Increase chunk size for speeding up file downloads

[Narsil]

Original fix: huggingface/huggingface_hub#1267
Not sure this function is actually still called though.

I haven't done benches on this. Is there a dataset where files are hosted on the hub through cloudfront so we can have the same setup as in hf_hub ?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[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)	0.008407 / 0.011353 (-0.002946)	0.004651 / 0.011008 (-0.006357)	0.100367 / 0.038508 (0.061859)	0.029107 / 0.023109 (0.005998)	0.302798 / 0.275898 (0.026900)	0.354379 / 0.323480 (0.030899)	0.006985 / 0.007986 (-0.001001)	0.003365 / 0.004328 (-0.000963)	0.078312 / 0.004250 (0.074062)	0.034205 / 0.037052 (-0.002847)	0.310431 / 0.258489 (0.051941)	0.346239 / 0.293841 (0.052398)	0.033800 / 0.128546 (-0.094747)	0.011515 / 0.075646 (-0.064131)	0.323588 / 0.419271 (-0.095684)	0.040766 / 0.043533 (-0.002767)	0.300914 / 0.255139 (0.045775)	0.332983 / 0.283200 (0.049784)	0.087500 / 0.141683 (-0.054182)	1.469505 / 1.452155 (0.017350)	1.505119 / 1.492716 (0.012403)

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)	0.187319 / 0.018006 (0.169313)	0.405498 / 0.000490 (0.405008)	0.001000 / 0.000200 (0.000800)	0.000069 / 0.000054 (0.000015)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.022583 / 0.037411 (-0.014828)	0.098096 / 0.014526 (0.083570)	0.104272 / 0.176557 (-0.072284)	0.142801 / 0.737135 (-0.594335)	0.109749 / 0.296338 (-0.186590)

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)	0.423343 / 0.215209 (0.208134)	4.215116 / 2.077655 (2.137461)	1.899714 / 1.504120 (0.395594)	1.689579 / 1.541195 (0.148384)	1.710292 / 1.468490 (0.241801)	0.690976 / 4.584777 (-3.893801)	3.432501 / 3.745712 (-0.313212)	1.899600 / 5.269862 (-3.370261)	1.279801 / 4.565676 (-3.285876)	0.082763 / 0.424275 (-0.341512)	0.012545 / 0.007607 (0.004938)	0.531381 / 0.226044 (0.305336)	5.320077 / 2.268929 (3.051148)	2.370705 / 55.444624 (-53.073919)	2.007089 / 6.876477 (-4.869388)	2.062412 / 2.142072 (-0.079661)	0.814998 / 4.805227 (-3.990229)	0.149822 / 6.500664 (-6.350842)	0.064399 / 0.075469 (-0.011070)

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)	1.226196 / 1.841788 (-0.615591)	13.823443 / 8.074308 (5.749134)	13.813667 / 10.191392 (3.622275)	0.161289 / 0.680424 (-0.519135)	0.028569 / 0.534201 (-0.505632)	0.390360 / 0.579283 (-0.188923)	0.396217 / 0.434364 (-0.038147)	0.483120 / 0.540337 (-0.057217)	0.570041 / 1.386936 (-0.816895)

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)	0.006422 / 0.011353 (-0.004931)	0.004528 / 0.011008 (-0.006481)	0.076043 / 0.038508 (0.037535)	0.027631 / 0.023109 (0.004522)	0.340622 / 0.275898 (0.064724)	0.376694 / 0.323480 (0.053214)	0.004993 / 0.007986 (-0.002992)	0.003403 / 0.004328 (-0.000926)	0.074521 / 0.004250 (0.070270)	0.037568 / 0.037052 (0.000516)	0.343423 / 0.258489 (0.084934)	0.387729 / 0.293841 (0.093888)	0.031790 / 0.128546 (-0.096757)	0.011767 / 0.075646 (-0.063879)	0.085182 / 0.419271 (-0.334090)	0.042867 / 0.043533 (-0.000666)	0.341269 / 0.255139 (0.086130)	0.368460 / 0.283200 (0.085261)	0.090153 / 0.141683 (-0.051530)	1.536490 / 1.452155 (0.084335)	1.596403 / 1.492716 (0.103686)

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)	0.222373 / 0.018006 (0.204367)	0.396145 / 0.000490 (0.395655)	0.000384 / 0.000200 (0.000184)	0.000062 / 0.000054 (0.000008)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.024801 / 0.037411 (-0.012610)	0.099711 / 0.014526 (0.085185)	0.106094 / 0.176557 (-0.070463)	0.147819 / 0.737135 (-0.589316)	0.110065 / 0.296338 (-0.186274)

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)	0.442863 / 0.215209 (0.227654)	4.420043 / 2.077655 (2.342388)	2.070136 / 1.504120 (0.566016)	1.862363 / 1.541195 (0.321168)	1.910890 / 1.468490 (0.442400)	0.702570 / 4.584777 (-3.882207)	3.435855 / 3.745712 (-0.309857)	1.871290 / 5.269862 (-3.398572)	1.169321 / 4.565676 (-3.396355)	0.083674 / 0.424275 (-0.340601)	0.012823 / 0.007607 (0.005216)	0.539330 / 0.226044 (0.313285)	5.403317 / 2.268929 (3.134389)	2.536508 / 55.444624 (-52.908117)	2.179629 / 6.876477 (-4.696847)	2.207586 / 2.142072 (0.065514)	0.812256 / 4.805227 (-3.992972)	0.152915 / 6.500664 (-6.347749)	0.068431 / 0.075469 (-0.007038)

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)	1.294982 / 1.841788 (-0.546806)	13.912811 / 8.074308 (5.838503)	13.415658 / 10.191392 (3.224266)	0.149531 / 0.680424 (-0.530893)	0.016785 / 0.534201 (-0.517416)	0.381055 / 0.579283 (-0.198228)	0.392084 / 0.434364 (-0.042280)	0.472614 / 0.540337 (-0.067724)	0.559799 / 1.386936 (-0.827137)

[lhoestq]

We simply do GET requests to hf.co to download files from the Hub right now. We may switch to hfh when we update how we do caching

You can try on any dataset hosted on the hub like imagenet-1k

[albertvillanova]

Thanks for the improvement, @Narsil.

Just a question below.

[albertvillanova]

To be aligned with huggingface_hub, shouldn't it be 10 MB instead?

Suggested change

	for chunk in response.iter_content(chunk_size=1024 * 1024):
	for chunk in response.iter_content(chunk_size=10 * 1024 * 1024):

[Narsil]

We could. In my experiements it made only tiny improvements over 1Mb but you're right let's keep consistent!

I would like to run benchmarks on this too, before claiming victory, however I'm short on time this week :(

[Narsil]

To do benchmarks, use a machine that has good network (> 50Mo) otherwise you don't see anything.

On DGX or large AWS instances, I was able to get ~500Mo/s of transfer speed where it was stuck at ~50Mo/s before.

For larger networks, it's impossible to do in Python afaik (but there are still ways to achieve it with hf_transfer but this is very much non official, since it's rare to get such bandwidth)

[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)	0.010931 / 0.011353 (-0.000422)	0.005730 / 0.011008 (-0.005278)	0.116653 / 0.038508 (0.078145)	0.041439 / 0.023109 (0.018330)	0.359559 / 0.275898 (0.083661)	0.408398 / 0.323480 (0.084918)	0.009193 / 0.007986 (0.001208)	0.006024 / 0.004328 (0.001695)	0.087743 / 0.004250 (0.083492)	0.048636 / 0.037052 (0.011584)	0.363133 / 0.258489 (0.104643)	0.407144 / 0.293841 (0.113303)	0.044610 / 0.128546 (-0.083936)	0.014075 / 0.075646 (-0.061571)	0.396506 / 0.419271 (-0.022766)	0.057014 / 0.043533 (0.013482)	0.358254 / 0.255139 (0.103115)	0.399887 / 0.283200 (0.116687)	0.115337 / 0.141683 (-0.026346)	1.731655 / 1.452155 (0.279500)	1.813276 / 1.492716 (0.320560)

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)	0.210197 / 0.018006 (0.192191)	0.475887 / 0.000490 (0.475397)	0.003323 / 0.000200 (0.003123)	0.000100 / 0.000054 (0.000045)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.031686 / 0.037411 (-0.005725)	0.131167 / 0.014526 (0.116641)	0.137919 / 0.176557 (-0.038637)	0.184843 / 0.737135 (-0.552293)	0.144998 / 0.296338 (-0.151340)

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)	0.471371 / 0.215209 (0.256162)	4.693739 / 2.077655 (2.616084)	2.251567 / 1.504120 (0.747447)	1.993653 / 1.541195 (0.452458)	2.053236 / 1.468490 (0.584746)	0.809226 / 4.584777 (-3.775551)	4.494120 / 3.745712 (0.748408)	2.436921 / 5.269862 (-2.832940)	1.541973 / 4.565676 (-3.023704)	0.098401 / 0.424275 (-0.325874)	0.014329 / 0.007607 (0.006722)	0.597813 / 0.226044 (0.371769)	5.964035 / 2.268929 (3.695107)	2.709283 / 55.444624 (-52.735341)	2.323537 / 6.876477 (-4.552940)	2.401707 / 2.142072 (0.259635)	0.976379 / 4.805227 (-3.828848)	0.194638 / 6.500664 (-6.306026)	0.076904 / 0.075469 (0.001435)

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)	1.516877 / 1.841788 (-0.324911)	18.228010 / 8.074308 (10.153702)	16.631750 / 10.191392 (6.440358)	0.176030 / 0.680424 (-0.504394)	0.033769 / 0.534201 (-0.500432)	0.520511 / 0.579283 (-0.058773)	0.531764 / 0.434364 (0.097400)	0.648658 / 0.540337 (0.108321)	0.779124 / 1.386936 (-0.607812)

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)	0.008635 / 0.011353 (-0.002718)	0.005785 / 0.011008 (-0.005223)	0.087042 / 0.038508 (0.048534)	0.039632 / 0.023109 (0.016523)	0.419719 / 0.275898 (0.143821)	0.463860 / 0.323480 (0.140380)	0.006621 / 0.007986 (-0.001364)	0.004655 / 0.004328 (0.000327)	0.087003 / 0.004250 (0.082753)	0.057122 / 0.037052 (0.020069)	0.417820 / 0.258489 (0.159331)	0.485981 / 0.293841 (0.192140)	0.042606 / 0.128546 (-0.085940)	0.014369 / 0.075646 (-0.061278)	0.101939 / 0.419271 (-0.317333)	0.058303 / 0.043533 (0.014770)	0.415053 / 0.255139 (0.159914)	0.439914 / 0.283200 (0.156714)	0.134628 / 0.141683 (-0.007055)	1.765464 / 1.452155 (0.313309)	1.843963 / 1.492716 (0.351247)

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)	0.307156 / 0.018006 (0.289150)	0.476657 / 0.000490 (0.476167)	0.019718 / 0.000200 (0.019518)	0.000160 / 0.000054 (0.000105)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.035286 / 0.037411 (-0.002125)	0.138094 / 0.014526 (0.123568)	0.144768 / 0.176557 (-0.031789)	0.191386 / 0.737135 (-0.545750)	0.151988 / 0.296338 (-0.144350)

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)	0.504733 / 0.215209 (0.289523)	5.027048 / 2.077655 (2.949394)	2.441571 / 1.504120 (0.937451)	2.198242 / 1.541195 (0.657047)	2.298473 / 1.468490 (0.829983)	0.848048 / 4.584777 (-3.736729)	4.613102 / 3.745712 (0.867390)	2.522824 / 5.269862 (-2.747037)	1.610159 / 4.565676 (-2.955517)	0.105197 / 0.424275 (-0.319078)	0.015195 / 0.007607 (0.007588)	0.626976 / 0.226044 (0.400932)	6.268459 / 2.268929 (3.999530)	3.014387 / 55.444624 (-52.430237)	2.554102 / 6.876477 (-4.322375)	2.656051 / 2.142072 (0.513979)	1.027978 / 4.805227 (-3.777249)	0.200686 / 6.500664 (-6.299978)	0.077104 / 0.075469 (0.001635)

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)	1.485228 / 1.841788 (-0.356560)	18.319949 / 8.074308 (10.245641)	15.855739 / 10.191392 (5.664347)	0.204365 / 0.680424 (-0.476059)	0.023824 / 0.534201 (-0.510377)	0.505000 / 0.579283 (-0.074283)	0.502866 / 0.434364 (0.068502)	0.629574 / 0.540337 (0.089237)	0.746602 / 1.386936 (-0.640334)