This model has the following configuration:

• 48-layer
• 1536 hidden dimension
• 1.5B parameters.

Refer to DeBERTa: Decoding-enhanced BERT with Disentangled Attention.

The training use fp16 and runs on 1 node with 16 Nvidia V100 GPUs. The autotuning uses the same hardware resource as the training. max_train_batch_size is not defined. The HF packages below are used.

HF examples require installing the transformers package from source:

    git clone https://github.com/huggingface/transformers.git
    cd transformers
    pip install .

The datasets package can be installed by pip install datasets

Below are the versions used in this test.

• transformers (4.12.0)
• datasets (1.11.0)

The table below shows the throughput (samples per second) comparison. The corresponding train micro-batch size per GPU (mbs or tmbspg) and ZeRO stage used to achieve the throughput value is also shown in the parentheses. Assume the strategy users would use in the handtuning process is to start from mbs = 1 and increase mbs by 2 each time until running out of GPU memory.

• baseline is the vanila HF without DeepSpeed (DS) and mbs is hand-tuned.
• HF + DS hand-tuned is HF with DS, and mbs is hand-tuned while other DS configuration uses default values.
• HF + DS autotuning is HF with DS, and the DS configuration is selected from autotuning.

Notation: Hugging Face (HF), DeepSpeed (DS), ZeRO stage (z), gradient accumulation steps (gas), train micro-batch size per GPU (mbs or tmbspg), reduce_bucket_size (rbs), allgather_bucket_size (abs).

Note that the performance metric used in autotuning is calculated using the timings captured within DeepSpeed forward, backward, and step functions. The sum of these timings is less than the actual training step latency, thus the throughput metric values used by autotuning would be higher than the end-to-end throughput in training.

• Autotuning time: 40 mins
• Number of experiments: 12
• Throughput Improvement over baseline: Inf

Tuning completed in 0:39:25.253998. Total number of experiments: 12.

• Autotuning time: 1 hr 2 mins
• Number of experiments: 24
• Throughput Improvement over baseline: Inf

Tuning completed in 1:02:32.759424. Total number of experiments: 24.