Accelerator: HPU training

Audience: Users looking to save money and run large models faster using single or multiple Gaudi devices.


What is an HPU?

Habana® Gaudi® AI Processor (HPU) training processors are built on a heterogeneous architecture with a cluster of fully programmable Tensor Processing Cores (TPC) along with its associated development tools and libraries, and a configurable Matrix Math engine.

The TPC core is a VLIW SIMD processor with an instruction set and hardware tailored to serve training workloads efficiently. The Gaudi memory architecture includes on-die SRAM and local memories in each TPC and, Gaudi is the first DL training processor that has integrated RDMA over Converged Ethernet (RoCE v2) engines on-chip.

On the software side, the PyTorch Habana bridge interfaces between the framework and SynapseAI software stack to enable the execution of deep learning models on the Habana Gaudi device.

Gaudi offers a substantial price/performance advantage – so you get to do more deep learning training while spending less.

For more information, check out Gaudi Architecture and Gaudi Developer Docs.


Run on Gaudi

To enable PyTorch Lightning to utilize the HPU accelerator, simply provide accelerator=HPUAccelerator()" parameter to the Trainer class.

from lightning_habana.pytorch.accelerator import HPUAccelerator

# run on as many Gaudi devices as available by default
trainer = Trainer(accelerator="auto", devices="auto", strategy="auto")
# equivalent to
trainer = Trainer()

# run on one Gaudi device
trainer = Trainer(accelerator=HPUAccelerator(), devices=1)
# run on multiple Gaudi devices
trainer = Trainer(accelerator=HPUAccelerator(), devices=8)
# choose the number of devices automatically
trainer = Trainer(accelerator=HPUAccelerator(), devices="auto")

The devices=1 parameter with HPUs enables the Habana accelerator for single card training. It uses SingleHPUStrategy.

The devices>1 parameter with HPUs enables the Habana accelerator for distributed training. It uses HPUDDPStrategy which is based on DDP strategy with the addition of Habana’s collective communication library (HCCL) to support scale-up within a node and scale-out across multiple nodes.

Note

HPUParallelStrategy is replaced by HPUDDPStrategy starting lightning-habana>=1.5.0 HPUParallelStrategy will be made as a base class for all hpu based parallel strategies like ddp, fsdp etc.


Scale-out on Gaudis

To train a Lightning model using multiple HPU nodes, set the num_nodes parameter with the available nodes in the Trainer class.

from lightning_habana.pytorch.accelerator import HPUAccelerator
from lightning_habana.pytorch.strategies import HPUDDPStrategy

hpus = 8
parallel_hpus = [torch.device("hpu")] * hpus
trainer = Trainer(accelerator=HPUAccelerator(), devices=hpus, strategy=HPUDDPStrategy(parallel_devices=parallel_hpus), num_nodes=2)

In addition to this, the following environment variables need to be set to establish communication across nodes.

  • MASTER_PORT - required; has to be a free port on machine with NODE_RANK 0

  • MASTER_ADDR - required (except for NODE_RANK 0); address of NODE_RANK 0 node

  • WORLD_SIZE - required; how many workers are in the cluster

  • NODE_RANK - required; id of the node in the cluster

The trainer needs to be instantiated on every node participating in the training.

On Node 1:

MASTER_ADDR=<MASTER_ADDR> MASTER_PORT=<MASTER_PORT> NODE_RANK=0 WORLD_SIZE=16
    python -m some_model_trainer.py (--arg1 ... train script args...)

On Node 2:

MASTER_ADDR=<MASTER_ADDR> MASTER_PORT=<MASTER_PORT> NODE_RANK=1 WORLD_SIZE=16
    python -m some_model_trainer.py (--arg1 ... train script args...)

How to access HPUs

To use HPUs, you must have access to a system with HPU devices.

AWS

You can either use Gaudi-based AWS EC2 DL1 instances or Supermicro X12 Gaudi server to get access to HPUs.

Check out the PyTorch Model on AWS DL1 Instance Quick Start.