FSDP (Fully Sharded Data Parallel) Setup for Training

Model LLaMA-2 70B does not fit into A100 80GB memory when using DDP. FSDP solves this by sharding parameters, gradients, and optimizer states across GPUs. We configure FSDP turnkey—the native fully sharded data parallelism implementation in PyTorch that saves up to 70% memory without losing speed. Certified engineers with many years of experience in distributed training. Over the course of our work, we have completed more than 50 projects for models ranging from 1B to 70B parameters. Our clients save up to 40% on cloud GPU budgets thanks to optimal configuration.

PyTorch FSDP documentation

Why FSDP over DeepSpeed?

FSDP is part of PyTorch core and requires no external dependencies. Unlike DeepSpeed ZeRO-3, integration with Hugging Face Transformers and Accelerate goes through native APIs. We use FSDP in every second project for fine-tuning large models—from LLaMA to Mistral. PyTorch FSDP documentation

How FSDP works

Principle of operation

During forward pass: parameters of each sharded layer are gathered (all-gather) from all GPUs before computation. After forward—immediately freed if reshard_after_forward is enabled. During backward pass: parameters are gathered again, gradients computed, then reduce-scatter distributes gradient shards across GPUs. This eliminates the situation where each GPU stores a full copy of the model, as in regular DDP.

Basic setup

import torch
import torch.distributed as dist
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.fsdp.fully_sharded_data_parallel import (
    CPUOffload,
    BackwardPrefetch,
)
from torch.distributed.fsdp.wrap import (
    size_based_auto_wrap_policy,
    enable_wrap,
    wrap,
)
import functools

def setup_fsdp(rank, world_size):
    dist.init_process_group("nccl", rank=rank, world_size=world_size)
    torch.cuda.set_device(rank)

def wrap_model_with_fsdp(model, rank):
    auto_wrap_policy = functools.partial(
        size_based_auto_wrap_policy,
        min_num_params=100_000_000
    )

    model = FSDP(
        model,
        auto_wrap_policy=auto_wrap_policy,
        cpu_offload=CPUOffload(offload_params=False),
        backward_prefetch=BackwardPrefetch.BACKWARD_PRE,
        device_id=torch.cuda.current_device(),
        sharding_strategy=ShardingStrategy.FULL_SHARD,
        mixed_precision=MixedPrecision(
            param_dtype=torch.bfloat16,
            reduce_dtype=torch.float32,
            buffer_dtype=torch.bfloat16,
        ),
    )
    return model

How to choose a sharding strategy

from torch.distributed.fsdp import ShardingStrategy

# FULL_SHARD — full sharding (analogous to ZeRO-3)
strategy = ShardingStrategy.FULL_SHARD
# SHARD_GRAD_OP — shard only gradients and optimizer (ZeRO-2)
strategy = ShardingStrategy.SHARD_GRAD_OP
# NO_SHARD — regular DDP
strategy = ShardingStrategy.NO_SHARD
# HYBRID_SHARD — FULL_SHARD within node, replication between nodes
strategy = ShardingStrategy.HYBRID_SHARD

The choice of strategy depends on model size, number of GPUs, and interconnect speed. For 8 GPUs with NVLink, FULL_SHARD is optimal; for multi-node, HYBRID_SHARD.

Sharding strategies: memory vs. speed comparison

How to configure FSDP: step-by-step instructions

1. Determine cluster topology: number of GPUs, nodes, interconnect type (NVLink, InfiniBand).
2. Choose sharding strategy: FULL_SHARD for a single node with NVLink, HYBRID_SHARD for multi-node.
3. Configure mixed precision: use bfloat16 for parameters, float32 for reductions.
4. Override wrap policy: for transformers, use transformer_auto_wrap_policy with the layer class specified.
5. Optimize checkpointing: enable offload_to_cpu when saving full state dict.
6. Profile performance: measure throughput, GPU utilization, and p99 latency.

Practical configuration aspects

Wrap policy for transformers

For transformers, it is important to wrap each Transformer block individually:

from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy
from transformers.models.llama.modeling_llama import LlamaDecoderLayer

llama_auto_wrap_policy = functools.partial(
    transformer_auto_wrap_policy,
    transformer_layer_cls={LlamaDecoderLayer},
)
model = FSDP(model, auto_wrap_policy=llama_auto_wrap_policy)

Saving and loading checkpoints

from torch.distributed.fsdp import FullStateDictConfig, StateDictType

save_policy = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, save_policy):
    cpu_state = model.state_dict()
    if rank == 0:
        torch.save(cpu_state, "checkpoint.pt")

if rank == 0:
    state_dict = torch.load("checkpoint.pt")
else:
    state_dict = {}
with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, save_policy):
    model.load_state_dict(state_dict)

Integration with Hugging Face Accelerate

from accelerate import Accelerator
from accelerate.utils import FullyShardedDataParallelPlugin
from torch.distributed.fsdp.fully_sharded_data_parallel import FullOptimStateDictConfig, FullStateDictConfig

fsdp_plugin = FullyShardedDataParallelPlugin(
    state_dict_config=FullStateDictConfig(offload_to_cpu=True, rank0_only=False),
    optim_state_dict_config=FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=False),
)
accelerator = Accelerator(fsdp_plugin=fsdp_plugin)

How we configured FSDP for LLaMA-70B

In one project, we needed to fine-tune LLaMA-2 70B on 8x A100 80GB. Initially, the model did not fit even with DeepSpeed ZeRO-3. We chose FSDP with FULL_SHARD and hybrid bfloat16 precision, configured transformer_auto_wrap_policy and backward prefetch. The throughput was 850 tokens/s with batch size 4 per GPU. Memory savings: 68% compared to DDP. Additionally, we reduced epoch time by 30% through communication optimization. The client saved over 35% on GPU rental costs.

What's included in FSDP setup

• Model and GPU configuration audit
• Selection of optimal sharding strategy and mixed precision
• Wrap policy tuning per architecture (transformers, CNNs, GNNs)
• Integration with Accelerate and Hugging Face Trainer
• Checkpoint optimization and loading
• Performance profiling (throughput, memory, GPU utilization)
• Documentation and training for your team
• Post-deployment support

Typical mistakes when configuring FSDP

• OOM when saving checkpoint: use FullStateDictConfig with offload_to_cpu=True.
• Slow initialization: try HYBRID_SHARD for multi-node.
• Incompatibility with some layers: verify auto_wrap_policy on all submodules.

Setup timeframe: 5 to 10 business days. Cost is calculated individually after a free consultation. Contact us to discuss your task. Order a turnkey FSDP setup—get a consultation from a certified engineer.

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.