1. Introduction
have a mannequin. You’ve a single GPU. Coaching takes 72 hours. You requisition a second machine with 4 extra GPUs — and now you want your code to really use them. That is the precise second the place most practitioners hit a wall. Not as a result of distributed coaching is conceptually exhausting, however as a result of the engineering required to do it appropriately — course of teams, rank-aware logging, sampler seeding, checkpoint boundaries — is scattered throughout dozens of tutorials that every cowl one piece of the puzzle.
This text is the information I want I had after I first scaled coaching past a single node. We’ll construct an entire, production-grade multi-node coaching pipeline from scratch utilizing PyTorch’s DistributedDataParallel (DDP). Each file is modular, each worth is configurable, and each distributed idea is made specific. By the tip, you’ll have a codebase you may drop into any cluster and begin coaching instantly.
What we are going to cowl: the psychological mannequin behind DDP, a clear modular undertaking construction, distributed lifecycle administration, environment friendly information loading throughout ranks, a coaching loop with blended precision and gradient accumulation, rank-aware logging and checkpointing, multi-node launch scripts, and the efficiency pitfalls that journey up even skilled engineers.
The total codebase is offered on GitHub. Each code block on this article is pulled straight from that repository.
2. How DDP Works — The Psychological Mannequin
Earlier than writing any code, we’d like a transparent psychological mannequin. DistributedDataParallel (DDP) just isn’t magic — it’s a well-defined communication sample constructed on prime of collective operations.
The setup is simple. You launch N processes (one per GPU, doubtlessly throughout a number of machines). Every course of initialises a course of group — a communication channel backed by NCCL (NVIDIA Collective Communications Library) for GPU-to-GPU transfers. Each course of will get three id numbers: its world rank (distinctive throughout all machines), its native rank (distinctive inside its machine), and the entire world dimension.
Every course of holds an equivalent copy of the mannequin. Knowledge is partitioned throughout processes utilizing a DistributedSampler — each rank sees a distinct slice of the dataset, however the mannequin weights begin (and keep) equivalent.
The important mechanism is what occurs throughout backward(). DDP registers hooks on each parameter. When a gradient is computed for a parameter, DDP buckets it with close by gradients and fires an all-reduce operation throughout the method group. This all-reduce computes the imply gradient throughout all ranks. As a result of each rank now has the identical averaged gradient, the next optimizer step produces equivalent weight updates, preserving all replicas in sync — with none specific synchronisation code from us.
For this reason DDP is strictly superior to the older DataParallel: there isn’t any single “grasp” GPU bottleneck, no redundant ahead passes, and gradient communication overlaps with backward computation.
Key terminology
| Time period | That means |
| Rank | Globally distinctive course of ID (0 to world_size – 1) |
| Native Rank | GPU index inside a single machine (0 to nproc_per_node – 1) |
| World Measurement | Complete variety of processes throughout all nodes |
| Course of Group | Communication channel (NCCL) connecting all ranks |
3. Structure Overview
A manufacturing coaching pipeline ought to by no means be a single monolithic script. Ours is break up into six targeted modules, every with a single duty. The dependency graph under exhibits how they join — be aware that config.py sits on the backside, appearing as the one supply of reality for each hyperparameter.
Right here is the undertaking construction:
pytorch-multinode-ddp/
├── prepare.py # Entry level — coaching loop
├── config.py # Dataclass configuration + argparse
├── ddp_utils.py # Distributed setup, teardown, checkpointing
├── mannequin.py # MiniResNet (light-weight ResNet variant)
├── dataset.py # Artificial dataset + DistributedSampler loader
├── utils/
│ ├── logger.py # Rank-aware structured logging
│ └── metrics.py # Operating averages + distributed all-reduce
├── scripts/
│ └── launch.sh # Multi-node torchrun wrapper
└── necessities.txtThis separation means you may swap in an actual dataset by enhancing solely dataset.py, or change the mannequin by enhancing solely mannequin.py. The coaching loop by no means wants to vary.
4. Centralized Configuration
Arduous-coded hyperparameters are the enemy of reproducibility. We use a Python dataclass as our single supply of configuration. Each different module imports TrainingConfig and reads from it — nothing is hard-coded.
The dataclass doubles as our CLI parser: the from_args() classmethod introspects the sphere names and kinds, routinely constructing argparse flags with defaults. This implies you get –batch_size 128 and –no-use_amp without spending a dime, with out writing a single parser line by hand.
@dataclass
class TrainingConfig:
"""Immutable bag of each parameter the coaching pipeline wants."""
# Mannequin
num_classes: int = 10
in_channels: int = 3
image_size: int = 32
# Knowledge
batch_size: int = 64 # per-GPU
num_workers: int = 4
# Optimizer / Scheduler
epochs: int = 10
lr: float = 0.01
momentum: float = 0.9
weight_decay: float = 1e-4
# Distributed
backend: str = "nccl"
# Blended Precision
use_amp: bool = True
# Gradient Accumulation
grad_accum_steps: int = 1
# Checkpointing
checkpoint_dir: str = "./checkpoints"
save_every: int = 1
resume_from: Optionally available[str] = None
# Logging & Profiling
log_interval: int = 10
enable_profiling: bool = False
seed: int = 42
@classmethod
def from_args(cls) -> "TrainingConfig":
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
defaults = cls()
for identify, val in vars(defaults).gadgets():
arg_type = sort(val) if val just isn't None else str
if isinstance(val, bool):
parser.add_argument(f"--{identify}", default=val,
motion=argparse.BooleanOptionalAction)
else:
parser.add_argument(f"--{identify}", sort=arg_type, default=val)
return cls(**vars(parser.parse_args()))Why a dataclass as a substitute of YAML or JSON? Three causes: (1) sort hints are enforced by the IDE and mypy, (2) there’s zero dependency on third-party config libraries, and (3) each parameter has a visual default proper subsequent to its declaration. For manufacturing methods that want hierarchical configs, you may all the time layer Hydra or OmegaConf on prime of this sample.
5. Distributed Lifecycle Administration
The distributed lifecycle has three phases: initialise, run, and tear down. Getting any of those flawed can produce silent hangs, so we wrap all the things in specific error dealing with.
Course of Group Initialization
The setup_distributed() operate reads the three surroundings variables that torchrun units routinely (RANK, LOCAL_RANK, WORLD_SIZE), pins the proper GPU with torch.cuda.set_device(), and initialises the NCCL course of group. It returns a frozen dataclass — DistributedContext — that the remainder of the codebase passes round as a substitute of re-reading os.environ.
@dataclass(frozen=True)
class DistributedContext:
"""Immutable snapshot of the present course of's distributed id."""
rank: int
local_rank: int
world_size: int
system: torch.system
def setup_distributed(config: TrainingConfig) -> DistributedContext:
required_vars = ("RANK", "LOCAL_RANK", "WORLD_SIZE")
lacking = [v for v in required_vars if v not in os.environ]
if lacking:
elevate RuntimeError(
f"Lacking surroundings variables: {lacking}. "
"Launch with torchrun or set them manually.")
if not torch.cuda.is_available():
elevate RuntimeError("CUDA is required for NCCL distributed coaching.")
rank = int(os.environ["RANK"])
local_rank = int(os.environ["LOCAL_RANK"])
world_size = int(os.environ["WORLD_SIZE"])
torch.cuda.set_device(local_rank)
system = torch.system("cuda", local_rank)
dist.init_process_group(backend=config.backend)
return DistributedContext(
rank=rank, local_rank=local_rank,
world_size=world_size, system=system)Checkpointing with Rank Guards
The most typical distributed checkpointing bug is all ranks writing to the identical file concurrently. We guard saving behind is_main_process(), and loading behind dist.barrier() — this ensures rank 0 finishes writing earlier than different ranks try to learn.
def save_checkpoint(path, epoch, mannequin, optimizer, scaler=None, rank=0):
"""Persist coaching state to disk (rank-0 solely)."""
if not is_main_process(rank):
return
Path(path).mum or dad.mkdir(dad and mom=True, exist_ok=True)
state = {
"epoch": epoch,
"model_state_dict": mannequin.module.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
}
if scaler just isn't None:
state["scaler_state_dict"] = scaler.state_dict()
torch.save(state, path)
def load_checkpoint(path, mannequin, optimizer=None, scaler=None, system="cpu"):
"""Restore coaching state. All ranks load after barrier."""
dist.barrier() # look ahead to rank 0 to complete writing
ckpt = torch.load(path, map_location=system, weights_only=False)
mannequin.load_state_dict(ckpt["model_state_dict"])
if optimizer and "optimizer_state_dict" in ckpt:
optimizer.load_state_dict(ckpt["optimizer_state_dict"])
if scaler and "scaler_state_dict" in ckpt:
scaler.load_state_dict(ckpt["scaler_state_dict"])
return ckpt.get("epoch", 0)6. Mannequin Design for DDP
We use a light-weight ResNet variant known as MiniResNet — three residual phases with rising channels (64, 128, 256), two blocks per stage, world common pooling, and a fully-connected head. It’s complicated sufficient to be practical however mild sufficient to run on any {hardware}.
The important DDP requirement: the mannequin should be moved to the proper GPU earlier than wrapping. DDP doesn’t transfer fashions for you.
def create_model(config: TrainingConfig, system: torch.system) -> nn.Module:
"""Instantiate a MiniResNet and transfer it to system."""
mannequin = MiniResNet(
in_channels=config.in_channels,
num_classes=config.num_classes,
)
return mannequin.to(system)
def wrap_ddp(mannequin: nn.Module, local_rank: int) -> DDP:
"""Wrap mannequin with DistributedDataParallel."""
return DDP(mannequin, device_ids=[local_rank])Be aware the two-step sample: create_model() → wrap_ddp(). This separation is intentional. When loading a checkpoint, you want the unwrapped mannequin (mannequin.module) to load state dicts, then re-wrap. If you happen to fuse creation and wrapping, checkpoint loading turns into awkward.
7. Distributed Knowledge Loading
DistributedSampler is what ensures every GPU sees a novel slice of knowledge. It partitions indices throughout world_size ranks and returns a non-overlapping subset for every. With out it, each GPU would prepare on equivalent batches — burning compute for zero profit.
There are three particulars that journey folks up:
First, sampler.set_epoch(epoch) should be known as initially of each epoch. The sampler makes use of the epoch quantity as a random seed for shuffling. If you happen to overlook this, each epoch will iterate over information in the identical order, which degrades generalisation.
Second, pin_memory=True within the DataLoader pre-allocates page-locked host reminiscence, enabling asynchronous CPU-to-GPU transfers whenever you name tensor.to(system, non_blocking=True). This overlap is the place actual throughput beneficial properties come from.
Third, persistent_workers=True avoids respawning employee processes each epoch — a major overhead discount when num_workers > 0.
def create_distributed_dataloader(dataset, config, ctx):
sampler = DistributedSampler(
dataset,
num_replicas=ctx.world_size,
rank=ctx.rank,
shuffle=True,
)
loader = DataLoader(
dataset,
batch_size=config.batch_size,
sampler=sampler,
num_workers=config.num_workers,
pin_memory=True,
drop_last=True,
persistent_workers=config.num_workers > 0,
)
return loader, sampler8. The Coaching Loop — The place It All Comes Collectively
That is the center of the pipeline. The loop under integrates each element now we have constructed to this point: DDP-wrapped mannequin, distributed information loader, blended precision, gradient accumulation, rank-aware logging, studying fee scheduling, and checkpointing.
Blended Precision (AMP)
Automated Blended Precision (AMP) retains grasp weights in FP32 however runs the ahead go and loss computation in FP16. This halves reminiscence bandwidth necessities and permits Tensor Core acceleration on fashionable NVIDIA GPUs, usually yielding a 1.5–2x throughput enchancment with negligible accuracy impression.
We use torch.autocast for the ahead go and torch.amp.GradScaler for loss scaling. A subtlety: we create the GradScaler with enabled=config.use_amp. When disabled, the scaler turns into a no-op — identical code path, zero overhead, no branching.
Gradient Accumulation
Generally you want a bigger efficient batch dimension than your GPU reminiscence permits. Gradient accumulation simulates this by operating a number of forward-backward passes earlier than stepping the optimizer. The secret’s to divide the loss by grad_accum_steps earlier than backward(), so the amassed gradient is appropriately averaged.
def train_one_epoch(mannequin, loader, criterion, optimizer, scaler, ctx, config, epoch, logger):
mannequin.prepare()
tracker = MetricTracker()
total_steps = len(loader)
use_amp = config.use_amp and ctx.system.sort == "cuda"
autocast_ctx = torch.autocast("cuda", dtype=torch.float16) if use_amp else nullcontext()
optimizer.zero_grad(set_to_none=True)
for step, (pictures, labels) in enumerate(loader):
pictures = pictures.to(ctx.system, non_blocking=True)
labels = labels.to(ctx.system, non_blocking=True)
with autocast_ctx:
outputs = mannequin(pictures)
loss = criterion(outputs, labels)
loss = loss / config.grad_accum_steps # scale for accumulation
scaler.scale(loss).backward()
if (step + 1) % config.grad_accum_steps == 0:
scaler.step(optimizer)
scaler.replace()
optimizer.zero_grad(set_to_none=True) # memory-efficient reset
# Observe uncooked (unscaled) loss for logging
raw_loss = loss.merchandise() * config.grad_accum_steps
acc = compute_accuracy(outputs, labels)
tracker.replace("loss", raw_loss, n=pictures.dimension(0))
tracker.replace("accuracy", acc, n=pictures.dimension(0))
if is_main_process(ctx.rank) and (step + 1) % config.log_interval == 0:
log_training_step(logger, epoch, step + 1, total_steps,
raw_loss, optimizer.param_groups[0]["lr"])
return trackerTwo particulars price highlighting. First, zero_grad(set_to_none=True) deallocates gradient tensors as a substitute of filling them with zeros, saving reminiscence proportional to the mannequin dimension. Second, information is moved to the GPU with non_blocking=True — this enables the CPU to proceed filling the following batch whereas the present one transfers, exploiting the pin_memory overlap.
The Important Operate
The principle() operate orchestrates the total pipeline. Be aware the attempt/lastly sample guaranteeing that the method group is torn down even when an exception happens — with out this, a crash on one rank can depart different ranks hanging indefinitely.
def major():
config = TrainingConfig.from_args()
ctx = setup_distributed(config)
logger = setup_logger(ctx.rank)
torch.manual_seed(config.seed + ctx.rank)
mannequin = create_model(config, ctx.system)
mannequin = wrap_ddp(mannequin, ctx.local_rank)
optimizer = torch.optim.SGD(mannequin.parameters(), lr=config.lr,
momentum=config.momentum,
weight_decay=config.weight_decay)
scheduler = CosineAnnealingLR(optimizer, T_max=config.epochs)
scaler = torch.amp.GradScaler(enabled=config.use_amp)
start_epoch = 1
if config.resume_from:
start_epoch = load_checkpoint(config.resume_from, mannequin.module,
optimizer, scaler, ctx.system) + 1
dataset = SyntheticImageDataset(dimension=50000, image_size=config.image_size,
num_classes=config.num_classes)
loader, sampler = create_distributed_dataloader(dataset, config, ctx)
criterion = nn.CrossEntropyLoss()
attempt:
for epoch in vary(start_epoch, config.epochs + 1):
sampler.set_epoch(epoch)
tracker = train_one_epoch(mannequin, loader, criterion, optimizer,
scaler, ctx, config, epoch, logger)
scheduler.step()
avg_loss = all_reduce_scalar(tracker.common("loss"),
ctx.world_size, ctx.system)
if is_main_process(ctx.rank):
log_epoch_summary(logger, epoch, {"loss": avg_loss})
if epoch % config.save_every == 0:
save_checkpoint(f"checkpoints/epoch_{epoch}.pt",
epoch, mannequin, optimizer, scaler, ctx.rank)
lastly:
cleanup_distributed()9. Launching Throughout Nodes
PyTorch’s torchrun (launched in v1.10 as a substitute for torch.distributed.launch) handles spawning one course of per GPU and setting the RANK, LOCAL_RANK, and WORLD_SIZE surroundings variables. For multi-node coaching, each node should specify the grasp node’s handle so that each one processes can set up the NCCL connection.
Right here is our launch script, which reads all tunables from surroundings variables:
#!/usr/bin/env bash
set -euo pipefail
NNODES="${NNODES:-2}"
NPROC_PER_NODE="${NPROC_PER_NODE:-4}"
NODE_RANK="${NODE_RANK:-0}"
MASTER_ADDR="${MASTER_ADDR:-127.0.0.1}"
MASTER_PORT="${MASTER_PORT:-12355}"
torchrun
--nnodes="${NNODES}"
--nproc_per_node="${NPROC_PER_NODE}"
--node_rank="${NODE_RANK}"
--master_addr="${MASTER_ADDR}"
--master_port="${MASTER_PORT}"
prepare.py "$@"For a fast single-node check on one GPU:
torchrun --standalone --nproc_per_node=1 prepare.py --epochs 2For 2-node coaching with 4 GPUs every, run on Node 0:
MASTER_ADDR=10.0.0.1 NODE_RANK=0 NNODES=2 NPROC_PER_NODE=4 bash scripts/launch.shAnd on Node 1:
MASTER_ADDR=10.0.0.1 NODE_RANK=1 NNODES=2 NPROC_PER_NODE=4 bash scripts/launch.sh
10. Efficiency Pitfalls and Suggestions
After constructing tons of of distributed coaching jobs, these are the errors I see most frequently:
Forgetting sampler.set_epoch(). With out it, information order is equivalent each epoch. That is the one most typical DDP bug and it silently hurts convergence.
CPU-GPU switch bottleneck. At all times use pin_memory=True in your DataLoader and non_blocking=True in your .to() calls. With out these, the CPU blocks on each batch switch.
Logging from all ranks. If each rank prints, output is interleaved rubbish. Guard all logging behind rank == 0 checks.
zero_grad() with out set_to_none=True. The default zero_grad() fills gradient tensors with zeros. set_to_none=True deallocates them as a substitute, lowering peak reminiscence.
Saving checkpoints from all ranks. A number of ranks writing the identical file causes corruption. Solely rank 0 ought to save, and all ranks ought to barrier earlier than loading.
Not seeding with rank offset. torch.manual_seed(seed + rank) ensures every rank’s information augmentation is totally different. With out the offset, augmentations are equivalent throughout GPUs.
When NOT to make use of DDP
DDP replicates the whole mannequin on each GPU. In case your mannequin doesn’t slot in a single GPU’s reminiscence, DDP alone is not going to assist. For such circumstances, look into Absolutely Sharded Knowledge Parallel (FSDP), which shards parameters, gradients, and optimizer states throughout ranks, or frameworks like DeepSpeed ZeRO.
11. Conclusion
We’ve gone from a single-GPU coaching mindset to a totally distributed, production-grade pipeline able to scaling throughout machines — with out sacrificing readability or maintainability.
However extra importantly, this wasn’t nearly making DDP work. It was about constructing it appropriately.
Let’s distill crucial takeaways:
Key Takeaways
- DDP is deterministic engineering, not magic
When you perceive course of teams, ranks, and all-reduce, distributed coaching turns into predictable and debuggable. - Construction issues greater than scale
A clear, modular codebase (config → information → mannequin → coaching → utils) is what makes scaling from 1 GPU to 100 GPUs possible. - Right information sharding is non-negotiable
DistributedSampler + set_epoch() is the distinction between true scaling and wasted compute. - Efficiency comes from small particulars
pin_memory, non_blocking, set_to_none=True, and AMP collectively ship large throughput beneficial properties. - Rank-awareness is crucial
Logging, checkpointing, and randomness should all respect rank — in any other case you get chaos. - DDP scales compute, not reminiscence
In case your mannequin doesn’t match on one GPU, you want FSDP or ZeRO — no more GPUs.
The Greater Image
What you’ve constructed right here isn’t just a coaching script — it’s a template for real-world ML methods.
This actual sample is utilized in:
- Manufacturing ML pipelines
- Analysis labs coaching massive fashions
- Startups scaling from prototype to infrastructure
And the perfect half?
Now you can:
- Plug in an actual dataset
- Swap in a Transformer or customized structure
- Scale throughout nodes with zero code modifications
What to Discover Subsequent
When you’re comfy with this setup, the following frontier is memory-efficient and large-scale coaching:
- Absolutely Sharded Knowledge Parallel (FSDP) → shard mannequin + gradients
- DeepSpeed ZeRO → shard optimizer states
- Pipeline Parallelism → break up fashions throughout GPUs
- Tensor Parallelism → break up layers themselves
These methods energy immediately’s largest fashions — however all of them construct on the precise DDP basis you now perceive.
Distributed coaching usually feels intimidating — not as a result of it’s inherently complicated, however as a result of it’s hardly ever introduced as an entire system.
Now you’ve seen the total image.
And when you see it end-to-end…
Scaling turns into an engineering determination, not a analysis downside.
What’s Subsequent
This pipeline handles data-parallel coaching — the most typical distributed sample. When your fashions outgrow single-GPU reminiscence, discover Absolutely Sharded Knowledge Parallel (FSDP) for parameter sharding, or DeepSpeed ZeRO for optimizer-state partitioning. For actually large fashions, pipeline parallelism (splitting the mannequin throughout GPUs layer by layer) and tensor parallelism (splitting particular person layers) change into needed.
However for the overwhelming majority of coaching workloads — from ResNets to medium-scale Transformers — the DDP pipeline we constructed right here is precisely what manufacturing groups use. Scale it by including nodes and GPUs; the code handles the remainder.
The whole, production-ready codebase for this undertaking is offered right here: pytorch-multinode-ddp
References
[1] PyTorch Distributed Overview, PyTorch Documentation (2024), https://pytorch.org/tutorials/newbie/dist_overview.html
[2] S. Li et al., PyTorch Distributed: Experiences on Accelerating Knowledge Parallel Coaching (2020), VLDB Endowment
[3] PyTorch DistributedDataParallel API, https://pytorch.org/docs/secure/generated/torch.nn.parallel.DistributedDataParallel.html
[4] NCCL: Optimized primitives for collective multi-GPU communication, NVIDIA, https://developer.nvidia.com/nccl
[5] PyTorch AMP: Automated Blended Precision, https://pytorch.org/docs/secure/amp.html
