Adding torch accelerator to ddp-tutorial-series example

distributed/ddp-tutorial-series/README.md

@@ -15,7 +15,27 @@ Each code file extends upon the previous one. The series starts with a non-distr
     * [slurm/setup_pcluster_slurm.md](slurm/setup_pcluster_slurm.md): instructions to set up an AWS cluster
     * [slurm/config.yaml.template](slurm/config.yaml.template): configuration to set up an AWS cluster
     * [slurm/sbatch_run.sh](slurm/sbatch_run.sh): slurm script to launch the training job
-
-
-
-
+## Installation
+```
+pip install -r requirements.txt
+```
+## Running Examples
+For running the examples to run for 20 Epochs and save checkpoints every 5 Epochs, you can use the following command:
+### Single GPU
+```
+python single_gpu.py 20 5
+```
+### Multi-GPU
+```
+python multigpu.py 20 5
+```
+### Multi-GPU Torchrun
+```
+torchrun --nnodes=1 --nproc_per_node=4 multigpu_torchrun.py 20 5
+```
+### Multi-Node
+```
+torchrun --nnodes=2 --nproc_per_node=4  multinode.py 20 5
+```
+
+For more details, check the [run_examples.sh](distributed/ddp-tutorial-series/run_examples.sh) script.

distributed/ddp-tutorial-series/multigpu.py

@@ -18,8 +18,14 @@ def ddp_setup(rank, world_size):
     """
     os.environ["MASTER_ADDR"] = "localhost"
     os.environ["MASTER_PORT"] = "12355"
-    torch.cuda.set_device(rank)
-    init_process_group(backend="nccl", rank=rank, world_size=world_size)
+
+    if torch.accelerator.is_available():       
+        device = torch.device(f"{torch.accelerator.current_accelerator()}:{rank}")
+        torch.accelerator.set_device_index(rank)
+        print(f"Running on rank {rank} on device {device}")
+
+    backend = torch.distributed.get_default_backend_for_device(device)
+    init_process_group(backend=backend, rank=rank, world_size=world_size)
 
 class Trainer:
     def __init__(
@@ -95,10 +101,10 @@ def main(rank: int, world_size: int, save_every: int, total_epochs: int, batch_s
 if __name__ == "__main__":
     import argparse
     parser = argparse.ArgumentParser(description='simple distributed training job')
-    parser.add_argument('total_epochs', type=int, help='Total epochs to train the model')
-    parser.add_argument('save_every', type=int, help='How often to save a snapshot')
+    parser.add_argument('total_epochs', default=50, type=int, help='Total epochs to train the model')
+    parser.add_argument('save_every', default=5, type=int, help='How often to save a snapshot')
     parser.add_argument('--batch_size', default=32, type=int, help='Input batch size on each device (default: 32)')
     args = parser.parse_args()
 
-    world_size = torch.cuda.device_count()
+    world_size = torch.accelerator.device_count()
     mp.spawn(main, args=(world_size, args.save_every, args.total_epochs, args.batch_size), nprocs=world_size)

distributed/ddp-tutorial-series/multigpu_torchrun.py

@@ -11,8 +11,18 @@
 
 
 def ddp_setup():
-    torch.cuda.set_device(int(os.environ["LOCAL_RANK"]))
-    init_process_group(backend="nccl")
+    rank = int(os.environ["LOCAL_RANK"])
+    if torch.accelerator.is_available():
+        device = torch.device(f"{torch.accelerator.current_accelerator()}:{rank}")
+        torch.accelerator.set_device_index(rank)
+        print(f"Running on rank {rank} on device {device}")
+    else:
+        print(f"Multi-GPU environment not detected")
+
+    backend = torch.distributed.get_default_backend_for_device(rank)
+    torch.distributed.init_process_group(backend=backend, rank=rank, device_id=rank)
+
+
 
 class Trainer:
     def __init__(
@@ -37,7 +47,7 @@ def __init__(
         self.model = DDP(self.model, device_ids=[self.gpu_id])
 
     def _load_snapshot(self, snapshot_path):
-        loc = f"cuda:{self.gpu_id}"
+        loc = str(torch.accelerator.current_accelerator())
         snapshot = torch.load(snapshot_path, map_location=loc)
         self.model.load_state_dict(snapshot["MODEL_STATE"])
         self.epochs_run = snapshot["EPOCHS_RUN"]
@@ -103,8 +113,8 @@ def main(save_every: int, total_epochs: int, batch_size: int, snapshot_path: str
 if __name__ == "__main__":
     import argparse
     parser = argparse.ArgumentParser(description='simple distributed training job')
-    parser.add_argument('total_epochs', type=int, help='Total epochs to train the model')
-    parser.add_argument('save_every', type=int, help='How often to save a snapshot')
+    parser.add_argument('total_epochs', default=50, type=int, help='Total epochs to train the model')
+    parser.add_argument('save_every', default=5, type=int, help='How often to save a snapshot')
     parser.add_argument('--batch_size', default=32, type=int, help='Input batch size on each device (default: 32)')
     args = parser.parse_args()
 

distributed/ddp-tutorial-series/multinode.py

@@ -11,8 +11,17 @@
 
 
 def ddp_setup():
-    torch.cuda.set_device(int(os.environ["LOCAL_RANK"]))
-    init_process_group(backend="nccl")
+    rank = int(os.environ["LOCAL_RANK"])
+    if torch.accelerator.is_available():
+        device = torch.device(f"{torch.accelerator.current_accelerator()}:{rank}")
+        torch.accelerator.set_device_index(rank)
+        print(f"Running on rank {rank} on device {device}")
+    else:
+        print(f"Multi-GPU environment not detected")
+
+    backend = torch.distributed.get_default_backend_for_device(rank)
+    torch.distributed.init_process_group(backend=backend, rank=rank, device_id=rank)
+
 
 class Trainer:
     def __init__(
@@ -38,7 +47,7 @@ def __init__(
         self.model = DDP(self.model, device_ids=[self.local_rank])
 
     def _load_snapshot(self, snapshot_path):
-        loc = f"cuda:{self.local_rank}"
+        loc = str(torch.accelerator.current_accelerator())
         snapshot = torch.load(snapshot_path, map_location=loc)
         self.model.load_state_dict(snapshot["MODEL_STATE"])
         self.epochs_run = snapshot["EPOCHS_RUN"]
@@ -104,8 +113,8 @@ def main(save_every: int, total_epochs: int, batch_size: int, snapshot_path: str
 if __name__ == "__main__":
     import argparse
     parser = argparse.ArgumentParser(description='simple distributed training job')
-    parser.add_argument('total_epochs', type=int, help='Total epochs to train the model')
-    parser.add_argument('save_every', type=int, help='How often to save a snapshot')
+    parser.add_argument('total_epochs', default=50, type=int, help='Total epochs to train the model')
+    parser.add_argument('save_every', default=5, type=int, help='How often to save a snapshot')
     parser.add_argument('--batch_size', default=32, type=int, help='Input batch size on each device (default: 32)')
     args = parser.parse_args()
 

distributed/ddp-tutorial-series/requirements.txt

@@ -1 +1 @@
-torch>=1.11.0
+torch>=2.7

distributed/ddp-tutorial-series/run_example.sh

@@ -0,0 +1,11 @@
+# /bin/bash
+# bash run_example.sh {file_to_run.py} {num_gpus}
+# where file_to_run = example to run. Default = 'example.py'
+# num_gpus = num local gpus to use (must be at least 2). Default = 2
+
+# samples to run include:
+# multigpu_torchrun.py
+# multinode.py
+
+echo "Launching ${1:-example.py} with ${2:-2} gpus"
+torchrun --nnodes=1 --nproc_per_node=${2:-2} --rdzv_id=101 --rdzv_endpoint="localhost:5972" ${1:-example.py} 10 1

distributed/ddp-tutorial-series/single_gpu.py

@@ -73,10 +73,10 @@ def main(device, total_epochs, save_every, batch_size):
 if __name__ == "__main__":
     import argparse
     parser = argparse.ArgumentParser(description='simple distributed training job')
-    parser.add_argument('total_epochs', type=int, help='Total epochs to train the model')
-    parser.add_argument('save_every', type=int, help='How often to save a snapshot')
+    parser.add_argument('total_epochs', default=50, type=int, help='Total epochs to train the model')
+    parser.add_argument('save_every', default=5, type=int, help='How often to save a snapshot')
     parser.add_argument('--batch_size', default=32, type=int, help='Input batch size on each device (default: 32)')
     args = parser.parse_args()
 
-    device = 0  # shorthand for cuda:0
+    device = 0  
     main(device, args.total_epochs, args.save_every, args.batch_size)

run_distributed_examples.sh

@@ -50,8 +50,16 @@ function distributed_tensor_parallelism() {
     uv run bash run_example.sh fsdp_tp_example.py || error "2D parallel example failed"
 }
 
+
+function distributed_ddp-tutorial-series() {
+    uv python multigpu.py 10 1 || error "ddp tutorial series multigpu example failed"
+    uv run bash run_example.sh multigpu_torchrun.py || error "ddp tutorial series multigpu torchrun example failed"
+    uv run bash run_example.sh multinode.py || error "ddp tutorial series multinode example failed"
+    uv python single_gpu.py 10 1 || error "ddp tutorial series single gpu example failed"
+
 function distributed_FSDP2() {
     uv run bash run_example.sh example.py || error "FSDP2 example failed"
+
 }
 
 function distributed_ddp() {
@@ -72,6 +80,7 @@ function distributed_rpc_rnn() {
 
 function run_all() {
   run distributed/tensor_parallelism
+  run distributed/ddp-tutorial-series
   run distributed/ddp
   run distributed/minGPT-ddp
   run distributed/rpc/ddp_rpc