diff --git a/.github/workflows/main_distributed.yaml b/.github/workflows/main_distributed.yaml
index b70da3617a..78afef687b 100644
--- a/.github/workflows/main_distributed.yaml
+++ b/.github/workflows/main_distributed.yaml
@@ -17,10 +17,10 @@ jobs:
 
     steps:
     - uses: actions/checkout@v2
-    - name: Set up Python 3.8
+    - name: Set up Python 3.10
       uses: actions/setup-python@v2
       with:
-        python-version: 3.8
+        python-version: 3.10
     - name: Install PyTorch
       uses: astral-sh/setup-uv@v6
     - name: Run Tests
diff --git a/distributed/tensor_parallelism/README.md b/distributed/tensor_parallelism/README.md
index b49d1672e8..ec61071e65 100644
--- a/distributed/tensor_parallelism/README.md
+++ b/distributed/tensor_parallelism/README.md
@@ -12,5 +12,5 @@ https://pytorch.org/docs/stable/distributed.tensor.parallel.html
 
 ```
 pip install -r requirements.txt
-python example.py
+torchrun --nnodes 1 --nproc-per-node 4 tensor_parallel_example.py
 ```
diff --git a/distributed/tensor_parallelism/fsdp_tp_example.py b/distributed/tensor_parallelism/fsdp_tp_example.py
index dbab48c1b8..87935f10f0 100644
--- a/distributed/tensor_parallelism/fsdp_tp_example.py
+++ b/distributed/tensor_parallelism/fsdp_tp_example.py
@@ -77,10 +77,11 @@
 # create a sharding plan based on the given world_size.
 dp_size = _world_size // tp_size
 
+device_type = torch.accelerator.current_accelerator().type
 # Create a device mesh with 2 dimensions.
 # First dim is the data parallel dimension
 # Second dim is the tensor parallel dimension.
-device_mesh = init_device_mesh("cuda", (dp_size, tp_size), mesh_dim_names=("dp", "tp"))
+device_mesh = init_device_mesh(device_type, (dp_size, tp_size), mesh_dim_names=("dp", "tp"))
 
 rank_log(_rank, logger, f"Device Mesh created: {device_mesh=}")
 tp_mesh = device_mesh["tp"]
@@ -92,10 +93,10 @@
 # to mimic the behavior of the dataloader.
 dp_rank = dp_mesh.get_local_rank()
 
-# create model and move it to GPU - init"cuda"_mesh has already mapped GPU ids.
+# create model and move it to GPU - initdevice_type_mesh has already mapped GPU ids.
 simple_llama2_config = ModelArgs(dim=256, n_layers=2, n_heads=16, vocab_size=32000)
 
-model = Transformer.from_model_args(simple_llama2_config).to("cuda")
+model = Transformer.from_model_args(simple_llama2_config).to(device_type)
 
 # init model weights
 model.init_weights()
@@ -170,7 +171,7 @@
 for i in range(num_iterations):
     # seeding with dp_rank to ensure identical inputs for TP groups
     torch.manual_seed(i + dp_rank)
-    inp = torch.randint(32000, (8, 256), device="cuda")
+    inp = torch.randint(32000, (8, 256), device=device_type)
 
     output = sharded_model(inp)
     output.sum().backward()
diff --git a/distributed/tensor_parallelism/log_utils.py b/distributed/tensor_parallelism/log_utils.py
index f16d46526d..d103df892b 100644
--- a/distributed/tensor_parallelism/log_utils.py
+++ b/distributed/tensor_parallelism/log_utils.py
@@ -17,6 +17,6 @@ def rank_log(_rank, logger, msg):
 
 def verify_min_gpu_count(min_gpus: int = 2) -> bool:
     """ verification that we have at least 2 gpus to run dist examples """
-    has_cuda = torch.cuda.is_available()
-    gpu_count = torch.cuda.device_count()
-    return has_cuda and gpu_count >= min_gpus
+    has_gpu = torch.accelerator.is_available()
+    gpu_count = torch.accelerator.device_count()
+    return has_gpu and gpu_count >= min_gpus
diff --git a/distributed/tensor_parallelism/requirements.txt b/distributed/tensor_parallelism/requirements.txt
index 80fad36bf2..4f47924993 100644
--- a/distributed/tensor_parallelism/requirements.txt
+++ b/distributed/tensor_parallelism/requirements.txt
@@ -1,6 +1,3 @@
 # Python dependencies required for running the example
 
---pre
---extra-index-url https://download.pytorch.org/whl/nightly/cu118
---extra-index-url https://download.pytorch.org/whl/nightly/cu121
-torch >= 2.3.0.dev0; sys_platform == "linux"
+torch >= 2.7.1; sys_platform == "linux"
diff --git a/distributed/tensor_parallelism/sequence_parallel_example.py b/distributed/tensor_parallelism/sequence_parallel_example.py
index 3324d28d4a..b145fbc95e 100644
--- a/distributed/tensor_parallelism/sequence_parallel_example.py
+++ b/distributed/tensor_parallelism/sequence_parallel_example.py
@@ -1,3 +1,5 @@
+# The following is an example command to run this code
+# torchrun --nnodes 1 --nproc-per-node 4 sequence_parallel_example.py
 import os
 import sys
 import torch
@@ -63,9 +65,10 @@ def forward(self, x):
 """
 logger = get_logger()
 
+device_type = torch.accelerator.current_accelerator().type
 # create a device mesh based on the given world_size.
 device_mesh = init_device_mesh(
-    device_type="cuda", mesh_shape=(int(os.environ["WORLD_SIZE"]),)
+    device_type=device_type, mesh_shape=(int(os.environ["WORLD_SIZE"]),)
 )
 
 _rank = device_mesh.get_rank()
@@ -75,7 +78,7 @@ def forward(self, x):
 rank_log(_rank, logger, f"Device Mesh created: {device_mesh=}")
 
 # create model and move it to GPU.  Init_device_mesh has already assigned gpu ids...
-model = ToyModel().to("cuda")
+model = ToyModel().to(device_type)
 
 # Custom parallelization plan for the model
 sp_model = parallelize_module(
@@ -100,7 +103,7 @@ def forward(self, x):
 
 for i in range(num_iters):
     # For SP, input can be different across all ranks.
-    inp = torch.rand(20, 10, device="cuda")
+    inp = torch.rand(20, 10, device=device_type)
     output = sp_model(inp)
     output.sum().backward()
     optimizer.step()
diff --git a/distributed/tensor_parallelism/tensor_parallel_example.py b/distributed/tensor_parallelism/tensor_parallel_example.py
index 0b9c884507..b96f982f0c 100755
--- a/distributed/tensor_parallelism/tensor_parallel_example.py
+++ b/distributed/tensor_parallelism/tensor_parallel_example.py
@@ -1,3 +1,5 @@
+# The following is an example command to run this code
+# torchrun --nnodes 1 --nproc-per-node 4 tensor_parallel_example.py
 import os
 import sys
 import torch
@@ -76,8 +78,8 @@ def forward(self, x):
 
 # create a device mesh based on the given world_size.
 _world_size = int(os.environ["WORLD_SIZE"])
-
-device_mesh = init_device_mesh(device_type="cuda", mesh_shape=(_world_size,))
+device_type = torch.accelerator.current_accelerator().type
+device_mesh = init_device_mesh(device_type=device_type, mesh_shape=(_world_size,))
 _rank = device_mesh.get_rank()
 
 
@@ -88,8 +90,8 @@ def forward(self, x):
 
 rank_log(_rank, logger, f"Device Mesh created: {device_mesh=}")
 
-# create model and move it to GPU - init"cuda"_mesh has already mapped GPU ids.
-tp_model = ToyModel().to("cuda")
+# create model and move it to GPU - initdevice_type_mesh has already mapped GPU ids.
+tp_model = ToyModel().to(device_type)
 
 
 # Custom parallelization plan for the model
@@ -116,7 +118,7 @@ def forward(self, x):
     # For TP, input needs to be same across all TP ranks.
     # Setting the random seed is to mimic the behavior of dataloader.
     torch.manual_seed(i)
-    inp = torch.rand(20, 10, device="cuda")
+    inp = torch.rand(20, 10, device=device_type)
     output = tp_model(inp)
     output.sum().backward()
     optimizer.step()
diff --git a/runtime.txt b/runtime.txt
index cc1923a40b..bd28b9c5c2 100644
--- a/runtime.txt
+++ b/runtime.txt
@@ -1 +1 @@
-3.8
+3.9