks up hub content by model_id, so a reference named differently from the public model id fails with ResourceNotFound. Add optional hub_content_name to JumpStartConfig and thread it through from_jumpstart_config; when set, _build_for_jumpstart resolves hub content by the reference's actual name. Verified E2E against a live private hub with an execution role that has zero S3 permissions: an aliased ModelReference now deploys successfully with SageMaker brokering artifact access through the hub content reference. * test(serve): private hub regression coverage for HubAccessConfig and aliased names Unit tests (extends the #5985 file, shared setup hoisted to a module-level _build_jumpstart_builder helper per package convention): - model_reference_arn propagation regression test guarding the #6036 fix (supersedes the subset assertions added there) - CreateModel payload chain test: drives the real unmocked path _build_for_jumpstart -> _prepare_container_def_base and asserts the container definition carries S3DataSource.HubAccessConfig.HubContentArn for hub builds and omits it for public builds -- the assertion that would have caught the v3.15.1-v3.16.0 regression - hub_content_name lookup and config threading tests Integration tests (slow_test, self-provisioning with full teardown): - test_deploy_with_no_s3_execution_role: zero-S3 execution role deploy, asserts HubAccessConfig on the created Model (service-side contract) - test_deploy_with_aliased_hub_content_name: aliased reference deploy Verified red->green at every layer: 4 unit tests and both integ tests fail on v3.16.0 baseline; all pass with the fixes (integ verified live in us-east-1 with a real hub, zero-S3 role, 2 passed in 38s). --------- Co-authored-by: Eli Davidson <elleedee@amazon.com>

lhnealreilly · 2026-07-01T23:06:41Z

Fix two defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket.

Defect 1: from_jumpstart_config sets hub_name AFTER init has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: call _initialize_jumpstart_config() again after setting hub_name so hub_arn is correctly derived.

Defect 2: _build_for_jumpstart does not forward hub_arn to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn=getattr(self, 'hub_arn', None) to all get_init_kwargs call sites in the build path (_build_for_jumpstart, _detect_jumpstart_image).

Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket.

Testing:

8 unit tests covering both defects and end-to-end flow
Integration test for private hub deployment (requires env config)

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Fix three defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket. Defect 1: from_jumpstart_config sets hub_name AFTER __init__ has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: derive hub_arn inline after setting hub_name. Defect 2: _build_for_jumpstart does not forward hub_arn or sagemaker_session to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn (when set) and sagemaker_session to all get_init_kwargs call sites in the build path. Defect 3: The v3 Session class (sagemaker.core.helper.session_helper.Session) is missing hub API methods (describe_hub_content, list_hub_content_versions, list_hub_contents) that the JumpStart cache calls during hub content resolution. Fix: add these methods as thin wrappers around the boto3 sagemaker_client calls. Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket. Testing: - 8 new unit tests covering hub_arn derivation, forwarding, and e2e flow - 2 existing unit tests updated to match new call signatures

…6039) * fix(serve): support aliased hub content names in private hub deploys CreateHubContentReference accepts a custom HubContentName, but the SDK always looks up hub content by model_id, so a reference named differently from the public model id fails with ResourceNotFound. Add optional hub_content_name to JumpStartConfig and thread it through from_jumpstart_config; when set, _build_for_jumpstart resolves hub content by the reference's actual name. Verified E2E against a live private hub with an execution role that has zero S3 permissions: an aliased ModelReference now deploys successfully with SageMaker brokering artifact access through the hub content reference. * test(serve): private hub regression coverage for HubAccessConfig and aliased names Unit tests (extends the #5985 file, shared setup hoisted to a module-level _build_jumpstart_builder helper per package convention): - model_reference_arn propagation regression test guarding the #6036 fix (supersedes the subset assertions added there) - CreateModel payload chain test: drives the real unmocked path _build_for_jumpstart -> _prepare_container_def_base and asserts the container definition carries S3DataSource.HubAccessConfig.HubContentArn for hub builds and omits it for public builds -- the assertion that would have caught the v3.15.1-v3.16.0 regression - hub_content_name lookup and config threading tests Integration tests (slow_test, self-provisioning with full teardown): - test_deploy_with_no_s3_execution_role: zero-S3 execution role deploy, asserts HubAccessConfig on the created Model (service-side contract) - test_deploy_with_aliased_hub_content_name: aliased reference deploy Verified red->green at every layer: 4 unit tests and both integ tests fail on v3.16.0 baseline; all pass with the fixes (integ verified live in us-east-1 with a real hub, zero-S3 role, 2 passed in 38s). --------- Co-authored-by: Eli Davidson <elleedee@amazon.com>

lhnealreilly · 2026-07-01T23:06:41Z

Fix two defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket.

Defect 1: from_jumpstart_config sets hub_name AFTER init has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: call _initialize_jumpstart_config() again after setting hub_name so hub_arn is correctly derived.

Defect 2: _build_for_jumpstart does not forward hub_arn to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn=getattr(self, 'hub_arn', None) to all get_init_kwargs call sites in the build path (_build_for_jumpstart, _detect_jumpstart_image).

Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket.

Testing:

8 unit tests covering both defects and end-to-end flow
Integration test for private hub deployment (requires env config)

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Fix three defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket. Defect 1: from_jumpstart_config sets hub_name AFTER __init__ has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: derive hub_arn inline after setting hub_name. Defect 2: _build_for_jumpstart does not forward hub_arn or sagemaker_session to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn (when set) and sagemaker_session to all get_init_kwargs call sites in the build path. Defect 3: The v3 Session class (sagemaker.core.helper.session_helper.Session) is missing hub API methods (describe_hub_content, list_hub_content_versions, list_hub_contents) that the JumpStart cache calls during hub content resolution. Fix: add these methods as thin wrappers around the boto3 sagemaker_client calls. Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket. Testing: - 8 new unit tests covering hub_arn derivation, forwarding, and e2e flow - 2 existing unit tests updated to match new call signatures

…6039) * fix(serve): support aliased hub content names in private hub deploys CreateHubContentReference accepts a custom HubContentName, but the SDK always looks up hub content by model_id, so a reference named differently from the public model id fails with ResourceNotFound. Add optional hub_content_name to JumpStartConfig and thread it through from_jumpstart_config; when set, _build_for_jumpstart resolves hub content by the reference's actual name. Verified E2E against a live private hub with an execution role that has zero S3 permissions: an aliased ModelReference now deploys successfully with SageMaker brokering artifact access through the hub content reference. * test(serve): private hub regression coverage for HubAccessConfig and aliased names Unit tests (extends the #5985 file, shared setup hoisted to a module-level _build_jumpstart_builder helper per package convention): - model_reference_arn propagation regression test guarding the #6036 fix (supersedes the subset assertions added there) - CreateModel payload chain test: drives the real unmocked path _build_for_jumpstart -> _prepare_container_def_base and asserts the container definition carries S3DataSource.HubAccessConfig.HubContentArn for hub builds and omits it for public builds -- the assertion that would have caught the v3.15.1-v3.16.0 regression - hub_content_name lookup and config threading tests Integration tests (slow_test, self-provisioning with full teardown): - test_deploy_with_no_s3_execution_role: zero-S3 execution role deploy, asserts HubAccessConfig on the created Model (service-side contract) - test_deploy_with_aliased_hub_content_name: aliased reference deploy Verified red->green at every layer: 4 unit tests and both integ tests fail on v3.16.0 baseline; all pass with the fixes (integ verified live in us-east-1 with a real hub, zero-S3 role, 2 passed in 38s). --------- Co-authored-by: Eli Davidson <elleedee@amazon.com>

lhnealreilly · 2026-07-01T23:06:41Z

Fix two defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket.

Defect 1: from_jumpstart_config sets hub_name AFTER init has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: call _initialize_jumpstart_config() again after setting hub_name so hub_arn is correctly derived.

Defect 2: _build_for_jumpstart does not forward hub_arn to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn=getattr(self, 'hub_arn', None) to all get_init_kwargs call sites in the build path (_build_for_jumpstart, _detect_jumpstart_image).

Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket.

Testing:

8 unit tests covering both defects and end-to-end flow
Integration test for private hub deployment (requires env config)

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Fix three defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket. Defect 1: from_jumpstart_config sets hub_name AFTER __init__ has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: derive hub_arn inline after setting hub_name. Defect 2: _build_for_jumpstart does not forward hub_arn or sagemaker_session to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn (when set) and sagemaker_session to all get_init_kwargs call sites in the build path. Defect 3: The v3 Session class (sagemaker.core.helper.session_helper.Session) is missing hub API methods (describe_hub_content, list_hub_content_versions, list_hub_contents) that the JumpStart cache calls during hub content resolution. Fix: add these methods as thin wrappers around the boto3 sagemaker_client calls. Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket. Testing: - 8 new unit tests covering hub_arn derivation, forwarding, and e2e flow - 2 existing unit tests updated to match new call signatures

…6039) * fix(serve): support aliased hub content names in private hub deploys CreateHubContentReference accepts a custom HubContentName, but the SDK always looks up hub content by model_id, so a reference named differently from the public model id fails with ResourceNotFound. Add optional hub_content_name to JumpStartConfig and thread it through from_jumpstart_config; when set, _build_for_jumpstart resolves hub content by the reference's actual name. Verified E2E against a live private hub with an execution role that has zero S3 permissions: an aliased ModelReference now deploys successfully with SageMaker brokering artifact access through the hub content reference. * test(serve): private hub regression coverage for HubAccessConfig and aliased names Unit tests (extends the #5985 file, shared setup hoisted to a module-level _build_jumpstart_builder helper per package convention): - model_reference_arn propagation regression test guarding the #6036 fix (supersedes the subset assertions added there) - CreateModel payload chain test: drives the real unmocked path _build_for_jumpstart -> _prepare_container_def_base and asserts the container definition carries S3DataSource.HubAccessConfig.HubContentArn for hub builds and omits it for public builds -- the assertion that would have caught the v3.15.1-v3.16.0 regression - hub_content_name lookup and config threading tests Integration tests (slow_test, self-provisioning with full teardown): - test_deploy_with_no_s3_execution_role: zero-S3 execution role deploy, asserts HubAccessConfig on the created Model (service-side contract) - test_deploy_with_aliased_hub_content_name: aliased reference deploy Verified red->green at every layer: 4 unit tests and both integ tests fail on v3.16.0 baseline; all pass with the fixes (integ verified live in us-east-1 with a real hub, zero-S3 role, 2 passed in 38s). --------- Co-authored-by: Eli Davidson <elleedee@amazon.com>

lhnealreilly · 2026-07-01T23:06:41Z

Fix two defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket.

Defect 1: from_jumpstart_config sets hub_name AFTER init has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: call _initialize_jumpstart_config() again after setting hub_name so hub_arn is correctly derived.

Defect 2: _build_for_jumpstart does not forward hub_arn to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn=getattr(self, 'hub_arn', None) to all get_init_kwargs call sites in the build path (_build_for_jumpstart, _detect_jumpstart_image).

Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket.

Testing:

8 unit tests covering both defects and end-to-end flow
Integration test for private hub deployment (requires env config)

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Fix three defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket. Defect 1: from_jumpstart_config sets hub_name AFTER __init__ has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: derive hub_arn inline after setting hub_name. Defect 2: _build_for_jumpstart does not forward hub_arn or sagemaker_session to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn (when set) and sagemaker_session to all get_init_kwargs call sites in the build path. Defect 3: The v3 Session class (sagemaker.core.helper.session_helper.Session) is missing hub API methods (describe_hub_content, list_hub_content_versions, list_hub_contents) that the JumpStart cache calls during hub content resolution. Fix: add these methods as thin wrappers around the boto3 sagemaker_client calls. Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket. Testing: - 8 new unit tests covering hub_arn derivation, forwarding, and e2e flow - 2 existing unit tests updated to match new call signatures

…6039) * fix(serve): support aliased hub content names in private hub deploys CreateHubContentReference accepts a custom HubContentName, but the SDK always looks up hub content by model_id, so a reference named differently from the public model id fails with ResourceNotFound. Add optional hub_content_name to JumpStartConfig and thread it through from_jumpstart_config; when set, _build_for_jumpstart resolves hub content by the reference's actual name. Verified E2E against a live private hub with an execution role that has zero S3 permissions: an aliased ModelReference now deploys successfully with SageMaker brokering artifact access through the hub content reference. * test(serve): private hub regression coverage for HubAccessConfig and aliased names Unit tests (extends the #5985 file, shared setup hoisted to a module-level _build_jumpstart_builder helper per package convention): - model_reference_arn propagation regression test guarding the #6036 fix (supersedes the subset assertions added there) - CreateModel payload chain test: drives the real unmocked path _build_for_jumpstart -> _prepare_container_def_base and asserts the container definition carries S3DataSource.HubAccessConfig.HubContentArn for hub builds and omits it for public builds -- the assertion that would have caught the v3.15.1-v3.16.0 regression - hub_content_name lookup and config threading tests Integration tests (slow_test, self-provisioning with full teardown): - test_deploy_with_no_s3_execution_role: zero-S3 execution role deploy, asserts HubAccessConfig on the created Model (service-side contract) - test_deploy_with_aliased_hub_content_name: aliased reference deploy Verified red->green at every layer: 4 unit tests and both integ tests fail on v3.16.0 baseline; all pass with the fixes (integ verified live in us-east-1 with a real hub, zero-S3 role, 2 passed in 38s). --------- Co-authored-by: Eli Davidson <elleedee@amazon.com>

lhnealreilly · 2026-07-01T23:06:41Z

Fix two defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket.

Defect 1: from_jumpstart_config sets hub_name AFTER init has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: call _initialize_jumpstart_config() again after setting hub_name so hub_arn is correctly derived.

Defect 2: _build_for_jumpstart does not forward hub_arn to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn=getattr(self, 'hub_arn', None) to all get_init_kwargs call sites in the build path (_build_for_jumpstart, _detect_jumpstart_image).

Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket.

Testing:

8 unit tests covering both defects and end-to-end flow
Integration test for private hub deployment (requires env config)

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Fix three defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket. Defect 1: from_jumpstart_config sets hub_name AFTER __init__ has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: derive hub_arn inline after setting hub_name. Defect 2: _build_for_jumpstart does not forward hub_arn or sagemaker_session to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn (when set) and sagemaker_session to all get_init_kwargs call sites in the build path. Defect 3: The v3 Session class (sagemaker.core.helper.session_helper.Session) is missing hub API methods (describe_hub_content, list_hub_content_versions, list_hub_contents) that the JumpStart cache calls during hub content resolution. Fix: add these methods as thin wrappers around the boto3 sagemaker_client calls. Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket. Testing: - 8 new unit tests covering hub_arn derivation, forwarding, and e2e flow - 2 existing unit tests updated to match new call signatures

…6039) * fix(serve): support aliased hub content names in private hub deploys CreateHubContentReference accepts a custom HubContentName, but the SDK always looks up hub content by model_id, so a reference named differently from the public model id fails with ResourceNotFound. Add optional hub_content_name to JumpStartConfig and thread it through from_jumpstart_config; when set, _build_for_jumpstart resolves hub content by the reference's actual name. Verified E2E against a live private hub with an execution role that has zero S3 permissions: an aliased ModelReference now deploys successfully with SageMaker brokering artifact access through the hub content reference. * test(serve): private hub regression coverage for HubAccessConfig and aliased names Unit tests (extends the #5985 file, shared setup hoisted to a module-level _build_jumpstart_builder helper per package convention): - model_reference_arn propagation regression test guarding the #6036 fix (supersedes the subset assertions added there) - CreateModel payload chain test: drives the real unmocked path _build_for_jumpstart -> _prepare_container_def_base and asserts the container definition carries S3DataSource.HubAccessConfig.HubContentArn for hub builds and omits it for public builds -- the assertion that would have caught the v3.15.1-v3.16.0 regression - hub_content_name lookup and config threading tests Integration tests (slow_test, self-provisioning with full teardown): - test_deploy_with_no_s3_execution_role: zero-S3 execution role deploy, asserts HubAccessConfig on the created Model (service-side contract) - test_deploy_with_aliased_hub_content_name: aliased reference deploy Verified red->green at every layer: 4 unit tests and both integ tests fail on v3.16.0 baseline; all pass with the fixes (integ verified live in us-east-1 with a real hub, zero-S3 role, 2 passed in 38s). --------- Co-authored-by: Eli Davidson <elleedee@amazon.com>

lhnealreilly · 2026-07-01T23:06:41Z

Fix two defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket.

Defect 1: from_jumpstart_config sets hub_name AFTER init has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: call _initialize_jumpstart_config() again after setting hub_name so hub_arn is correctly derived.

Defect 2: _build_for_jumpstart does not forward hub_arn to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn=getattr(self, 'hub_arn', None) to all get_init_kwargs call sites in the build path (_build_for_jumpstart, _detect_jumpstart_image).

Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket.

Testing:

8 unit tests covering both defects and end-to-end flow
Integration test for private hub deployment (requires env config)

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Fix three defects causing ModelBuilder to ignore private hub when resolving model artifacts, forcing the execution role to access the public JumpStart S3 cache bucket. Defect 1: from_jumpstart_config sets hub_name AFTER __init__ has already called _initialize_jumpstart_config(), which takes the else branch and sets hub_arn = None. Fix: derive hub_arn inline after setting hub_name. Defect 2: _build_for_jumpstart does not forward hub_arn or sagemaker_session to get_init_kwargs, causing model data to resolve from the public catalog. Fix: pass hub_arn (when set) and sagemaker_session to all get_init_kwargs call sites in the build path. Defect 3: The v3 Session class (sagemaker.core.helper.session_helper.Session) is missing hub API methods (describe_hub_content, list_hub_content_versions, list_hub_contents) that the JumpStart cache calls during hub content resolution. Fix: add these methods as thin wrappers around the boto3 sagemaker_client calls. Impact: Customers deploying from private hubs via ModelBuilder no longer need to grant their execution role s3:GetObject on the public JumpStart cache bucket. Testing: - 8 new unit tests covering hub_arn derivation, forwarding, and e2e flow - 2 existing unit tests updated to match new call signatures

…6039) * fix(serve): support aliased hub content names in private hub deploys CreateHubContentReference accepts a custom HubContentName, but the SDK always looks up hub content by model_id, so a reference named differently from the public model id fails with ResourceNotFound. Add optional hub_content_name to JumpStartConfig and thread it through from_jumpstart_config; when set, _build_for_jumpstart resolves hub content by the reference's actual name. Verified E2E against a live private hub with an execution role that has zero S3 permissions: an aliased ModelReference now deploys successfully with SageMaker brokering artifact access through the hub content reference. * test(serve): private hub regression coverage for HubAccessConfig and aliased names Unit tests (extends the #5985 file, shared setup hoisted to a module-level _build_jumpstart_builder helper per package convention): - model_reference_arn propagation regression test guarding the #6036 fix (supersedes the subset assertions added there) - CreateModel payload chain test: drives the real unmocked path _build_for_jumpstart -> _prepare_container_def_base and asserts the container definition carries S3DataSource.HubAccessConfig.HubContentArn for hub builds and omits it for public builds -- the assertion that would have caught the v3.15.1-v3.16.0 regression - hub_content_name lookup and config threading tests Integration tests (slow_test, self-provisioning with full teardown): - test_deploy_with_no_s3_execution_role: zero-S3 execution role deploy, asserts HubAccessConfig on the created Model (service-side contract) - test_deploy_with_aliased_hub_content_name: aliased reference deploy Verified red->green at every layer: 4 unit tests and both integ tests fail on v3.16.0 baseline; all pass with the fixes (integ verified live in us-east-1 with a real hub, zero-S3 role, 2 passed in 38s). --------- Co-authored-by: Eli Davidson <elleedee@amazon.com>

lhnealreilly temporarily deployed to auto-approve July 1, 2026 23:06 — with GitHub Actions Inactive

lhnealreilly force-pushed the fix/private-hub-artifact-resolution branch from 73c3510 to c85e4c4Compare July 2, 2026 02:36

lhnealreilly temporarily deployed to auto-approve July 2, 2026 02:36 — with GitHub Actions Inactive

lhnealreilly force-pushed the fix/private-hub-artifact-resolution branch from c85e4c4 to f265dceCompare July 2, 2026 14:50

lhnealreilly temporarily deployed to auto-approve July 2, 2026 14:50 — with GitHub Actions Inactive

lhnealreilly force-pushed the fix/private-hub-artifact-resolution branch from f265dce to 5cc330bCompare July 2, 2026 15:56

lhnealreilly temporarily deployed to auto-approve July 2, 2026 15:56 — with GitHub Actions Inactive

lhnealreilly temporarily deployed to auto-approve July 2, 2026 15:57 — with GitHub Actions Inactive

lhnealreilly force-pushed the fix/private-hub-artifact-resolution branch from 5cc330b to 93c38e7Compare July 2, 2026 17:59

lhnealreilly temporarily deployed to auto-approve July 2, 2026 17:59 — with GitHub Actions Inactive

lhnealreilly force-pushed the fix/private-hub-artifact-resolution branch from 93c38e7 to ae1f6cfCompare July 2, 2026 19:47

lhnealreilly temporarily deployed to auto-approve July 2, 2026 19:47 — with GitHub Actions Inactive

jam-jee approved these changes Jul 8, 2026
View reviewed changes

jam-jee merged commit b27221b into aws:masterJul 8, 2026
15 of 24 checks passed

This was referenced Jul 16, 2026
fix(serve): private hub deploys attach HubAccessConfig and support aliased content names #6035
Closed
fix(serve): support aliased hub content names in private hub deploys #6039
Merged