sagemaker.train.rft.adapters.strands

Contents

sagemaker.train.rft.adapters.strands#

Strands framework adapter for automatic header and inference param injection.

Provides wrap_model() which wraps a Strands model to automatically inject RFT headers and inference parameters into requests using the rollout context.

The wrapper intercepts stream() and injects headers via client_args["default_headers"] because Strands OpenAIModel creates a new OpenAI client per request from client_args.

Functions

wrap_model(model)

Wrap a Strands model to auto-inject headers and inference params from context.
sagemaker.train.rft.adapters.strands.wrap_model(model: Any) → Any[source]#

Wrap a Strands model to auto-inject headers and inference params from context.

Creates a transparent proxy that: 1. Injects the X-RFT-Metadata header (containing job_id, experiment_id,

rollout_id) via client_args[“default_headers”] on every stream() call

  1. Injects inference parameters (temperature, max_tokens, top_p)

Parameters:

model – A Strands model instance (e.g., OpenAIModel).

Returns:

A wrapped model that transparently injects RFT headers.

Example:

from strands.models.openai import OpenAIModel
from strands import Agent
from sagemaker.train.rft.adapters.strands import wrap_model

model = OpenAIModel(
    client_args={"api_key": "...", "base_url": "..."},
    model_id="my-model",
)
wrapped = wrap_model(model)
agent = Agent(model=wrapped, tools=[...])
result = agent("Solve this task")
Contents