Persist IterableDataset epoch in workers

[lhoestq]

Use shared memory for the IterableDataset epoch.

This way calling ds.set_epoch() in the main process will update the epoch in the DataLoader workers as well.
This is useful especially because the epoch is used to compute the effective_seed used for shuffling.

I used torch's shared memory in case users want to send dataset copies without shared memory using pickle. I also find it easier to use than using multiprocessing.shared_memory than requires unlinking only in the main process, or mp.Value that is not picklable.

close #6673

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.

[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)	0.005283 / 0.011353 (-0.006070)	0.003866 / 0.011008 (-0.007142)	0.063124 / 0.038508 (0.024616)	0.030240 / 0.023109 (0.007131)	0.232855 / 0.275898 (-0.043043)	0.257538 / 0.323480 (-0.065942)	0.004165 / 0.007986 (-0.003820)	0.002826 / 0.004328 (-0.001502)	0.049735 / 0.004250 (0.045485)	0.045297 / 0.037052 (0.008244)	0.251831 / 0.258489 (-0.006658)	0.277812 / 0.293841 (-0.016029)	0.030004 / 0.128546 (-0.098542)	0.012319 / 0.075646 (-0.063328)	0.206881 / 0.419271 (-0.212391)	0.036561 / 0.043533 (-0.006972)	0.234364 / 0.255139 (-0.020775)	0.258316 / 0.283200 (-0.024884)	0.017815 / 0.141683 (-0.123867)	1.114111 / 1.452155 (-0.338043)	1.165428 / 1.492716 (-0.327288)

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)	0.099302 / 0.018006 (0.081296)	0.309195 / 0.000490 (0.308705)	0.000261 / 0.000200 (0.000061)	0.000044 / 0.000054 (-0.000010)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.018765 / 0.037411 (-0.018646)	0.063123 / 0.014526 (0.048597)	0.075437 / 0.176557 (-0.101119)	0.122570 / 0.737135 (-0.614566)	0.076637 / 0.296338 (-0.219702)

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)	0.289965 / 0.215209 (0.074756)	2.839053 / 2.077655 (0.761398)	1.503463 / 1.504120 (-0.000657)	1.390833 / 1.541195 (-0.150361)	1.401918 / 1.468490 (-0.066572)	0.711000 / 4.584777 (-3.873777)	2.325513 / 3.745712 (-1.420199)	2.831630 / 5.269862 (-2.438231)	1.908370 / 4.565676 (-2.657307)	0.077867 / 0.424275 (-0.346408)	0.005509 / 0.007607 (-0.002098)	0.336494 / 0.226044 (0.110450)	3.358587 / 2.268929 (1.089658)	1.901067 / 55.444624 (-53.543558)	1.590130 / 6.876477 (-5.286347)	1.753850 / 2.142072 (-0.388223)	0.792458 / 4.805227 (-4.012769)	0.135584 / 6.500664 (-6.365080)	0.042028 / 0.075469 (-0.033441)

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)	0.966162 / 1.841788 (-0.875625)	11.705310 / 8.074308 (3.631002)	9.158842 / 10.191392 (-1.032550)	0.128793 / 0.680424 (-0.551631)	0.014422 / 0.534201 (-0.519779)	0.299009 / 0.579283 (-0.280274)	0.262873 / 0.434364 (-0.171491)	0.340836 / 0.540337 (-0.199501)	0.464440 / 1.386936 (-0.922496)

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)	0.005951 / 0.011353 (-0.005402)	0.003984 / 0.011008 (-0.007024)	0.051432 / 0.038508 (0.012924)	0.033223 / 0.023109 (0.010113)	0.263972 / 0.275898 (-0.011926)	0.289060 / 0.323480 (-0.034420)	0.004446 / 0.007986 (-0.003540)	0.002891 / 0.004328 (-0.001438)	0.049347 / 0.004250 (0.045096)	0.041191 / 0.037052 (0.004138)	0.278334 / 0.258489 (0.019844)	0.314065 / 0.293841 (0.020224)	0.032020 / 0.128546 (-0.096526)	0.012472 / 0.075646 (-0.063174)	0.061288 / 0.419271 (-0.357984)	0.033489 / 0.043533 (-0.010044)	0.266831 / 0.255139 (0.011692)	0.283008 / 0.283200 (-0.000192)	0.018491 / 0.141683 (-0.123192)	1.133634 / 1.452155 (-0.318521)	1.154627 / 1.492716 (-0.338089)

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)	0.101831 / 0.018006 (0.083825)	0.317942 / 0.000490 (0.317452)	0.000217 / 0.000200 (0.000018)	0.000056 / 0.000054 (0.000002)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.022608 / 0.037411 (-0.014803)	0.076776 / 0.014526 (0.062250)	0.088686 / 0.176557 (-0.087870)	0.129092 / 0.737135 (-0.608044)	0.090780 / 0.296338 (-0.205558)

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)	0.286762 / 0.215209 (0.071553)	2.824307 / 2.077655 (0.746652)	1.547215 / 1.504120 (0.043095)	1.424522 / 1.541195 (-0.116673)	1.446414 / 1.468490 (-0.022076)	0.723683 / 4.584777 (-3.861094)	0.974129 / 3.745712 (-2.771583)	2.952552 / 5.269862 (-2.317309)	1.903663 / 4.565676 (-2.662013)	0.078786 / 0.424275 (-0.345489)	0.005130 / 0.007607 (-0.002477)	0.338925 / 0.226044 (0.112881)	3.378557 / 2.268929 (1.109629)	1.892951 / 55.444624 (-53.551674)	1.599844 / 6.876477 (-5.276633)	1.611963 / 2.142072 (-0.530109)	0.793614 / 4.805227 (-4.011613)	0.133795 / 6.500664 (-6.366869)	0.040777 / 0.075469 (-0.034692)

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)	1.001391 / 1.841788 (-0.840397)	12.166811 / 8.074308 (4.092503)	10.588180 / 10.191392 (0.396788)	0.141609 / 0.680424 (-0.538815)	0.020941 / 0.534201 (-0.513260)	0.340149 / 0.579283 (-0.239134)	0.122988 / 0.434364 (-0.311376)	0.339747 / 0.540337 (-0.200591)	0.434338 / 1.386936 (-0.952598)