Transformer model on 2 GPU vs. 4 GPU


I want to run the transformer model with the parameters mentioned at:

However, the machine I am currently using has only 2 GPU. Should I adjust any of the recommended values to match the expected result?

Many thanks,

I have two GPUs. So add the parameters like:

-world_size 2 -gpu_ranks 0 1

Then keep the two GPUs working.
I also use watch -n 1 -d nvidia-smi to moniter the performance of them.
Just do it, experience is the best teacher

Thanks, Yaren, for your reply.

Yes, sure about -world_size 2 -gpu_ranks 0 1 I just wondered if I should change other parameters like -batch_size

Thanks indeed for the tip!

Kind regards,

-batch_size should be changed to fit your GPU RAM,
-train_steps depends the number of pairs of your corpus.
Otherwise, you can use the command like here:

Copyed here for your:

python -data /tmp/de2/data -save_model /tmp/extra \
        -layers 6 -rnn_size 512 -word_vec_size 512 -transformer_ff 2048 -heads 8  \
        -encoder_type transformer -decoder_type transformer -position_encoding \
        -train_steps 200000  -max_generator_batches 2 -dropout 0.1 \
        -batch_size 4096 -batch_type tokens -normalization tokens  -accum_count 2 \
        -optim adam -adam_beta2 0.998 -decay_method noam -warmup_steps 8000 -learning_rate 2 \
        -max_grad_norm 0 -param_init 0  -param_init_glorot \
        -label_smoothing 0.1 -valid_steps 10000 -save_checkpoint_steps 10000 \
        -world_size 4 -gpu_ranks 0 1 2 3 

Thanks, Yaren, for your insights!
I am currently training a model with the Transformer recommended parameters and will compare the results to the model I already trained with the default parameters for the same corpus.

@ymoslem I have the same question, did you change any of the hyperparameter values when training on a fewer number of GPUs?

You can also “replicate” 4 GPUs with 2 GPUs, by playing with the accum related parameters. Search for accum_count, gradient accumulation, etc.

1 Like

@francoishernandez can I divide the batch_size by 4, if I want to use 1 gpu instead of changing the acc_count? Does that make sense?

No. The batch_size is always considered “per GPU”.
When training on multiple GPUs, the “real” batch size is actually batch_size * num_gpus.
Gradient accumulation allows you to “simulate” bigger batches, hence real batch size si batch_size * num_gpus * accum_count.
So, switching from 4 GPUs to 1, you can simply multiply accum_count by 4.

If you divide your batch_size by 4, in addition to switching from 4 GPUs to 1, your “real batch size” will be 16x smaller. And, your GPU will probably be under utilized.