Faster webdataset streaming

[lhoestq]

requests.get(..., streaming=True) is faster than using HTTP range requests when streaming large TAR files

it can be enabled using block_size=0 in fsspec

cc @rwightman

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[lhoestq]

I added faster streaming support using streaming Requests instances in huggingface_hub and will be available in 0.21.

This PR can be used with huggingface/huggingface_hub#1967 to get fast WebDataset streaming

[github-actions]

Show benchmarks

PyArrow==8.0.0

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)	0.004941 / 0.011353 (-0.006412)	0.003431 / 0.011008 (-0.007577)	0.062768 / 0.038508 (0.024260)	0.029212 / 0.023109 (0.006103)	0.253053 / 0.275898 (-0.022845)	0.273061 / 0.323480 (-0.050419)	0.004114 / 0.007986 (-0.003871)	0.002713 / 0.004328 (-0.001616)	0.048481 / 0.004250 (0.044231)	0.040001 / 0.037052 (0.002949)	0.268461 / 0.258489 (0.009971)	0.287767 / 0.293841 (-0.006074)	0.027885 / 0.128546 (-0.100661)	0.010474 / 0.075646 (-0.065172)	0.207989 / 0.419271 (-0.211282)	0.035893 / 0.043533 (-0.007640)	0.256833 / 0.255139 (0.001694)	0.274197 / 0.283200 (-0.009003)	0.017283 / 0.141683 (-0.124400)	1.133597 / 1.452155 (-0.318558)	1.206661 / 1.492716 (-0.286055)

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)	0.089610 / 0.018006 (0.071604)	0.306051 / 0.000490 (0.305562)	0.000217 / 0.000200 (0.000017)	0.000042 / 0.000054 (-0.000012)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.018686 / 0.037411 (-0.018725)	0.061253 / 0.014526 (0.046727)	0.073654 / 0.176557 (-0.102903)	0.120499 / 0.737135 (-0.616637)	0.074827 / 0.296338 (-0.221511)

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)	0.293756 / 0.215209 (0.078547)	2.897755 / 2.077655 (0.820100)	1.558146 / 1.504120 (0.054026)	1.458020 / 1.541195 (-0.083174)	1.453489 / 1.468490 (-0.015001)	0.576666 / 4.584777 (-4.008111)	2.423441 / 3.745712 (-1.322271)	2.727760 / 5.269862 (-2.542102)	1.750287 / 4.565676 (-2.815390)	0.062094 / 0.424275 (-0.362181)	0.004940 / 0.007607 (-0.002667)	0.338815 / 0.226044 (0.112770)	3.342677 / 2.268929 (1.073748)	1.928335 / 55.444624 (-53.516290)	1.629965 / 6.876477 (-5.246511)	1.651836 / 2.142072 (-0.490236)	0.644354 / 4.805227 (-4.160874)	0.117890 / 6.500664 (-6.382774)	0.041907 / 0.075469 (-0.033562)

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)	0.984399 / 1.841788 (-0.857389)	11.516572 / 8.074308 (3.442264)	10.326922 / 10.191392 (0.135530)	0.130821 / 0.680424 (-0.549603)	0.014084 / 0.534201 (-0.520117)	0.287078 / 0.579283 (-0.292205)	0.263466 / 0.434364 (-0.170898)	0.326867 / 0.540337 (-0.213470)	0.425313 / 1.386936 (-0.961623)

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)	0.005305 / 0.011353 (-0.006048)	0.003646 / 0.011008 (-0.007362)	0.049402 / 0.038508 (0.010894)	0.031719 / 0.023109 (0.008610)	0.272579 / 0.275898 (-0.003319)	0.295241 / 0.323480 (-0.028239)	0.004309 / 0.007986 (-0.003677)	0.002781 / 0.004328 (-0.001548)	0.048134 / 0.004250 (0.043883)	0.044702 / 0.037052 (0.007650)	0.288201 / 0.258489 (0.029712)	0.320351 / 0.293841 (0.026510)	0.051327 / 0.128546 (-0.077219)	0.011019 / 0.075646 (-0.064628)	0.057983 / 0.419271 (-0.361288)	0.034211 / 0.043533 (-0.009322)	0.272856 / 0.255139 (0.017717)	0.290007 / 0.283200 (0.006807)	0.018656 / 0.141683 (-0.123027)	1.135017 / 1.452155 (-0.317138)	1.183904 / 1.492716 (-0.308813)

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)	0.090854 / 0.018006 (0.072847)	0.299654 / 0.000490 (0.299165)	0.000224 / 0.000200 (0.000024)	0.000063 / 0.000054 (0.000009)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.021882 / 0.037411 (-0.015529)	0.075297 / 0.014526 (0.060771)	0.086620 / 0.176557 (-0.089937)	0.127125 / 0.737135 (-0.610011)	0.088622 / 0.296338 (-0.207717)

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)	0.287104 / 0.215209 (0.071895)	2.802723 / 2.077655 (0.725068)	1.570137 / 1.504120 (0.066017)	1.452234 / 1.541195 (-0.088961)	1.465457 / 1.468490 (-0.003033)	0.564965 / 4.584777 (-4.019812)	2.416724 / 3.745712 (-1.328988)	2.645057 / 5.269862 (-2.624805)	1.727599 / 4.565676 (-2.838078)	0.063338 / 0.424275 (-0.360937)	0.005018 / 0.007607 (-0.002589)	0.345280 / 0.226044 (0.119235)	3.384323 / 2.268929 (1.115395)	1.957227 / 55.444624 (-53.487397)	1.667620 / 6.876477 (-5.208856)	1.795339 / 2.142072 (-0.346733)	0.642049 / 4.805227 (-4.163178)	0.114853 / 6.500664 (-6.385811)	0.040459 / 0.075469 (-0.035010)

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)	1.023640 / 1.841788 (-0.818147)	11.998130 / 8.074308 (3.923822)	10.858137 / 10.191392 (0.666744)	0.130235 / 0.680424 (-0.550189)	0.016201 / 0.534201 (-0.518000)	0.289743 / 0.579283 (-0.289540)	0.275100 / 0.434364 (-0.159264)	0.329299 / 0.540337 (-0.211039)	0.418632 / 1.386936 (-0.968304)