Distributed support

[lhoestq]

To split your dataset across your training nodes, you can use the new [datasets.distributed.split_dataset_by_node]:

import os
from datasets.distributed import split_dataset_by_node

ds = split_dataset_by_node(ds, rank=int(os.environ["RANK"]), world_size=int(os.environ["WORLD_SIZE"]))

This works for both map-style datasets and iterable datasets.
The dataset is split for the node at rank rank in a pool of nodes of size world_size.

For map-style datasets:

Each node is assigned a chunk of data, e.g. rank 0 is given the first chunk of the dataset.

For iterable datasets:

If the dataset has a number of shards that is a factor of world_size (i.e. if dataset.n_shards % world_size == 0),
then the shards are evenly assigned across the nodes, which is the most optimized.
Otherwise, each node keeps 1 example out of world_size, skipping the other examples.

This can also be combined with a torch.utils.data.DataLoader if you want each node to use multiple workers to load the data.

This also supports shuffling. At each epoch, the iterable dataset shards are reshuffled across all the nodes - you just have to call iterable_ds.set_epoch(epoch_number).

TODO:

• docs for usage in PyTorch
• unit tests
• integration tests with torch.distributed.launch

Related to huggingface/transformers#20770
Close #5360

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[lhoestq]

Alright all the tests are passing - this is ready for review

[mariosasko]

One nit.

[github-actions]

Show benchmarks

PyArrow==6.0.0

Show updated benchmarks!

Benchmark: benchmark_array_xd.json

metric	read_batch_formatted_as_numpy after write_array2d	read_batch_formatted_as_numpy after write_flattened_sequence	read_batch_formatted_as_numpy after write_nested_sequence	read_batch_unformated after write_array2d	read_batch_unformated after write_flattened_sequence	read_batch_unformated after write_nested_sequence	read_col_formatted_as_numpy after write_array2d	read_col_formatted_as_numpy after write_flattened_sequence	read_col_formatted_as_numpy after write_nested_sequence	read_col_unformated after write_array2d	read_col_unformated after write_flattened_sequence	read_col_unformated after write_nested_sequence	read_formatted_as_numpy after write_array2d	read_formatted_as_numpy after write_flattened_sequence	read_formatted_as_numpy after write_nested_sequence	read_unformated after write_array2d	read_unformated after write_flattened_sequence	read_unformated after write_nested_sequence	write_array2d	write_flattened_sequence	write_nested_sequence
new / old (diff)	0.015146 / 0.011353 (0.003793)	0.006683 / 0.011008 (-0.004326)	0.125994 / 0.038508 (0.087486)	0.041345 / 0.023109 (0.018235)	0.378609 / 0.275898 (0.102711)	0.483139 / 0.323480 (0.159659)	0.009669 / 0.007986 (0.001684)	0.005143 / 0.004328 (0.000814)	0.092015 / 0.004250 (0.087765)	0.052728 / 0.037052 (0.015676)	0.397166 / 0.258489 (0.138677)	0.465820 / 0.293841 (0.171979)	0.051025 / 0.128546 (-0.077521)	0.018451 / 0.075646 (-0.057196)	0.397311 / 0.419271 (-0.021960)	0.054842 / 0.043533 (0.011309)	0.391203 / 0.255139 (0.136064)	0.412743 / 0.283200 (0.129543)	0.111356 / 0.141683 (-0.030327)	1.697526 / 1.452155 (0.245372)	1.795017 / 1.492716 (0.302301)

Benchmark: benchmark_getitem_100B.json

metric	get_batch_of_1024_random_rows	get_batch_of_1024_rows	get_first_row	get_last_row
new / old (diff)	0.253737 / 0.018006 (0.235731)	0.583071 / 0.000490 (0.582581)	0.005958 / 0.000200 (0.005758)	0.000110 / 0.000054 (0.000056)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.030397 / 0.037411 (-0.007014)	0.112242 / 0.014526 (0.097716)	0.138807 / 0.176557 (-0.037749)	0.209820 / 0.737135 (-0.527316)	0.139530 / 0.296338 (-0.156808)

Benchmark: benchmark_iterating.json

metric	read 5000	read 50000	read_batch 50000 10	read_batch 50000 100	read_batch 50000 1000	read_formatted numpy 5000	read_formatted pandas 5000	read_formatted tensorflow 5000	read_formatted torch 5000	read_formatted_batch numpy 5000 10	read_formatted_batch numpy 5000 1000	shuffled read 5000	shuffled read 50000	shuffled read_batch 50000 10	shuffled read_batch 50000 100	shuffled read_batch 50000 1000	shuffled read_formatted numpy 5000	shuffled read_formatted_batch numpy 5000 10	shuffled read_formatted_batch numpy 5000 1000
new / old (diff)	0.574111 / 0.215209 (0.358902)	5.623713 / 2.077655 (3.546058)	2.416880 / 1.504120 (0.912760)	1.951013 / 1.541195 (0.409819)	2.124565 / 1.468490 (0.656075)	1.268854 / 4.584777 (-3.315923)	5.942368 / 3.745712 (2.196656)	5.413814 / 5.269862 (0.143952)	2.931638 / 4.565676 (-1.634038)	0.135070 / 0.424275 (-0.289205)	0.014290 / 0.007607 (0.006683)	0.708384 / 0.226044 (0.482340)	7.487994 / 2.268929 (5.219065)	3.074210 / 55.444624 (-52.370414)	2.380583 / 6.876477 (-4.495893)	2.522298 / 2.142072 (0.380226)	1.336741 / 4.805227 (-3.468486)	0.236761 / 6.500664 (-6.263903)	0.076592 / 0.075469 (0.001123)

Benchmark: benchmark_map_filter.json

metric	filter	map fast-tokenizer batched	map identity	map identity batched	map no-op batched	map no-op batched numpy	map no-op batched pandas	map no-op batched pytorch	map no-op batched tensorflow
new / old (diff)	1.629415 / 1.841788 (-0.212373)	19.000640 / 8.074308 (10.926332)	21.474058 / 10.191392 (11.282666)	0.231227 / 0.680424 (-0.449197)	0.046213 / 0.534201 (-0.487988)	0.565703 / 0.579283 (-0.013580)	0.662956 / 0.434364 (0.228592)	0.656475 / 0.540337 (0.116137)	0.762534 / 1.386936 (-0.624402)

PyArrow==latest

Show updated benchmarks!

Benchmark: benchmark_array_xd.json

metric	read_batch_formatted_as_numpy after write_array2d	read_batch_formatted_as_numpy after write_flattened_sequence	read_batch_formatted_as_numpy after write_nested_sequence	read_batch_unformated after write_array2d	read_batch_unformated after write_flattened_sequence	read_batch_unformated after write_nested_sequence	read_col_formatted_as_numpy after write_array2d	read_col_formatted_as_numpy after write_flattened_sequence	read_col_formatted_as_numpy after write_nested_sequence	read_col_unformated after write_array2d	read_col_unformated after write_flattened_sequence	read_col_unformated after write_nested_sequence	read_formatted_as_numpy after write_array2d	read_formatted_as_numpy after write_flattened_sequence	read_formatted_as_numpy after write_nested_sequence	read_unformated after write_array2d	read_unformated after write_flattened_sequence	read_unformated after write_nested_sequence	write_array2d	write_flattened_sequence	write_nested_sequence
new / old (diff)	0.010952 / 0.011353 (-0.000400)	0.006259 / 0.011008 (-0.004749)	0.132430 / 0.038508 (0.093922)	0.037920 / 0.023109 (0.014811)	0.483565 / 0.275898 (0.207667)	0.528190 / 0.323480 (0.204710)	0.008116 / 0.007986 (0.000130)	0.006768 / 0.004328 (0.002440)	0.100520 / 0.004250 (0.096270)	0.055208 / 0.037052 (0.018155)	0.484672 / 0.258489 (0.226183)	0.556937 / 0.293841 (0.263096)	0.057938 / 0.128546 (-0.070609)	0.020821 / 0.075646 (-0.054826)	0.430735 / 0.419271 (0.011464)	0.066317 / 0.043533 (0.022785)	0.496652 / 0.255139 (0.241513)	0.502004 / 0.283200 (0.218804)	0.125403 / 0.141683 (-0.016280)	1.833396 / 1.452155 (0.381241)	1.974517 / 1.492716 (0.481800)

Benchmark: benchmark_getitem_100B.json

metric	get_batch_of_1024_random_rows	get_batch_of_1024_rows	get_first_row	get_last_row
new / old (diff)	0.269198 / 0.018006 (0.251191)	0.620314 / 0.000490 (0.619824)	0.000535 / 0.000200 (0.000335)	0.000083 / 0.000054 (0.000029)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.032373 / 0.037411 (-0.005039)	0.130043 / 0.014526 (0.115517)	0.146217 / 0.176557 (-0.030339)	0.200187 / 0.737135 (-0.536948)	0.152839 / 0.296338 (-0.143499)

Benchmark: benchmark_iterating.json

metric	read 5000	read 50000	read_batch 50000 10	read_batch 50000 100	read_batch 50000 1000	read_formatted numpy 5000	read_formatted pandas 5000	read_formatted tensorflow 5000	read_formatted torch 5000	read_formatted_batch numpy 5000 10	read_formatted_batch numpy 5000 1000	shuffled read 5000	shuffled read 50000	shuffled read_batch 50000 10	shuffled read_batch 50000 100	shuffled read_batch 50000 1000	shuffled read_formatted numpy 5000	shuffled read_formatted_batch numpy 5000 10	shuffled read_formatted_batch numpy 5000 1000
new / old (diff)	0.677478 / 0.215209 (0.462268)	6.678856 / 2.077655 (4.601201)	3.025870 / 1.504120 (1.521750)	2.678196 / 1.541195 (1.137001)	2.740640 / 1.468490 (1.272150)	1.237163 / 4.584777 (-3.347614)	5.752621 / 3.745712 (2.006908)	3.170435 / 5.269862 (-2.099427)	2.049174 / 4.565676 (-2.516502)	0.147663 / 0.424275 (-0.276612)	0.016107 / 0.007607 (0.008500)	0.849666 / 0.226044 (0.623621)	8.395212 / 2.268929 (6.126283)	3.741120 / 55.444624 (-51.703505)	3.102926 / 6.876477 (-3.773550)	3.233655 / 2.142072 (1.091583)	1.520349 / 4.805227 (-3.284878)	0.267159 / 6.500664 (-6.233505)	0.083646 / 0.075469 (0.008177)

Benchmark: benchmark_map_filter.json

metric	filter	map fast-tokenizer batched	map identity	map identity batched	map no-op batched	map no-op batched numpy	map no-op batched pandas	map no-op batched pytorch	map no-op batched tensorflow
new / old (diff)	1.640458 / 1.841788 (-0.201330)	19.043169 / 8.074308 (10.968861)	22.786126 / 10.191392 (12.594734)	0.218040 / 0.680424 (-0.462384)	0.032948 / 0.534201 (-0.501253)	0.569574 / 0.579283 (-0.009710)	0.658746 / 0.434364 (0.224382)	0.650501 / 0.540337 (0.110164)	0.730588 / 1.386936 (-0.656348)

[lhoestq]

just added a note :)

[mariosasko]

Thanks, looks all good now!

[rishabhm12]

Hi @lhoestq ,
Can you please throw some light on the following statement
If the dataset has a number of shards that is a factor of world_size (i.e. if dataset.n_shards % world_size == 0), then the shards are evenly assigned across the nodes, which is the most optimized. Otherwise, each node keeps 1 example out of world_size, skipping the other examples.

Let's assume I have 127 parquet files and world_size is 4. I was not able to fully comprehend the above statement
What does this statement mean?
each node keeps 1 example out of world_size, skipping the other examples.
Thank you!

[lhoestq]

If you have 128 parquet files, then dataset.n_shards % world_size == 0. In this case each worker can take care of 32 parquet files.

On the other hand if you have dataset.n_shards % world_size != 0 (in your case 127 files), then we can't assign the same number of files to each worker. This is an issue because it may under-utilize your GPU at the end of your training since some workers will take longer to iterate on the dataset than others.

Therefore in this case, all the workers take care of the 127 parquet files but workers will skip examples to not end up with duplicates. That's what "each node keeps 1 example out of world_size, skipping the other examples" means, and in your case it implies:

• rank=0 will read the samples with idx=0, 4, 8 etc.
• rank=1 will read the samples with idx=1, 5, 9 etc.
• rank=2 will read the samples with idx=2, 6, 10 etc.
• rank=3 will read the samples with idx=3, 7, 11 etc.

[rishabhm12]

Thanks a lot @lhoestq , this helps!

[KatarinaYuan]

Hi, in the case above, if we use keep_in_memory=True for Dataset, then we still need to read in n times the dataset if we use DDP on n GPUs (1 node), right? That means we need n times the memory. Is there any way to only load the data once, to save memory?

[lhoestq]

Dataset objects are memory mapped from disk so they use almost no RAM (only the current batch)

Also they are perfectly sharded using split_dataset_by_node so it's going to be read exactly once in total using DDP.
You can also achieve the same thing using a DistributedSampler in pytorch for DDP instead of using split_dataset_by_node.

[KatarinaYuan]

Hi, please correct if I mistake anything:

1. Dataset with keep_in_memory=True would explicitly pre-load the data into memory, instead of reading from disk via the memory map for every batch. The former way should be faster than the latter.
2. When using DDP, before sending the Dataset object into split_dataset_by_node or incorporate it with DistributedSampler, every process still needs to pre-load the entire data into memory (when keep_in_memory=True) and then select the chunked indices from the loaded data.

Generally, the dilemma I'm facing is:
Suppose we have a data around 120GB, and we want to use DistributedLengthGroupedSampler to optimize batching. When using DDP and keep_in_memory=True, every process loads 120GB which is not acceptable. For now, I turned off keep_in_memory and try to increase the number of workers for DataLoader to get better pipelining.

But is it possible to load 120GB once into 4 * A100 (which has around 4*120GB memory) and make each process read from this shared data from memory? Theoretically, maybe it should be faster?

[lhoestq]

Feel free to ask your questions on the forum if you don't mind, this way the discussions may be useful to other people ;)