Complete Guide to Distributed Training Setup for Any Model Size

We configure end-to-end distributed training for models of any size — from millions to hundreds of billions of parameters. We will assess your project and propose the optimal parallelism strategy, guaranteeing GPU utilization >85% and MFU up to 50% on A100/H100 clusters. Our experience spans 5+ years training large models on clusters of up to 256 GPUs, with over 30 successful projects for AI startups and enterprises.

Why Distributed Training Is Critical for Large Models

A core problem: the model does not fit in a single GPU memory. For instance, LLaMA 3 70B requires about 140GB for weights alone in FP16, exceeding even the H100 80GB capacity. Distributing across several GPUs is the only option. Additionally, training on a single GPU would take weeks; parallelism cuts time by orders of magnitude. In practice, we achieved an 8x speedup on an 8-H100 cluster for a 13B parameter model — with the right configuration. Proper distributed training can reduce GPU costs by 30-50% compared to naive single-GPU approaches. Our clients typically save $5k+ per month on cloud bills through optimized parallelism.

Main Parallelism Strategies

• Data Parallelism — each GPU holds a full model copy and processes different parts of the batch. Gradients are aggregated (all-reduce) after each step. Suitable for models that fit into a single GPU memory.
• Model Parallelism (Tensor Parallelism) — the model is split across layers or tensors among GPUs. Necessary when the model is too large for one GPU. Used in Megatron-LM and DeepSpeed.
• Pipeline Parallelism — model layers are distributed across GPUs sequentially. Different GPUs process different micro-batches simultaneously. Used in GPipe, PipeDream.
• 3D Parallelism — a combination of all three strategies. Used by DeepSpeed and Megatron-LM for training LLMs with hundreds of billions of parameters. DDP (Data Parallel) outperforms naive all-reduce by 30-40% in speed for batches >128 per GPU, and with ZeRO-3 efficiency approaches Model Parallel on large models.

How to Choose a Parallelism Strategy

For most projects, the optimal choice is to start with DDP and DeepSpeed ZeRO, and only move to Model/Pipeline Parallel if the model does not fit in memory. More details on DeepSpeed ZeRO (https://www.deepspeed.ai/tutorials/zero/) can be found in the official documentation.

What Is 3D Parallelism and When to Use It?

3D Parallelism combines Data, Model, and Pipeline parallelism. It is the only way to train models with hundreds of billions of parameters (e.g., GPT-4 or LLaMA 3 405B). It requires careful tuning: balancing the number of micro-batches, tensor sizes, and pipeline stages. In DeepSpeed and Megatron-LM, this is automated through JSON configs. We use 3D Parallelism on clusters of 64+ GPUs. Typical configuration: 8-way tensor, 4-way pipeline, 8-way data (ZeRO-1). Proper 3D parallelism can achieve up to 2x higher MFU compared to naive DDP for models >10B parameters.

Comparison of Parallelism Methods

Data Parallel with PyTorch DDP

DistributedDataParallel (DDP) — the recommended approach for data parallelism in PyTorch. PyTorch DDP Documentation (https://pytorch.org/docs/stable/distributed.html)

import torch
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP

def setup(rank, world_size):
    dist.init_process_group(
        backend='nccl',  # nccl for GPU, gloo for CPU
        rank=rank,
        world_size=world_size
    )
    torch.cuda.set_device(rank)

def train(rank, world_size, model, dataset):
    setup(rank, world_size)

    model = model.to(rank)
    ddp_model = DDP(model, device_ids=[rank])

    sampler = DistributedSampler(dataset, num_replicas=world_size, rank=rank)
    loader = DataLoader(dataset, sampler=sampler, batch_size=32)

    optimizer = torch.optim.AdamW(ddp_model.parameters(), lr=1e-4)

    for epoch in range(num_epochs):
        sampler.set_epoch(epoch)
        for batch in loader:
            optimizer.zero_grad()
            loss = ddp_model(batch)
            loss.backward()
            optimizer.step()

Running on a single node (8 GPUs):

torchrun --nproc_per_node=8 train.py

Running on multiple nodes:

# On node 0 (master):
torchrun --nnodes=4 --nproc_per_node=8 \
  --node_rank=0 \
  --master_addr="10.0.0.1" --master_port=29500 \
  train.py

# On nodes 1-3 (worker):
torchrun --nnodes=4 --nproc_per_node=8 \
  --node_rank=1 \  # 2, 3 respectively
  --master_addr="10.0.0.1" --master_port=29500 \
  train.py

Accelerate from Hugging Face

For simpler configuration with support for mixed precision, gradient accumulation, and various distributed backends:

from accelerate import Accelerator

accelerator = Accelerator(
    mixed_precision='bf16',
    gradient_accumulation_steps=4
)

model, optimizer, train_dataloader = accelerator.prepare(
    model, optimizer, train_dataloader
)

for batch in train_dataloader:
    with accelerator.accumulate(model):
        outputs = model(**batch)
        loss = outputs.loss
        accelerator.backward(loss)
        optimizer.step()
        optimizer.zero_grad()

Step-by-Step Guide to Setting Up Distributed Training

1. Analyze model size and memory requirements.
2. Select parallelism strategy (use the table above).
3. Configure network: NCCL backend, InfiniBand or RoCE, master node.
4. Prepare training script with DDP, DeepSpeed or Megatron.
5. Launch with torchrun or slurm on the cluster.
6. Monitor GPU utilization and MFU, adjust batch size and gradient accumulation.

Deliverables

• Detailed documentation of the parallelism strategy and configuration.
• Configuration files for DDP, DeepSpeed, or Megatron.
• Monitoring dashboard (W&B or MLflow) with key metrics.
• Training session (2 hours) to walk through the setup.
• 1 month of post-launch support.

What Is Included in Distributed Training Setup

• Analysis of model architecture and selection of parallelism strategy.
• Configuration of distributed environment (NCCL, InfiniBand, MPI).
• Setup of DDP/DeepSpeed/Megatron tailored to your hardware.
• Hyperparameter optimization (batch size, learning rate, gradient accumulation).
• Integration of monitoring (W&B, MLflow) and logging.
• Deployment documentation and post-launch support.

Our Experience and Guarantees

We guarantee:

• GPU utilization >85% and MFU no lower than 40% for typical configurations.
• Support for all popular frameworks: PyTorch DDP, DeepSpeed ZeRO-2/3, Megatron-LM, Hugging Face Accelerate.
• Experience with clusters on AWS, GCP, On-premise (NVIDIA DGX, Supermicro).
• Individual approach: we do not propose template solutions but adapt to your model and hardware.

Trusted by 15+ AI startups, with 5+ years in distributed training and 30+ completed projects. Typical setup fee starts at $3,000. Contact us for an assessment of your project. Get a consultation on strategy selection and configuration.

MLOps: Infrastructure for Training, Deploying, and Monitoring ML Models

The model is trained, metrics — F1 0.94 on validation. Three months later in production, quality drops by 12%. No one knows when — there is no monitoring. It's impossible to retrain quickly — the training script is in a Jupyter notebook of a data scientist who has already left. Data for retraining is collected manually from three disparate systems. About half of the projects come to us with this pain. We build a turnkey MLOps platform: from experiment tracking to automatic deployment and data drift monitoring. We will assess your infrastructure in 1–2 weeks, and in 4–6 weeks you will get a basic MLOps core running in production. Our team has 10+ years of experience in ML infrastructure, over 50 implementations.

How does MLOps infrastructure benefit your ML projects?

Experiment Tracking and Reproducibility

Without tracking, an ML project turns into chaos: it's unclear which checkpoint is better, which hyperparameters were used, which dataset. Reproducing a result a month later is a quest.

Why is experiment tracking the foundation of reproducibility?

MLflow is an open source standard for tracking. It logs parameters, metrics, artifacts (models, graphs), and code. MLflow Model Registry is a centralized model storage with versioning and lifecycle stages (Staging → Production → Archived). Deployment via MLflow Serving or integration with external systems.

Typical initialization in code:

import mlflow

mlflow.set_experiment("fraud-detection-v2")
with mlflow.start_run():
    mlflow.log_params({"learning_rate": 3e-4, "batch_size": 64, "epochs": 10})
    mlflow.log_metric("val_f1", val_f1, step=epoch)
    mlflow.pytorch.log_model(model, "model")

This is the minimum. In production, we add logging of system metrics (GPU utilization, memory), dataset (hash, version), code (git commit hash). Weights & Biases — richer UI, collaboration features, sweep for hyperparameter optimization. MLflow — for on-premise deployment without external dependencies.

DVC (Data Version Control) — versioning of data and models on top of git. Data is stored in S3/GCS/Azure Blob, only metadata (hashes) in git. dvc repro reproduces the entire pipeline from raw data to metrics.

To ensure reproducibility of training, fix random seeds (torch.manual_seed, numpy.random.seed, random.seed) and record them in experiment metadata. Without this, debugging irregular results is painful. Log the dataset version (DVC hash) and git commit — then any experiment can be reproduced down to the byte.

Pipeline Orchestration: Kubeflow, Airflow, Prefect

A pipeline orchestrator becomes necessary when: A 100-line training script in cron is fine for simple tasks. But as soon as you have a multi-step pipeline (data loading → preprocessing → feature engineering → training → validation → deployment if quality above threshold), you need an orchestrator with retry logic, visualization, and alerts.

Kubeflow — Kubernetes-native orchestrator for ML (see Kubeflow). Each step is a Docker container. Supports parallel steps, conditional branches, artifacts between steps. Integrates with Katib (AutoML), KServe (serving), Feast (feature store).

Apache Airflow — more general DAG orchestrator. Wide ecosystem of operators (S3, Spark, DBT, Kubernetes). Easier to deploy if Airflow already exists in the company.

Prefect / Metaflow — less boilerplate. Prefect 2.x with @flow and @task decorators — quick start for small teams.

Typical training pipeline architecture on Kubeflow:

1. Data ingestion component — fetches data from S3/DB, validates schema via Great Expectations
2. Preprocessing component — transformations, normalization, train/val/test split
3. Training component — training on GPU, logging to MLflow
4. Evaluation component — metric calculation, comparison with baseline in Model Registry
5. Conditional deployment — deploy only if new model is better than current by >2% F1

Each component is a separate Docker image. Pipeline is versioned in git. Scheduled run (retraining once a week on new data) or manual.

Model Registry and Lifecycle Management

Model Registry is not just a checkpoint store. It is a centralized system that knows:

• Which model is currently in production (and with what metrics)
• History of all versions with training parameters
• Metadata: dataset, git commit, validation results
• Lifecycle stage: None → Staging → Production → Archived

MLflow Model Registry — standard. For enterprise — Vertex AI Model Registry (GCP), SageMaker Model Registry (AWS), Azure ML Model Registry.

Model promotion through stages: automatically move model to Staging after successful eval, then manual or automatic (during A/B test) promotion to Production. Rollback — switch to previous Production version in seconds.

Serving: From FastAPI to Triton Inference Server

Simple case. FastAPI + PyTorch/ONNX on one server — 80% of production ML deployments are exactly that. Sufficient for most tasks with load up to 100 req/s.

from fastapi import FastAPI
import onnxruntime as ort

app = FastAPI()
session = ort.InferenceSession("model.onnx", providers=["CUDAExecutionProvider"])

@app.post("/predict")
async def predict(request: PredictRequest):
    inputs = preprocess(request.text)
    outputs = session.run(None, {"input_ids": inputs})
    return {"label": postprocess(outputs)}

Triton Inference Server — production standard for high loads (500+ req/s). Dynamic batching, concurrent model execution, model ensemble. Supports TensorRT, ONNX, PyTorch TorchScript, TensorFlow SavedModel.

KServe — Kubernetes-native ML serving with autoscaling, canary deployments, A/B testing out of the box. Scale-to-zero for inactive models — savings on infrastructure up to 40% annually for a project with 10 models.

Monitoring: Data Drift, Model Drift, Infrastructure Metrics

Monitoring — what is usually done last and regretted first. Three levels.

Infrastructure monitoring. Latency (P50/P95/P99), throughput (req/s), error rate (4xx, 5xx), GPU/CPU utilization. Prometheus + Grafana — standard. Alert when P99 latency > threshold or error rate > 1%.

Data drift monitoring. Distribution of input data changes over time. Detect via PSI (Population Stability Index) for numerical features: PSI > 0.2 — strong drift. Chi-squared test for categorical, Kolmogorov-Smirnov test for continuous. Evidently AI — open source library with ready-made drift tests.

Model drift monitoring. If ground truth is delayed (e.g., we know conversion after a week) — monitor real metrics. If not — surrogate metrics: distribution of prediction scores, proportion of confident predictions.

Alerting. Three levels: INFO (minor drift, log it), WARNING (significant, notify team), CRITICAL (quality dropped below threshold — automatic switch to fallback model).

Why is data drift monitoring important?

Without it, you learn about model degradation only from user complaints or ringing SLA. A drift alert allows you to retrain the model in advance, before errors start causing losses. In one of our projects, PSI monitoring detected drift 2 days after a data source change — this saved the campaign.

Feature Store

Feature Store solves the training-serving skew problem. If preprocessing during training and inference is implemented in two different places — divergence is inevitable.

A Feature Store is needed when:

• Several models use the same features
• Features are computed from streaming data (real-time)
• Large team with different people on feature engineering and model training

Feast — open source Feature Store. Offline store (S3 + Parquet) for training, online store (Redis, DynamoDB) for low-latency inference. Feature definitions as code, materialization job syncs offline → online.

Tecton (commercial), Vertex AI Feature Store (GCP), SageMaker Feature Store (AWS) — managed options with less ops overhead.

CI/CD for ML

ML CI/CD is regular CI/CD plus specific ML steps.

ML-specific checks in CI:

• Reproducibility check: run training with a fixed seed, result must match
• Data validation: Great Expectations or Pandera on schema/distribution checks
• Model performance check: automatic eval on holdout, block merge if degradation > threshold
• Latency regression test: inference must meet SLA

GitOps for deployment. Merge to main → CI triggers training → eval → if passes → automatic deployment to Staging → smoke tests → manual promotion to Production or automatic upon successful canary.

Tools: GitHub Actions / GitLab CI for CI, ArgoCD for GitOps deployment on Kubernetes.

What's Included in MLOps Platform Development

We provide a full cycle of work, documentation, and team training.

Total time from audit to full MLOps platform: 3–5 months. Also possible phased launch: basic level (tracking + serving) in 4–6 weeks.

Cost is calculated individually based on data volume, number of models, and infrastructure requirements. Order an MLOps infrastructure audit — get a roadmap in 1–2 weeks. Contact us for a project assessment — we will send a preliminary estimate within 2 business days.

Note: warranty on architectural solutions — 12 months. We provide integration certificates with major cloud providers (AWS, GCP, Azure). During our work, we have not lost a single client after the first implementation — the experience of 50+ successful MLOps projects speaks for itself. Get a consultation on building an MLOps platform today.