[rsareddy0329]

Issue

canaries-v3-master (sagemaker-train) has been failing with 15 failures, all in tests/integ/train/test_recipe_override_integration.py. None are SDK regressions — the JumpStart Hub SFT recipe for meta-textgeneration-llama-3-2-1b-instruct was updated (hub content 2.17.0, template v1.0.3) and the tests pinned the old recipe shape:

• training_config.training_args.* nesting removed → 10× KeyError: 'training_args'
• max_len renamed to dataset_max_len → KeyError: 'max_len' / max_len missing from [...]
• default learning_rate changed 5e-06 → 0.0001 → assert '0.0001' == '5e-06'
• validation bounds relaxed → 2× DID NOT RAISE ValueError

These tests assert a live, externally-owned artifact, so any Hub recipe bump breaks the canary even when the SDK is correct.

Fix

• Remove the 15 structure-pinning SFT tests. Their merge/override/flatten/validation behavior is already covered Hub-independently by the pinned-fixture unit tests in tests/unit/train/test_recipe_resolver.py (TestRecipeResolver, TestFullRecipeTemplate, TestFlattenResolvedRecipe, TestValidateValue(Extended), TestValidateStepConstraints, TestRecipeResolverInstanceTypeValidation).
• Add test_sft_llama_serverless_only_user_override_keys_applied, mirroring the existing test_sft_nova_serverless_only_user_override_keys_applied / test_rlvr_serverless_only_user_override_keys_applied: override keys are discovered dynamically from the live Hub spec (hyperparameters._specs) and overrides are asserted by value, so the test survives recipe leaf renames (e.g. learning_rate → training_config.optim.lr) and future drift.
• Untouched: the already recipe-proof suites — TestModelTrainerRecipeOverrideInteg (local recipe), TestBenchMarkEvaluatorRecipeOverrideInteg, and the existing RLVR/Nova serverless tests.

Net: one consistent, recipe-proof pattern across RLVR, Nova SFT, and llama SFT. Tests 35 → 21 (+100 / −459).

Testing

• black (line-length 100) and flake8 pass on the updated file.
• Resolver logic unit tests (tests/unit/train/test_recipe_resolver.py) unchanged and already green.

Follow-up (not in this PR)

_apply_recipe_to_hyperparameters emits recipe leaf names (lr, total_epochs, max_length) rather than user-facing spec names (learning_rate, max_epochs, dataset_max_len) for serverless hyperparameters. Worth confirming with the recipe owners that this is what the training container expects.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[rsareddy0329]

Ran the sagemaker-train integ tests for the changed file on this PR head (3f87292) via CodeBuild — SUCCEEDED (18 passed, 2 skipped):
│ arn:aws:codebuild:us-west-2::build/sagemaker-python-sdk-ci-sagemaker-train-integ-tests:3afe119a-2677-463e-954f-988f341d4d21

[github-actions]

🤖 AI code review

This PR is test-only: it removes 14 integration tests in test_recipe_override_integration.py that hard-pinned the JumpStart Hub SFT recipe shape (nesting under training_config.training_args, max_len, fixed learning_rate), and adds test_sft_llama_serverless_only_user_override_keys_applied. The new test mirrors the existing Nova/RLVR serverless-override tests, discovers override keys dynamically from the live Hub spec (hyperparameters._specs), and asserts overrides by value via flatten_resolved_recipe, making it robust to recipe leaf renames. The new test is internally consistent, uses imports already present (TrainingJobCompute, flatten_resolved_recipe), and faithfully follows the established pattern. No correctness, security, or backward-compatibility concerns; looks clean.

_{Reviewed commit 3f87292. Automated review; verify before acting.}