Adding torch accelerator to ddp-tutorial-series example

Signed-off-by: dggaytan <diana.gaytan.munoz@intel.com>

Author: dggaytan

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

@@ -19,14 +19,10 @@ def ddp_setup(rank, world_size):
     os.environ["MASTER_ADDR"] = "localhost"
     os.environ["MASTER_PORT"] = "12355"
 
-    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:
-        device = torch.device("cpu")
-        print(f"Running on device {device}")
 
     backend = torch.distributed.get_default_backend_for_device(device)
     init_process_group(backend=backend, rank=rank, world_size=world_size)

distributed/ddp-tutorial-series/multigpu_torchrun.py

@@ -17,12 +17,11 @@ def ddp_setup():
         torch.accelerator.set_device_index(rank)
         print(f"Running on rank {rank} on device {device}")
     else:
-        device = torch.device("cpu")
-        print(f"Running on device {device}")
-                
-    backend = torch.distributed.get_default_backend_for_device(device)
-    torch.distributed.init_process_group(backend=backend, device_id=device)
-    return device
+        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:
@@ -33,7 +32,6 @@ def __init__(
         optimizer: torch.optim.Optimizer,
         save_every: int,
         snapshot_path: str,
-        device: torch.device, 
     ) -> None:
         self.gpu_id = int(os.environ["LOCAL_RANK"])
         self.model = model.to(self.gpu_id)
@@ -42,15 +40,14 @@ def __init__(
         self.save_every = save_every
         self.epochs_run = 0
         self.snapshot_path = snapshot_path
-        self.device = device
         if os.path.exists(snapshot_path):
             print("Loading snapshot")
             self._load_snapshot(snapshot_path)
 
         self.model = DDP(self.model, device_ids=[self.gpu_id])
 
     def _load_snapshot(self, snapshot_path):
-        loc = str(self.device)
+        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"]
@@ -105,10 +102,10 @@ def prepare_dataloader(dataset: Dataset, batch_size: int):
 
 
 def main(save_every: int, total_epochs: int, batch_size: int, snapshot_path: str = "snapshot.pt"):
-    device = ddp_setup()
+    ddp_setup()
     dataset, model, optimizer = load_train_objs()
     train_data = prepare_dataloader(dataset, batch_size)
-    trainer = Trainer(model, train_data, optimizer, save_every, snapshot_path, device)
+    trainer = Trainer(model, train_data, optimizer, save_every, snapshot_path)
     trainer.train(total_epochs)
     destroy_process_group()
 

distributed/ddp-tutorial-series/multinode.py

@@ -17,12 +17,11 @@ def ddp_setup():
         torch.accelerator.set_device_index(rank)
         print(f"Running on rank {rank} on device {device}")
     else:
-        device = torch.device("cpu")
-        print(f"Running on device {device}")
-                
-    backend = torch.distributed.get_default_backend_for_device(device)
-    torch.distributed.init_process_group(backend=backend, device_id=device)
-    return device
+        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__(
@@ -32,7 +31,6 @@ def __init__(
         optimizer: torch.optim.Optimizer,
         save_every: int,
         snapshot_path: str,
-        device: torch.device,
     ) -> None:
         self.local_rank = int(os.environ["LOCAL_RANK"])
         self.global_rank = int(os.environ["RANK"])
@@ -42,15 +40,14 @@ def __init__(
         self.save_every = save_every
         self.epochs_run = 0
         self.snapshot_path = snapshot_path
-        self.device = device
         if os.path.exists(snapshot_path):
             print("Loading snapshot")
             self._load_snapshot(snapshot_path)
 
         self.model = DDP(self.model, device_ids=[self.local_rank])
 
     def _load_snapshot(self, snapshot_path):
-        loc = str(self.device)
+        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"]
@@ -105,10 +102,10 @@ def prepare_dataloader(dataset: Dataset, batch_size: int):
 
 
 def main(save_every: int, total_epochs: int, batch_size: int, snapshot_path: str = "snapshot.pt"):
-    device = ddp_setup()
+    ddp_setup()
     dataset, model, optimizer = load_train_objs()
     train_data = prepare_dataloader(dataset, batch_size)
-    trainer = Trainer(model, train_data, optimizer, save_every, snapshot_path, device)
+    trainer = Trainer(model, train_data, optimizer, save_every, snapshot_path)
     trainer.train(total_epochs)
     destroy_process_group()
 

distributed/ddp-tutorial-series/run_example.sh

@@ -4,7 +4,8 @@
 # num_gpus = num local gpus to use (must be at least 2). Default = 2
 
 # samples to run include:
-# example.py
+# 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" ${1:-example.py}
+torchrun --nnodes=1 --nproc_per_node=${2:-2} --rdzv_id=101 --rdzv_endpoint="localhost:5972" ${1:-example.py} 10 1

distributed/ddp-tutorial-series/single_gpu.py

@@ -78,5 +78,5 @@ def main(device, total_epochs, save_every, batch_size):
     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

@@ -51,9 +51,10 @@ function distributed_tensor_parallelism() {
 }
 
 function distributed_ddp-tutorial-series() {
-    uv run bash run_example.sh multigpu.py || error "ddp tutorial series multigpu example failed"
+    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_ddp() {