feat(sagemaker): add train_dapo with dynamic sampling and Clip-Higher

src/agentcore_rl_toolkit/backends/experimental/sagemaker/SETUP.md

@@ -48,3 +48,15 @@ cp config.yaml.example config.yaml
 cd src/agentcore_rl_toolkit/backends/experimental/sagemaker
 python train_grpo.py --config config.yaml
 ```
+
+To train with DAPO instead, run `train_dapo.py` with the same config:
+
+```bash
+python train_dapo.py --config config.yaml
+```
+
+`train_dapo.py` is the GRPO loop plus dynamic sampling (each step rolls out batches of
+prompts and keeps only groups whose rewards differ, until `batch_size` groups are kept) and
+Clip-Higher (`clip_low_threshold` / `clip_high_threshold`). Its extra settings are in the
+DAPO section of `config.yaml.example`. Each step logs a `sampling:` line with the number of
+generation batches and the filtered fraction.

src/agentcore_rl_toolkit/backends/experimental/sagemaker/config.py

@@ -1,5 +1,5 @@
 from dataclasses import dataclass, field
-from typing import Literal
+from typing import Literal, TypeVar
 
 import yaml
 
@@ -58,9 +58,18 @@ class Config:
     rollout_log_path: str = ""  # JSONL file for per-rollout diagnostics; auto-named if empty
 
 
-def load_config(path: str) -> Config:
+ConfigT = TypeVar("ConfigT", bound=Config)
+
+
+def load_config(path: str, config_cls: type[ConfigT] = Config) -> ConfigT:
+    """Load a YAML file into ``config_cls``: ``Config``, or a recipe's subclass that adds fields.
+
+    Keys that ``config_cls`` does not define are ignored, so one YAML file can serve several
+    recipes (e.g. ``train_grpo`` skips the fields only ``train_dapo`` reads). Returns an
+    instance of ``config_cls``.
+    """
     with open(path) as f:
         raw: dict = yaml.safe_load(f) or {}
 
-    fields = Config.__dataclass_fields__
-    return Config(**{k: v for k, v in raw.items() if k in fields})
+    fields = config_cls.__dataclass_fields__
+    return config_cls(**{k: v for k, v in raw.items() if k in fields})

src/agentcore_rl_toolkit/backends/experimental/sagemaker/config.yaml.example

@@ -1,4 +1,5 @@
-# GRPO training — rollout gateway + SageMaker Training Sessions
+# GRPO / DAPO training — rollout gateway + SageMaker Training Sessions
+# Both train_grpo.py and train_dapo.py read this file; each ignores keys it does not use.
 
 # --- AWS resources (required) ---
 role_arn: "arn:aws:iam::<account-id>:role/<your-sagemaker-role>"
@@ -51,3 +52,8 @@ acr_tps_limit: 5
 # --- Logging ---
 wandb_project: ""
 wandb_run_name: ""  # defaults to the SageMaker training session name
+
+# --- DAPO (train_dapo.py only; train_grpo.py ignores these) ---
+clip_low_threshold: 0.8    # PPO ratio lower bound, 1 - eps_low
+clip_high_threshold: 1.28  # PPO ratio upper bound, 1 + eps_high (Clip-Higher)
+max_num_gen_batches: 10    # dynamic sampling: max generation batches per step (0 = no cap)