feat: add post_training RuntimeConfig

certain APIs require a bunch of runtime arguments per-provider. The best way currently to pass these arguments in is via the
provider config. This is tricky because it requires a provider to be pre-configured with certain arguments that a client side user should be able to pass in at runtime

Especially with the advent of out-of-tree providers, it would be great for a generic RuntimeConfig class to allow for providers to add and validate their own runtime arguments for things like supervised_fine_tune

For example: https://github.com/opendatahub-io/llama-stack-provider-kft has things like `input-pvc`, `model-path`, etc in the Provider Config.
This is not sustainable nor is adding each and every field needed to the post_training API spec. RuntimeConfig has a sub-class called Config which allows for extra fields to arbitrarily be specified. It is the providers job to create its own class based on this one and add valid options, parse them, etc

Signed-off-by: Charlie Doern <[email protected]>

Author: cdoern

llama_stack/apis/post_training/post_training.py

@@ -169,6 +169,16 @@ class PostTrainingJobArtifactsResponse(BaseModel):
     # TODO(ashwin): metrics, evals
 
 
+@json_schema_type
+class RuntimeConfig(BaseModel):
+    """
+    Provider-specific runtime configuration. Providers should document and parse their own expected fields.
+    This model allows arbitrary extra fields for maximum flexibility.
+    """
+    class Config:
+        extra = "allow"
+
+
 class PostTraining(Protocol):
     @webmethod(route="/post-training/supervised-fine-tune", method="POST")
     async def supervised_fine_tune(
@@ -183,6 +193,7 @@ async def supervised_fine_tune(
         ),
         checkpoint_dir: Optional[str] = None,
         algorithm_config: Optional[AlgorithmConfig] = None,
+        runtime_config: Optional[RuntimeConfig] = None,
     ) -> PostTrainingJob: ...
 
     @webmethod(route="/post-training/preference-optimize", method="POST")
@@ -194,6 +205,7 @@ async def preference_optimize(
         training_config: TrainingConfig,
         hyperparam_search_config: Dict[str, Any],
         logger_config: Dict[str, Any],
+        runtime_config: Optional[RuntimeConfig] = None,
     ) -> PostTrainingJob: ...
 
     @webmethod(route="/post-training/jobs", method="GET")

llama_stack/providers/inline/post_training/torchtune/post_training.py

@@ -19,6 +19,7 @@
     PostTrainingJobArtifactsResponse,
     PostTrainingJobStatusResponse,
     TrainingConfig,
+    RuntimeConfig,
 )
 from llama_stack.providers.inline.post_training.torchtune.config import (
     TorchtunePostTrainingConfig,
@@ -80,6 +81,7 @@ async def supervised_fine_tune(
         model: str,
         checkpoint_dir: Optional[str],
         algorithm_config: Optional[AlgorithmConfig],
+        runtime_config: Optional[RuntimeConfig] = None,
     ) -> PostTrainingJob:
         if isinstance(algorithm_config, LoraFinetuningConfig):