feat(train): Auto-detect subscription recipe hyperparameters in SFTTr…

[haardm]

Auto-detect subscription recipe hyperparameters in SFTTrainer

Issue

When using SFTTrainer with models that have subscription-gated recipes (e.g., Nova Forge datamix), the trainer's hyperparameter schema (_specs) only includes keys from the standard recipe. Subscription recipe hyperparameters like customer_data_percent and nova_*_percent are rejected with AttributeError.

This is because the subscription recipe's SmtjOverrideParamsS3Uri points to an S3 access point that requires the customer to be subscribed. The SDK currently only downloads override_params from the standard (publicly accessible) recipe.

Solution

After loading the standard recipe's override_params, auto-detect if any recipe in RecipeCollection has IsSubscriptionModel: true. If found, attempt to download its override_params from the S3 access point using the customer's credentials and merge the additional keys into _specs with default: None.

Key behaviors:

• Subscription recipe keys are added to _specs (settable) but with default: None
• Keys with None default are NOT serialized in to_dict() unless explicitly set by the user
• This ensures non-datamix jobs don't accidentally send datamix HPs
• For non-subscribed users, the download fails silently (AccessDenied) — no behavior change

Changes

• sagemaker-train/src/sagemaker/train/common_utils/finetune_utils.py:

  • After standard recipe override_params are loaded, find subscription recipes
  • Resolve {customer_id} placeholder in the S3 URI with caller's account ID
  • Handle S3 access point ARN URI format for GetObject
  • Merge subscription keys with default: None (non-destructive, won't serialize unless set)
  • Silent fallback on any exception

• sagemaker-train/tests/unit/train/common_utils/test_finetune_utils.py:

  • Test: subscription recipe HPs available when user is subscribed (default is None)
  • Test: subscription recipe HPs NOT available when no subscription recipe exists
  • Test: graceful fallback when user is not subscribed (AccessDenied)

Testing

Validated end-to-end in IAD gamma with both cases:

Datamix job (user sets datamix HPs):

• trainer.hyperparameters.customer_data_percent = 70 works natively
• Only datamix HPs appear in the API request
• API selects datamix recipe (nova_lite_2_0_p5_gpu_sft_text_with_datamix)

Non-datamix job (user only sets standard HPs):

• No datamix HPs in the API request
• API selects standard recipe (nova_lite_2_0_p5_gpu_sft)
• No regression

All 58 unit tests passing.

Customer Experience

Subscribed user wanting datamix:

trainer = SFTTrainer(model="nova-textgeneration-lite-v2", ...)
trainer.hyperparameters.customer_data_percent = 70
trainer.hyperparameters.nova_code_percent = 30
# ... set all nova categories summing to 100
trainer.train()  # → datamix recipe selected

Subscribed user NOT wanting datamix:

trainer = SFTTrainer(model="nova-textgeneration-lite-v2", ...)
trainer.hyperparameters.max_steps = 4
trainer.train()  # → standard recipe selected, no datamix HPs sent

Non-subscribed user:

trainer = SFTTrainer(model="nova-textgeneration-lite-v2", ...)
# Datamix HPs not in schema (access point download failed silently)
# Standard HPs work as before — no regression