Train your team

AIC-230 · GPU & AI Compute

Distributed Training with PyTorch

From a single-GPU training loop to sharded multi-node training that survives a node failure.

Practitioner 4 days in person8 half-days online Max 14 in person

Who this course is for

ML and platform engineers scaling PyTorch training beyond one GPU — who need the distributed machinery (DDP, FSDP, DeepSpeed, Megatron) to actually work, not just to import.

Prerequisites

Python proficiencyBasic deep-learning concepts (training loop, backprop)AIC-220 helpful but not required

Course outline

Day 1 — Deep learning foundations for systems engineers

  • Architectures: MLP, CNN, RNN, Transformer
  • Backpropagation and automatic differentiation
  • Optimizers: SGD, Adam, AdamW, LAMB; LR scheduling
  • Loss functions and evaluation metrics
  • Regularization; complete training workflow: DataLoader, loop, validation, checkpointing

Day 2 — PyTorch for AI development

  • Tensors, autograd, nn.Module
  • Data loading: Dataset, DataLoader, multi-process workers
  • GPU acceleration: .to(device), pin_memory
  • Mixed precision with torch.cuda.amp and GradScaler
  • PyTorch Profiler and Nsight integration
  • Export: TorchScript, torch.compile/TorchInductor

Day 3 — Data-parallel training

  • PyTorch DDP: init_process_group, DistributedSampler, all-reduce
  • DDP debugging and common failure modes
  • PyTorch FSDP: FULL_SHARD, SHARD_GRAD_OP, CPU offloading
  • Mixed precision at scale: BF16, FP8 with Transformer Engine
  • Checkpointing: sync/async, sharded formats

Day 4 — Large-model frameworks and fault tolerance

  • DeepSpeed ZeRO stages 1/2/3; CPU/NVMe offloading; 3D parallelism
  • Megatron-LM: tensor and pipeline parallelism, interleaved schedules
  • Fault-tolerant training: torchrun --max_restarts, elastic training
  • Choosing a parallelism strategy for a given model and cluster
  • Capstone workshop

Hands-on labs

Labs follow the academy model — 35% principles, 20% guided investigation, 45% engineering studio. Every claim you make in a lab is backed by a trace, a counter or a measurement you captured yourself. How we teach

  1. Lab: build a complete PyTorch training pipeline with mixed precision and profiling; fix the data-loading bottleneck you find
  2. Lab: convert single-GPU training to DDP across multiple GPUs; verify gradient synchronisation
  3. Lab: shard a model with FSDP and measure memory savings vs DDP
  4. Lab: configure DeepSpeed ZeRO for a model that does not fit on one GPU; compare stage 2 vs 3
  5. Lab: kill a worker mid-training and prove elastic recovery from a sharded checkpoint

Capstone project

Scale a transformer training job from one GPU to a multi-GPU (and multi-node, where available) setup: choose and justify the parallelism strategy, reach a target scaling efficiency, and demonstrate fault-tolerant checkpoint/restart — with profiler evidence for each decision.

What you leave with

  • A production PyTorch pipeline: data, mixed precision, profiling, export
  • Hands-on DDP, FSDP and DeepSpeed ZeRO experience
  • A decision framework for data/tensor/pipeline parallelism
  • Fault-tolerant training patterns that survive real clusters

How it runs

Every course follows the same model: 35% principles, 20% guided investigation, 45% engineering studio. You leave with working code, raw measurements and an evidence-based report — not a certificate of attendance. Read the methodology or see a full sample lesson.

Material is adapted to your kernel version, hardware and workload before a private delivery. For public cohorts, the environment is provided and configured.

Questions

Who is this course for?

ML and platform engineers scaling PyTorch training beyond one GPU — who need the distributed machinery (DDP, FSDP, DeepSpeed, Megatron) to actually work, not just to import. It sits at practitioner level within the GPU & AI Compute track.

What do I need to know already?

Specific prerequisites for this course: Python proficiency; Basic deep-learning concepts (training loop, backprop); AIC-220 helpful but not required. We confirm levels before the cohort starts and adapt if a group is stronger or weaker than expected.

Can this run privately for my team?

Yes. Any course runs on-site at your offices anywhere, or live online for a distributed team, with labs adapted to your hardware and codebase.

What is the difference between in-person and online?

In person is 4 full days with hardware on your desk, capped at 14. Online is 8 half-day sessions across about two weeks so you can keep working, capped at 20, with remote lab access.

Do you invoice companies?

Yes. Purchase orders are accepted and invoicing is available in USD, EUR, GBP, SAR and CAD.

Upcoming dates

DatesWhereSeatsEarly birdRegular
15 Nov – 18 Nov 20264 full days RiyadhIn person · KAFD Conference Centre 4 of 14 SAR 9,450until 16 OctSAR 10,500
22 Nov – 25 Nov 20264 full days Kuwait CityIn person · Al Hamra Tower 9 of 14 KWD 780until 23 OctKWD 870
22 Nov – 25 Nov 20264 full days MuscatIn person · Knowledge Oasis Muscat 4 of 14 OMR 970until 23 OctOMR 1,080
29 Nov – 8 Dec 20268 half-days Gulf bandLive online · 09:00–13:00 GMT+3 16 of 20 US$1,800until 30 OctUS$2,000
30 Nov – 3 Dec 20264 full days OttawaIn person · Kanata North Tech Park 9 of 14 CAD 3,430until 31 OctCAD 3,810
30 Nov – 9 Dec 20268 half-days Europe bandLive online · 09:00–13:00 CET 5 of 20 US$1,800until 31 OctUS$2,000
7 Dec – 10 Dec 20264 full days TorontoIn person · MaRS Discovery District 4 of 14 CAD 3,430until 7 NovCAD 3,810
7 Dec – 10 Dec 20264 full days LondonIn person · Shoreditch Works 9 of 14 GBP 1,960until 7 NovGBP 2,180
7 Dec – 16 Dec 20268 half-days Americas bandLive online · 13:00–17:00 ET 10 of 20 US$1,800until 7 NovUS$2,000
14 Dec – 17 Dec 20264 full days BerlinIn person · Factory Görlitzer Park 4 of 14 EUR 2,320until 14 NovEUR 2,580

Dates shown for the next few months. If nothing fits, tell us where and when — cohorts are added on demand, and private delivery can be scheduled any week.

More in GPU & AI Compute

AIC-1003 days Foundations for AI Compute The architecture, operating system and networking groundwork every GPU systems engineer is assumed to have and often does not. Foundation Practitioner-taught SAR 6,750Next 25 Oct AIC-1103 days GPU Architecture, Memory & Interconnects How the hardware constrains your workload: SIMT execution, the memory hierarchy and the fabric between GPUs. FoundationNCA-AIIO Practitioner-taught SAR 6,750Next 11 Oct AIC-2004 days Linux for GPU Systems What the kernel is doing underneath your training job, and how to tune it. The layer almost nobody teaches. Practitioner Practitioner-taught SAR 10,500Next 15 Nov AIC-2104 days RDMA & AI Cluster Networking Build and debug the fabric distributed training runs on, from queue pairs up to a tuned NCCL all-reduce. PractitionerNCP-AIN Practitioner-taught SAR 10,500Next 1 Nov AIC-2205 days CUDA & HIP Programming Write, profile and optimise GPU kernels on both vendors, including the CUDA-to-HIP porting path. Practitioner Practitioner-taught SAR 13,120Next 11 Oct AIC-3004 days Containers, Kubernetes & GPU Schedulers Run a shared GPU cluster multiple teams can actually use: partitioning, scheduling, quotas and isolation. PractitionerNCP-AII Practitioner-taught SAR 10,500Next 18 Oct AIC-3103 days Storage & Data Pipelines for AI Stop starving your GPUs: parallel filesystems, GPUDirect Storage and pipelines built for sustained throughput. Practitioner Practitioner-taught SAR 7,880Next 22 Nov AIC-3203 days MLOps & Inference Serving Get models off a laptop and onto a GPU endpoint that scales, with the pipeline machinery around them. PractitionerNCP-AIO Practitioner-taught SAR 7,880Next 1 Nov AIC-3302 days AI Infrastructure Security & Observability Harden a multi-tenant GPU platform and see what it is doing before users report a problem. Advanced Practitioner-taught SAR 6,000Next 18 Oct AIC-4003 days Large-Scale Training & Datacenter Architecture The thousand-GPU conversation: 3D parallelism, reference architectures, TCO and where the hardware is heading. Advanced Practitioner-taught SAR 9,000Next 8 Nov