ZeRO++ is a system of communication optimization strategies built on top of ZeRO to offer unmatched efficiency for large model training regardless of the scale or cross-device bandwidth constraints. Read our ZeRO++ blog and paper to learn more!

We recommend that you read the tutorials on Getting Started, ZeRO and Megatron-DeepSpeed before stepping through this tutorial.

Three Components of ZeRO++

ZeRO++ consists of three key designs, namely quantized weights (qwZ), hiearchical partitioning ZeRO (hpZ), and quantized gradients (qgZ):

  • qwZ applies block-based quantization to reduce ZeRO parameter all-gather communication volume by half from FP16 to INT8)
  • hpZ eliminates inter-node backward parameter all-gather communication through data remapping and recomputation
  • qgZ replaces gradients allreduce collective with a new communication efficient all-to-all based quantized gradient averaging.

Collectively, the three optimization reduces communication volume by 4x compared to ZeRO baseline. Each of the three components can be enabled independent of each other and collectively as a group as described in the next section.

Training Environment

For this tutorial, we will configure a 18 billion parameter GPT-2 model using the DeepSpeed Megatron-DeepSpeed GPT-2 code. We will use 4 nodes of 16x NVIDIA Tesla V100-SXM3 Tensor Core GPU with 32GB RAM per node for this exercise.

Training a 18B parameter GPT-2 with ZeRO++

There are no change needed to the user code. However, since ZeRO++ extends ZeRO Stage 3 (ZeRO-3), appropriate flags need to be added to activate each or all of the three ZeRO++ communication collective optimizations. The three flags and their meanings and defaults and preferred values:

  • zero_quantized_weights: Boolean indicating whether to use quantized zero weights (qwZ), default is false
  • zero_hpz_partition_size: number of ranks in hpZ (secondary partition) group, default is 1 meaning no hpZ, ideal is number of ranks (gpus) per node
  • zero_quantized_gradients: Boolean indicating whether to use quantized zero gradients (qgZ), default is false

DeepSpeed Configuration Changes

An example snippet of deepspeed configurations with all three ZeRO++ optimization enable is shown below:

    "zero_optimization": {
        "stage": 3,
        "reduce_bucket_size": 10000000,
        "reduce_scatter": true,

        "zero_quantized_weights": true,
        "zero_hpz_partition_size": 16,
        "zero_quantized_gradients": true,

        "contiguous_gradients": true,
        "overlap_comm": true

Finally, to launch your experiment, issue the following command:

       deepspeed \
       --tensor-model-parallel-size 1 \
       --pipeline-model-parallel-size 1 \
       --num-layers 40 \
       --hidden-size 6144 \
       --seq-length 512 \
       --num-attention-heads 32 \
       --batch-size 1 \
       --zero-stage 3 \
       --deepspeed_config ds_zeropp_config.json \
       --deepspeed-activation-checkpointing \
       --fp16 \

See more details on Megatron-DeepSpeed tutorial examples on how to launch a Megatron-DeepSpeed job.

Here is a screenshots of the training log for both ZeRO baseline and ZeRO++:

ZeRO baseline


Congratulations! You have completed the ZeRO++ tutorial.