[Release] Use NCCL backend for release tests (#20677)

* use nccl for release tests

* link issue

python/ray/train/examples/train_linear_example.py

@@ -5,7 +5,6 @@ import torch
 import torch.nn as nn
 import ray.train as train
 from ray.train import Trainer
-from ray.train.torch import TorchConfig
 from ray.train.callbacks import JsonLoggerCallback, TBXLoggerCallback
 
 
@@ -89,9 +88,7 @@ def train_func(config):
 
 def train_linear(num_workers=2, use_gpu=False, epochs=3):
     trainer = Trainer(
-        backend=TorchConfig(backend="gloo"),
-        num_workers=num_workers,
-        use_gpu=use_gpu)
+        backend="torch", num_workers=num_workers, use_gpu=use_gpu)
     config = {"lr": 1e-2, "hidden_size": 1, "batch_size": 4, "epochs": epochs}
     trainer.start()
     results = trainer.run(

release/ml_user_tests/ray-lightning/app_config.yaml

@@ -1,6 +1,4 @@
 base_image: "anyscale/ray-ml:nightly-py37-gpu"
-env_vars:
-  PL_TORCH_DISTRIBUTED_BACKEND: gloo
 
 debian_packages:
   - curl

release/ml_user_tests/ray-lightning/app_config_master.yaml

@@ -1,6 +1,4 @@
 base_image: "anyscale/ray-ml:nightly-py37-gpu"
-env_vars:
-  PL_TORCH_DISTRIBUTED_BACKEND: gloo
 
 debian_packages:
   - curl

release/ml_user_tests/ray-lightning/ray_lightning_user_test.py

@@ -9,11 +9,19 @@ if __name__ == "__main__":
     start = time.time()
 
     addr = os.environ.get("RAY_ADDRESS")
-    job_name = os.environ.get("RAY_JOB_NAME", "horovod_user_test")
+    job_name = os.environ.get("RAY_JOB_NAME", "ray_lightning_user_test")
+
+    # Manually set NCCL_SOCKET_IFNAME to "ens3" so NCCL training works on
+    # anyscale_default_cloud.
+    # See https://github.com/pytorch/pytorch/issues/68893 for more details.
+    # Passing in runtime_env to ray.init() will also set it for all the
+    # workers.
+    runtime_env = {"env_vars": {"NCCL_SOCKET_IFNAME": "ens3"}}
+
     if addr is not None and addr.startswith("anyscale://"):
-        ray.init(address=addr, job_name=job_name)
+        ray.init(address=addr, job_name=job_name, runtime_env=runtime_env)
     else:
-        ray.init(address="auto")
+        ray.init(address="auto", runtime_env=runtime_env)
 
     main(num_workers=6, use_gpu=True, max_steps=50)
 

release/ml_user_tests/train/train_torch_linear_test.py

@@ -12,10 +12,17 @@ if __name__ == "__main__":
     addr = os.environ.get("RAY_ADDRESS")
     job_name = os.environ.get("RAY_JOB_NAME", "train_torch_linear_test")
 
+    # Manually set NCCL_SOCKET_IFNAME to "ens3" so NCCL training works on
+    # anyscale_default_cloud.
+    # See https://github.com/pytorch/pytorch/issues/68893 for more details.
+    # Passing in runtime_env to ray.init() will also set it for all the
+    # workers.
+    runtime_env = {"env_vars": {"NCCL_SOCKET_IFNAME": "ens3"}}
+
     if addr is not None and addr.startswith("anyscale://"):
-        ray.init(address=addr, job_name=job_name)
+        ray.init(address=addr, job_name=job_name, runtime_env=runtime_env)
     else:
-        ray.init(address="auto")
+        ray.init(address="auto", runtime_env=runtime_env)
 
     results = train_linear(num_workers=6, use_gpu=True, epochs=20)