# Benchmark results

**URL:** </static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fbenchmark-results%2F385%3E%5C
**Category:** 🤗Transformers\
**Created:** [July 19, 2020, 9:36pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fbenchmark-results%2F385 "2020-07-19T21:36:19Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![sshleifer](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/sshleifer/32/3304_2.png) [@sshleifer](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fsshleifer)%5C
**Post date:** [July 19, 2020, 9:36pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fbenchmark-results%2F385%2F1 "2020-07-19T21:36:19Z")

</div>

Command:

```bash
export d=mbart_benchmark_data
python examples/benchmarking/run_benchmark.py \
  --models facebook/bart-large-cnn \
  --log_filename $d/log.txt \
  --inference_memory_csv \
  $d/inference_memory.csv \
  --train_memory_csv $d/train_memory.csv \
  --train_time_csv $d/train_time.csv \
  --inference_time_csv $d/inference_time.csv \
  --fp16 --log_print --training --save_to_csv \
  --batch_sizes 4 8 12 16

```

Results:  
`bart-large-cnn`

```auto
============== TRAIN - MEMORY - RESULTS =======
--------------------------------------------------------------------------------
 Model Name Batch Size Seq Length Mem in MB
--------------------------------------------------------------------------------
   facebook/bart-large-cnn 4 8 2795
   facebook/bart-large-cnn 4 32 2897
   facebook/bart-large-cnn 4 128 3169
   facebook/bart-large-cnn 4 512 6873
   facebook/bart-large-cnn 8 8 2827
   facebook/bart-large-cnn 8 32 2933
   facebook/bart-large-cnn 8 128 3465
   facebook/bart-large-cnn 8 512 12195          
   facebook/bart-large-cnn 12 8 2859
   facebook/bart-large-cnn 12 32 3137
   facebook/bart-large-cnn 12 128 4371
   facebook/bart-large-cnn 12 512 N/A
   facebook/bart-large-cnn 16 8 2891
   facebook/bart-large-cnn 16 32 3105
   facebook/bart-large-cnn 16 128 5153
   facebook/bart-large-cnn 16 512 N/A
--------------------------------------------------------------------------------

```

`mbart`

```auto
========= TRAIN - MEMORY - RESULTS =======
--------------------------------------------------------------------------------
          Model Name Batch Size Seq Length Memory in MB
--------------------------------------------------------------------------------
  facebook/mbart-large-en-ro 4 8 4355
  facebook/mbart-large-en-ro 4 32 4947
  facebook/mbart-large-en-ro 4 128 5117
  facebook/mbart-large-en-ro 4 512 10383
  facebook/mbart-large-en-ro 8 8 4877
  facebook/mbart-large-en-ro 8 32 4493
  facebook/mbart-large-en-ro 8 128 5857
  facebook/mbart-large-en-ro 8 512 N/A
  facebook/mbart-large-en-ro 12 8 4909
  facebook/mbart-large-en-ro 12 32 5085
  facebook/mbart-large-en-ro 12 128 7079
  facebook/mbart-large-en-ro 12 512 N/A
  facebook/mbart-large-en-ro 16 8 4941
  facebook/mbart-large-en-ro 16 32 4663
  facebook/mbart-large-en-ro 16 128 8655
  facebook/mbart-large-en-ro 16 512 N/A
--------------------------------------------------------------------------------

```

---

<div class="post-metadata">

**Author:** ![sshleifer](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/sshleifer/32/3304_2.png) [@sshleifer](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fsshleifer)%5C
**Post date:** [July 19, 2020, 9:42pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fbenchmark-results%2F385%2F2 "2020-07-19T21:42:59Z")

</div>

- This assumes that len(input\_ids) == len(decoder\_input\_ids), which is not true for summarization
- since bart-large-cnn has smaller embeddings, it is less likely to OOM (have an N/A entry)
- env: v100 16GB GPU, fp16  
Train times show bart more than 2x faster

```auto
model,batch_size,sequence_length,result
facebook/mbart-large-en-ro,4,8,0.0669
facebook/mbart-large-en-ro,4,32,0.0699
facebook/mbart-large-en-ro,4,128,0.1377
facebook/mbart-large-en-ro,4,512,0.3529
facebook/mbart-large-en-ro,8,8,0.0672
facebook/mbart-large-en-ro,8,32,0.0831
facebook/mbart-large-en-ro,8,128,0.1928
facebook/mbart-large-en-ro,8,512,N/A
facebook/mbart-large-en-ro,12,8,0.0687
facebook/mbart-large-en-ro,12,32,0.1156
facebook/mbart-large-en-ro,12,128,0.2629
facebook/mbart-large-en-ro,12,512,N/A
facebook/mbart-large-en-ro,16,8,0.0705
facebook/mbart-large-en-ro,16,32,0.1392
facebook/mbart-large-en-ro,16,128,0.3334
facebook/mbart-large-en-ro,16,512,N/A

```

```auto
model,batch_size,sequence_length,result
facebook/bart-large-cnn,4,8,0.0619
facebook/bart-large-cnn,4,32,0.0629
facebook/bart-large-cnn,4,128,0.0623
facebook/bart-large-cnn,4,512,0.1274
facebook/bart-large-cnn,8,8,0.0699
facebook/bart-large-cnn,8,32,0.0628
facebook/bart-large-cnn,8,128,0.0705
facebook/bart-large-cnn,8,512,0.2347
facebook/bart-large-cnn,12,8,0.0614
facebook/bart-large-cnn,12,32,0.0620
facebook/bart-large-cnn,12,128,0.0884
facebook/bart-large-cnn,12,512,N/A
facebook/bart-large-cnn,16,8,0.0667
facebook/bart-large-cnn,16,32,0.0668
facebook/bart-large-cnn,16,128,0.1075
facebook/bart-large-cnn,16,512,N/A

```
