Alternative to https://github.com/pytorch/pytorch/pull/107034, implements @ezyang 's suggestion from https://github.com/pytorch/pytorch/pull/107034#discussion_r1292857201.
This PR addresses https://fb.workplace.com/groups/pytorch.oss.dev/posts/1699944830430051 and does a bunch of stacked changes:
- Make `Generator` class support GC;this makes all `Generator` instances tracked and accessile through Python's GC.
- Use the GC to retrieve all existing Generator instances in Dataloader's `_worker_loop` and re-seed them: this extends what is already applied to the global/default Generator, which is already re-seeded.
~TODO: a bit of docs and justification, which I'll do if this PR is mergeable.~ -- Done
CC @albanD @ezyang as previously discussed
BC-Breaking Note
-------------------
We now re-seed all `Generator` instances within the `Dataloader` workers' loop to ensure that their RNG is different across workers.
Previously, the RNG of user-defined `Generators` would be the same across workers, which could lead to wrong training procedures. This only affects user-defined `Generators`, not the default `Generator` (which was already re-seeded).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107131
Approved by: https://github.com/ezyang
Torch wrapping datasets list has:
`TensorDataset`
`ConcatDataset`
`ChainDataset`
`TensorDataset` is useful for stacking sets of tensors but can't work with objects without `.size()` method.
This PR proposes `StackDataset`, similar to `TensorDataset` but for a general case like `ConcatDataset`.
Possible usage of `StackDataset` is multimodal networks with different input like image+text or for staking non-tensor input and property to predict.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/101338
Approved by: https://github.com/ejguan, https://github.com/NivekT
Common advice we give for handling memory fragmentation issues is to
allocate a big block upfront to reserve memory which will get split up later.
For programs with changing tensor sizes this can be especially helpful to
avoid OOMs that happen the first time we see a new largest input and would
otherwise have to allocate new segments.
However the issue with allocating a block upfront is that is nearly impossible
to correctly estimate the size of that block. If too small, space in the block
will run out and the allocator will allocate separate blocks anyway. Too large,
and other non-PyTorch libraries might stop working because they cannot allocate
any memory.
This patch provides the same benefits as using a pre-allocating block but
without having to choose its size upfront. Using the cuMemMap-style APIs,
it adds the ability to expand the last block in a segment when more memory is
needed.
Compared to universally using cudaMallocAsync to avoid fragmentation,
this patch can fix this common fragmentation issue while preserving most
of the existing allocator behavior. This behavior can be enabled and disabled dynamically.
This should allow users to, for instance, allocate long-lived parameters and state in individual buffers,
and put temporary state into the large expandable blocks, further reducing
fragmentation.
See inline comments for information about the implementation and its limitations.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96995
Approved by: https://github.com/eellison
Applies the remaining flake8-comprehension fixes and checks. This changes replace all remaining unnecessary generator expressions with list/dict/set comprehensions which are more succinct, performant, and better supported by our torch.jit compiler. It also removes useless generators such as 'set(a for a in b)`, resolving it into just the set call.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94676
Approved by: https://github.com/ezyang
Rerun all disabled test to gather their latest result so that we can close disabled tickets automatically. When running under this mode (RERUN_DISABLED_TESTS=true), only disabled tests are run while the rest are skipped `<skipped message="Test is enabled but --rerun-disabled-tests verification mode is set, so only disabled tests are run" type="skip"/>`
The logic is roughly as follows, the test runs multiple times (n=50)
* If the disabled test passes, and it's flaky, do nothing because it's still flaky. In the test report, we'll see the test passes with the following skipped message:
```
<testcase classname="TestMultiprocessing" file="test_multiprocessing.py" line="357" name="test_fs" time="0.000" timestamp="0001-01-01T00:00:00">
<skipped message="{"flaky": True, "num_red": 4, "num_green": 0, "max_num_retries": 3, "rerun_disabled_test": true}" type="skip"/>
</testcase>
```
* If the disabled test passes every single time, and it is not flaky anymore, mark it so that it can be closed later. We will see the test runs and passes, i.e.
```
<testcase classname="TestCommonCUDA" name="test_out_warning_linalg_lu_factor_cuda" time="0.170" file="test_ops.py" />
```
* If the disabled test fails after all retries, this is also expected. So only report this but don't fail the job (because we don't care about red signals here), we'll see the test is skipped (without the `flaky` field), i.e.
```
<testcase classname="TestMultiprocessing" file="test_multiprocessing.py" line="357" name="test_fs" time="0.000" timestamp="0001-01-01T00:00:00">
<skipped message="{"num_red": 4, "num_green": 0, "max_num_retries": 3, "rerun_disabled_test": true}" type="skip"/>
</testcase>
```
This runs at the same schedule as `mem_leak_check` (daily). The change to update test stats, and (potentially) grouping on HUD will come in separated PRs.
### Testing
* pull https://github.com/pytorch/pytorch/actions/runs/3447434434
* trunk https://github.com/pytorch/pytorch/actions/runs/3447434928
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88646
Approved by: https://github.com/clee2000
This PR changes the behavior of `IterDataPipe` to always invoke `reset` for the state of `NotStarted`. The main reason is we normally put lazy initialization code into `reset` function. Even for the state of `NotStarted`, we should invoke `reset` to initialize those lazy variables. Otherwise, we have to manually determine if the state is `NotStarted` or `Iterating` in `__iter__` function and only manually invoke `reset` in the state of `NotStarted`.
This PR also makes `Shuffler` is able to serialize with `buffer` and `rng_state`.
The following part is removed:
~I am also add `_snapshot_state` into serialization state and during `__setstate__` only change the state to `Restored` if the original state is `Iterating`. Especially, for the case of deserializing/serializing `NotStarted` DataPipe (multiprocessing), we would invoke `set_seed` for `Shuffler`. We need the `DataPipe` remains as `NotStarted` to properly `reset`.~
I am listing all the expected behavior state transition below:
- Initial state: `NotStarted`
- `iter` -> Call `reset` and change the state to `Iterating`
- serialize/deserialize -> Keep the state as `NotStarted` (will `reset` if `iter` is called afterwards)
- Initial state: `Iterating`
- `iter` -> Call `reset` and keep the state to `Iterating`
- serialize/deserialize -> Change the state as `Restored`
- Initial state: `Restored`
- `iter` -> Only change the state to `Iterating`
- serialize/deserialize -> Not allowed
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83535
Approved by: https://github.com/NivekT
Fixes#70517#70547#74598
I basically disable the `loader_kill` test that would prevent proper clean up, which would be handled by Python interpreter in normal python script execution. However, in pytest framework, the python interpreter never exits and execute clean up on the shared memory using `/multiprocessing/resource_tracker.py`. So, we might encounter the running out of shared memory in some cases.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81660
Approved by: https://github.com/NivekT
Fixes#78510
This PR adds support for using fractions with `random_split`. This should be completely backwards-compatible as the fractional-style splitting is only applied when the sum across the input lengths is lower than 1.0
Pull Request resolved: https://github.com/pytorch/pytorch/pull/78877
Approved by: https://github.com/ejguan
Summary:
pin_memory, has optional device parameter to specify
which device you want to pin for. With this above change
the Dataloader will work only for CUDA backend. To add
support for other backend which supports pinned memory,
dataloader is updated with device as optional parameter.
Fixes #{issue number}
Pull Request resolved: https://github.com/pytorch/pytorch/pull/65402
Reviewed By: zou3519
Differential Revision: D32282204
Pulled By: VitalyFedyunin
fbshipit-source-id: e2e09876969af108d0db38af7c2d1b2f1cfa9858
(cherry picked from commit 3b76e151964fce442e27fe8fb5c37af930da4fa1)
Without this, `DataLoader2` will just add an `Shuffler` to the end of the datapipe if `shuffle=True`:
```py
from torch.utils.data.dataloader_experimental import DataLoader2
from torchdata.datapipes.iter import IterableWrapper, IterDataPipe, Shuffler
class Sorter(IterDataPipe):
def __init__(self, datapipe):
self.datapipe = datapipe
def __iter__(self):
return iter(sorted(self.datapipe))
data = list(range(1000))
dp = IterableWrapper(data)
dp = Shuffler(dp).set_shuffle(False)
dp = Sorter(dp)
dl2 = DataLoader2(dp, shuffle=True, batch_size=None)
assert list(dl2) == data # fails unless you hit a lucky random seed
```
This example is somewhat non-sensical, but demonstrates we cannot simply add a `Shuffler`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75014
Approved by: https://github.com/ejguan
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/72896
Fixing the issue described here: https://github.com/pytorch/data/issues/214
There will be a follow-up PR in TorchData as well
Test Plan: Imported from OSS
Reviewed By: gchanan
Differential Revision: D34258669
Pulled By: NivekT
fbshipit-source-id: 6dd88250ed14ebe779915dc46139be7e012e9d1b
(cherry picked from commit 025b8ed98019e576bfef04c33a3f33ed1a426a66)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/71579Fixes#1551
As the comment in the code, register a function to terminate persistent workers.
By adding a reference of these workers in `atexit`, it would prevent Python interpreter kills these persistent worker processes before `pin_memorh_thread` exits.
And, if users explicitly kills DataLoader iterator, such function in `atexit` would be a no-op.
Test Plan: Imported from OSS
Reviewed By: VitalyFedyunin
Differential Revision: D33896537
Pulled By: ejguan
fbshipit-source-id: 36b57eac7523d8aa180180c2b61fc693ea4638ae
(cherry picked from commit 05add2ae0f)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/71579Fixes#1551
As the comment in the code, register a function to terminate persistent workers. Using `atexit` to make sure termination of persistent workers always happens at the end (after pin_memory_thread exits).
We need such mechanism because Python interpreter would clean up worker process before DataLoader iterator in some rare cases.
Test Plan: Imported from OSS
Reviewed By: VitalyFedyunin
Differential Revision: D33694867
Pulled By: ejguan
fbshipit-source-id: 0847f4d424a0cd6b3c0be8235d505415970254e8
(cherry picked from commit 18ad4621af)
Summary:
`default_collate`, `default_convert`, and `pin_memory` convert sequences into lists. I believe they should keep the original type when possible (e.g., I have a class that inherits from `list`, which comes from a 3rd party library that I can't change, and provides extra functionality).
Note it's easy to do when the type supports an iterable in its creation but it's not always the case (e.g., `range`).
Even though this can be accomplished if using a custom `default_collate`/`default_convert`, 1) this is behavior they should support out-of-the-box IMHO, and 2) `pin_memory` still does it.
cc VitalyFedyunin ejguan NivekT
Pull Request resolved: https://github.com/pytorch/pytorch/pull/68779
Reviewed By: wenleix
Differential Revision: D32651129
Pulled By: ejguan
fbshipit-source-id: 17c390934bacc0e4ead060469cf15dde815550b4
