chore: import upstream snapshot with attribution

release/train_tests/torchft/compute_aws.yaml

@@ -0,0 +1,8 @@
+cloud: {{env["ANYSCALE_CLOUD_NAME"]}}
+
+# Single GPU node. The torchft linear example trains on CPU (gloo) but the
+# image ships GPU torch (torch==2.9.0+cu128), so a CUDA node is representative.
+head_node:
+    instance_type: g4dn.4xlarge
+
+worker_nodes: []

release/train_tests/torchft/torchft_hello_world.py

@@ -0,0 +1,39 @@
+"""Hello-world release test for the torchft Ray ML image variant.
+
+This is a reference test showing how to run a release test on the custom
+torchft BYOD image. The image is the core Ray CUDA image (py3.13) with the
+torchft ML dependency lock installed on top:
+
+  - lock: release/ray_release/byod/ml_torchft_py3.13.lock
+          (generated by ci/raydepsets/configs/release_ml_torchft_tests.depsets.yaml)
+
+The lock is referenced from the release test via `byod.python_depset`; the BYOD
+build installs it automatically (`uv pip install --system --no-deps -r
+python_depset.lock`), so no post_build_script is required.
+
+See the matching `torchft_hello_world` entry in release/release_tests.yaml for
+the cluster wiring (byod.type / byod.python_depset).
+
+The test simply proves the image works end to end: torch 2.9.0 + torchft import
+cleanly and a tiny Ray Train v2 + torchft training loop runs to completion.
+"""
+
+import ray
+import torch
+import torchft  # noqa: F401  -- provided by torchft-nightly in the image
+
+from ray.train.v2.examples.pytorch.torchft_linear_example import train_torchft
+
+
+def main() -> None:
+    print(f"ray=={ray.__version__} torch=={torch.__version__} torchft OK")
+
+    # torchft_linear_example trains a tiny linear model on CPU (gloo backend)
+    # under Ray Train v2 with a torchft TorchftConfig. A short run is enough to
+    # validate that the image's torch + torchft + Ray Train stack works.
+    metrics = train_torchft(num_workers=2, num_steps=20)
+    print(f"torchft hello world succeeded: {metrics}")
+
+
+if __name__ == "__main__":
+    main()