Summary:
Improved roundup_power2_divisions knob so it allows better control of rouding in the PyTorch CUDA Caching Allocator.
This new version allows setting the number of divisions per power of two interval starting from 1MB and ending at 64GB and above. An example use case is when rouding is desirable for small allocations but there are also very large allocations which are persistent, thus would not benefit from rounding and take up extra space.
Test Plan: Tested locally
Differential Revision: D40103909
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87290
Approved by: https://github.com/zdevito
Fixes#83973 (This is a substitute PR for https://github.com/pytorch/pytorch/pull/85024)
First of all, thanks for your invaluable contributions to PyTorch everyone!
Given how extensively `torch.cuda.is_available` is used in the PyTorch ecosystem, IMHO it's worthwhile to provide downstream libraries/frameworks/users the ability to alter the default behavior of `torch.cuda.is_available` in the context of their PyTorch usage.
I'm confident there are many current and future such use cases which could benefit from leveraging a weakened, NVML-based `torch.cuda.is_available` assessment at a downstream framework's explicit direction (thanks @malfet 81da50a972 !). Though one could always patch out the `torch.cuda.is_available` function with another implementation in a downstream library, I think this environmental variable based configuration option is more convenient and the cost to including the option is quite low.
As discussed in https://github.com/pytorch/pytorch/pull/85024#issuecomment-1261542045, this PR gates new non-default NVML-based CUDA behavior with an environmental variable (PYTORCH_NVML_BASED_CUDA_CHK) that allows a user/framework to invoke non-default, NVML-based `is_available()` assessments if desired.
Thanks again for your work everyone!
@ngimel @malfet @awaelchli
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85951
Approved by: https://github.com/ngimel
Summary:
Added an additional roundup knob( ``roundup_bypass_threshold_mb``) to bypass rounding the requested allocation size, for allocation requests larger than the threshold value (in MB). This can help reduce the memory footprint when making large allocations that are expected to be persistent or have a large lifetime.
Differential Revision: D39868104
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85940
Approved by: https://github.com/zdevito
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/74261
### Goal
Implement a cheap way to reclaim GPU memory (garbage collection) without incurring GPU sync.
### Why do we need this?
Currently, there are only two ways to reclaim GPU memory block already assigned to a particular stream.
- `release_available_cached_blocks(params)`: Free blocks exceeding the `CachingAllocatorConfig::max_split_size()` until we can satisfy the request.
Issue: If the `max_split_size` is unset (default), this function is a no-op. Even if this is set, the reclamation is quite conservative (e.g., never frees blocks under max_split_size).
- `release_cached_blocks()`: Waits for all the in-flight events and then reclaim blocks.
Issue: 'waiting for all event' is very expensive as it will likely stall all the GPU operations. Many GPU applications without a proper handling of potential GPU throttling would suffer/crash.
### Proposed idea
- If the garbage collection threshold is set, try to reclaim some memory blocks *without* synchronization. It should be safe to do so, as `release_available_cached_blocks` essentially does the same thing (but less aggressively).
- GC is triggered only when we fail to serve a `malloc` request from the block pool. No need to free blocks when the block pool is functioning just fine.
- Prioritize reclaiming blocks that weren't reused for long time. Reclamation stops once the used memory capacity < threshold.
- This code path is totally optional; by default it won't be invoked.
Test Plan:
- Unit tests
- Manually checked that the GPU memory usage stays as indicated by the garbage collector. If not the caching allocator at least tries to keep freeing the blocks.
Reviewed By: jianyuh
Differential Revision: D34482514
fbshipit-source-id: d5eae62ac60b94b0bca851f9d233a092d086e3c2
(cherry picked from commit 05780f1ed4b176f05e765b2411c9eaa2eaeb48b0)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/74213
In the current CUDACachingAllocator, the sizes are rounded up in multiple of blocks size of 512, so this works for smaller sizes. However for large sizes, we can have lots of different size blocks in the larger pool. This is problematic when we have variable batch sizes 1001, 1021, 1023 -> all will go to different block size and will create different size of blocks. This will create lots of unused blocks and will waste GPU memory capacity.
This diff adds a rounding approach to allocation size. It rounds up the size to nearest power-of-2 divisions and the power2-division can be changed with env variable setting.
For example, if we need to round-up size of1200 and if number of divisions is 4,
the size 1200 lies between 1024 and 2048 and if we do 4 divisions between
them, the values are 1024, 1280, 1536, and 1792. So the function will
return 1280 as the nearest ceiling of power-2 division.
env setting:
export PYTORCH_CUDA_ALLOC_CONF=roundup_power2_divisions:4
ghstack-source-id: 151446017
Reviewed By: ezyang
Differential Revision: D34868036
fbshipit-source-id: 494785add16e6b37c920dcb5a2b81d4c637b554a
(cherry picked from commit 548454ccacbd8700e7ffd2d762e40b4ba37abbae)
Summary:
https://github.com/pytorch/pytorch/issues/67578 disabled reduced precision reductions for FP16 GEMMs. After benchmarking, we've found that this has substantial performance impacts for common GEMM shapes (e.g., those found in popular instantiations of multiheaded-attention) on architectures such as Volta. As these performance regressions may come as a surprise to current users, this PR adds a toggle to disable reduced precision reductions
`torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = `
rather than making it the default behavior.
CC ngimel ptrblck
stas00 Note that the behavior after the previous PR can be replicated with
`torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/67946
Reviewed By: zou3519
Differential Revision: D32289896
Pulled By: ngimel
fbshipit-source-id: a1ea2918b77e27a7d9b391e030417802a0174abe
Summary:
Powers have decided this API should be listed as beta.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/65247
Reviewed By: malfet
Differential Revision: D31057940
Pulled By: ngimel
fbshipit-source-id: 137b63cbd2c7409fecdc161a22135619bfc96bfa
Summary:
Puts memory sharing intro under Sharing memory... header, where it should have been all along.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/64996
Reviewed By: mruberry
Differential Revision: D30948619
Pulled By: ngimel
fbshipit-source-id: 5d9dd267b34e9d3fc499d4738377b58a22da1dc2
Summary:
Before https://github.com/pytorch/pytorch/pull/57833, calls to backward() or grad() synced only the calling thread's default stream with autograd leaf streams at the end of backward. This made the following weird pattern safe:
```python
with torch.cuda.stream(s):
# imagine forward used many streams, so backward leaf nodes may run on many streams
loss.backward()
# no sync
use grads
```
but a more benign-looking pattern was unsafe:
```python
with torch.cuda.stream(s):
# imagine forward used a lot of streams, so backward leaf nodes may run on many streams
loss.backward()
# backward() syncs the default stream with all the leaf streams, but does not sync s with anything,
# so counterintuitively (even though we're in the same stream context as backward()!)
# it is NOT SAFE to use grads here, and there's no easy way to make it safe,
# unless you manually sync on all the streams you used in forward,
# or move "use grads" back to default stream outside the context.
use grads
```
mruberry ngimel and I decided backward() should have the [same user-facing stream semantics as any cuda op](https://pytorch.org/docs/master/notes/cuda.html#stream-semantics-of-backward-passes).** In other words, the weird pattern should be unsafe, and the benign-looking pattern should be safe. Implementationwise, this meant backward() should sync its calling thread's current stream, not default stream, with the leaf streams.
After https://github.com/pytorch/pytorch/pull/57833, backward syncs the calling thread's current stream AND default stream with all leaf streams at the end of backward. The default stream syncs were retained for temporary backward compatibility.
This PR finishes https://github.com/pytorch/pytorch/pull/57833's work by deleting syncs on the default stream.
With this PR, graph-capturing an entire backward() call should be possible (see the [test_graph_grad_scaling diffs](https://github.com/pytorch/pytorch/compare/master...mcarilli:streaming_backwards_remove_default_syncs?expand=1#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450R3641-R3642)).
** first paragraph has a formatting error which this PR should also fix.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/60421
Reviewed By: albanD
Differential Revision: D29370344
Pulled By: ngimel
fbshipit-source-id: 3248bc5fb92fc517db0c15c897e5d7250f67d7fe
Summary:
Before https://github.com/pytorch/pytorch/pull/57833, calls to backward() or grad() synced only the calling thread's default stream with autograd leaf streams at the end of backward. This made the following weird pattern safe:
```python
with torch.cuda.stream(s):
# imagine forward used many streams, so backward leaf nodes may run on many streams
loss.backward()
# no sync
use grads
```
but a more benign-looking pattern was unsafe:
```python
with torch.cuda.stream(s):
# imagine forward used a lot of streams, so backward leaf nodes may run on many streams
loss.backward()
# backward() syncs the default stream with all the leaf streams, but does not sync s with anything,
# so counterintuitively (even though we're in the same stream context as backward()!)
# it is NOT SAFE to use grads here, and there's no easy way to make it safe,
# unless you manually sync on all the streams you used in forward,
# or move "use grads" back to default stream outside the context.
use grads
```
mruberry ngimel and I decided backward() should have the [same user-facing stream semantics as any cuda op](https://pytorch.org/docs/master/notes/cuda.html#stream-semantics-of-backward-passes).** In other words, the weird pattern should be unsafe, and the benign-looking pattern should be safe. Implementationwise, this meant backward() should sync its calling thread's current stream, not default stream, with the leaf streams.
After https://github.com/pytorch/pytorch/pull/57833, backward syncs the calling thread's current stream AND default stream with all leaf streams at the end of backward. The default stream syncs were retained for temporary backward compatibility.
This PR finishes https://github.com/pytorch/pytorch/pull/57833's work by deleting syncs on the default stream.
With this PR, graph-capturing an entire backward() call should be possible (see the [test_graph_grad_scaling diffs](https://github.com/pytorch/pytorch/compare/master...mcarilli:streaming_backwards_remove_default_syncs?expand=1#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450R3641-R3642)).
** first paragraph has a formatting error which this PR should also fix.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/60421
Reviewed By: VitalyFedyunin, albanD
Differential Revision: D29342234
Pulled By: ngimel
fbshipit-source-id: 98e6be7fdd8550872f0a78f9a66cb8dfe75abf63
Summary:
Fixes https://github.com/pytorch/pytorch/issues/35901
This change is designed to prevent fragmentation in the Caching Allocator. Permissive block splitting in the allocator allows very large blocks to be split into many pieces. Once split too finely it is unlikely all pieces will be 'free' at that same time so the original allocation can never be returned. Anecdotally, we've seen a model run out of memory failing to alloc a 50 MB block on a 32 GB card while the caching allocator is holding 13 GB of 'split free blocks'
Approach:
- Large blocks above a certain size are designated "oversize". This limit is currently set 1 decade above large, 200 MB
- Oversize blocks can not be split
- Oversize blocks must closely match the requested size (e.g. a 200 MB request will match an existing 205 MB block, but not a 300 MB block)
- In lieu of splitting oversize blocks there is a mechanism to quickly free a single oversize block (to the system allocator) to allow an appropriate size block to be allocated. This will be activated under memory pressure and will prevent _release_cached_blocks()_ from triggering
Initial performance tests show this is similar or quicker than the original strategy. Additional tests are ongoing.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/44742
Reviewed By: zou3519
Differential Revision: D29186394
Pulled By: ezyang
fbshipit-source-id: c88918836db3f51df59de6d1b3e03602ebe306a9
Summary:
Fixes https://github.com/pytorch/pytorch/issues/35901
This change is designed to prevent fragmentation in the Caching Allocator. Permissive block splitting in the allocator allows very large blocks to be split into many pieces. Once split too finely it is unlikely all pieces will be 'free' at that same time so the original allocation can never be returned. Anecdotally, we've seen a model run out of memory failing to alloc a 50 MB block on a 32 GB card while the caching allocator is holding 13 GB of 'split free blocks'
Approach:
- Large blocks above a certain size are designated "oversize". This limit is currently set 1 decade above large, 200 MB
- Oversize blocks can not be split
- Oversize blocks must closely match the requested size (e.g. a 200 MB request will match an existing 205 MB block, but not a 300 MB block)
- In lieu of splitting oversize blocks there is a mechanism to quickly free a single oversize block (to the system allocator) to allow an appropriate size block to be allocated. This will be activated under memory pressure and will prevent _release_cached_blocks()_ from triggering
Initial performance tests show this is similar or quicker than the original strategy. Additional tests are ongoing.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/44742
Reviewed By: ngimel
Differential Revision: D23752058
Pulled By: ezyang
fbshipit-source-id: ccb7c13e3cf8ef2707706726ac9aaac3a5e3d5c8
Summary:
Minor doc fix in clarifying that the input data is rounded not truncated.
CC zasdfgbnm ngimel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/49625
Reviewed By: mruberry
Differential Revision: D25668244
Pulled By: ngimel
fbshipit-source-id: ac97e41e0ca296276544f9e9f85b2cf1790d9985
Summary:
Ref https://github.com/pytorch/pytorch/issues/42175
This removes the 4 deprecated spectral functions: `torch.{fft,rfft,ifft,irfft}`. `torch.fft` is also now imported by by default.
The actual `at::native` functions are still used in `torch.stft` so can't be full removed yet. But will once https://github.com/pytorch/pytorch/issues/47601 has been merged.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48594
Reviewed By: heitorschueroff
Differential Revision: D25298929
Pulled By: mruberry
fbshipit-source-id: e36737fe8192fcd16f7e6310f8b49de478e63bf0
Summary:
I have been asked several times how to toggle this flag on libtorch. I think it would be good to mention it in the docs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/47331
Reviewed By: glaringlee
Differential Revision: D24777576
Pulled By: mruberry
fbshipit-source-id: cc2a338c477bb57e0bb74b8960c47fde99665e41
Summary:
Currently, a GraphRoot instance doesn't have an associated stream. Streaming backward synchronization logic assumes the instance ran on the default stream, and tells consumer ops to sync with the default stream. If the gradient the GraphRoot instance passes to consumer backward ops was populated on a non-default stream, we have a race condition.
The race condition can exist even if the user doesn't give a manually populated gradient:
```python
with torch.cuda.stream(side_stream):
# loss.backward() implicitly synthesizes a one-element 1.0 tensor on side_stream
# GraphRoot passes it to consumers, but consumers first sync on default stream, not side_stream.
loss.backward()
# Internally to backward(), streaming-backward logic takes over, stuff executes on the same stream it ran on in forward,
# and the side_stream context is irrelevant. GraphRoot's interaction with its first consumer(s) is the spot where
# the side_stream context causes a problem.
```
This PR fixes the race condition by associating a GraphRoot instance, at construction time, with the current stream(s) on the device(s) of the grads it will pass to consumers. (i think this relies on GraphRoot executing in the main thread, before backward thread(s) fork, because the grads were populated on the main thread.)
The test demonstrates the race condition. It fails reliably without the PR's GraphRoot diffs and passes with the GraphRoot diffs.
With the GraphRoot diffs, manually populating an incoming-gradient arg for `backward` (or `torch.autograd.grad`) and the actual call to `autograd.backward` will have the same stream-semantics relationship as any other pair of ops:
```python
# implicit population is safe
with torch.cuda.stream(side_stream):
loss.backward()
# explicit population in side stream then backward in side stream is safe
with torch.cuda.stream(side_stream):
kickoff_grad = torch.ones_like(loss)
loss.backward(gradient=kickoff_grad)
# explicit population in one stream then backward kickoff in another stream
# is NOT safe, even with this PR's diffs, but that unsafety is consistent with
# stream-semantics relationship of any pair of ops
kickoff_grad = torch.ones_like(loss)
with torch.cuda.stream(side_stream):
loss.backward(gradient=kickoff_grad)
# Safe, as you'd expect for any pair of ops
kickoff_grad = torch.ones_like(loss)
side_stream.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(side_stream):
loss.backward(gradient=kickoff_grad)
```
This PR also adds the last three examples above to cuda docs and references them from autograd docstrings.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/45787
Reviewed By: nairbv
Differential Revision: D24138376
Pulled By: albanD
fbshipit-source-id: bc4cd9390f9f0358633db530b1b09f9c1080d2a3
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/45294
While tracking down a recent memory corruption bug we found that
cuda-memcheck wasn't finding the bad accesses, and ngimel pointed out that
it's because we use a caching allocator so a lot of "out of bounds" accesses
land in a valid slab.
This PR adds a runtime knob (`PYTORCH_NO_CUDA_MEMORY_CACHING`) that, when set,
bypasses the caching allocator's caching logic so that allocations go straight
to cudaMalloc. This way, cuda-memcheck will actually work.
Test Plan:
Insert some memory errors and run a test under cuda-memcheck;
observe that cuda-memcheck flags an error where expected.
Specifically I removed the output-masking logic here:
https://github.com/pytorch/pytorch/blob/master/torch/csrc/jit/tensorexpr/cuda_codegen.cpp#L819-L826
And ran:
```
PYTORCH_NO_CUDA_MEMORY_CACHING=1 cuda-memcheck pytest -k test_superslomo test_jit_fuser_te.py
```
Reviewed By: ngimel
Differential Revision: D23964734
Pulled By: bertmaher
fbshipit-source-id: 04efd11e8aff037b9edde80c70585cb820ee6e39
Summary:
I ran `make linkcheck` using `sphinx.builders.linkcheck` on the documentation and noticed a few links weren't using HTTPS so I quickly updated them all.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/40878
Differential Revision: D22404647
Pulled By: ngimel
fbshipit-source-id: 9c9756db59197304023fddc28f252314f6cf4af3
Summary:
We should recommend DDP instead of DP. Hope we can also cherry-pick this for 1.5
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35063
Differential Revision: D20549621
Pulled By: ngimel
fbshipit-source-id: 86b1b2134664065cc6070ea4212895f993eaf543
Summary:
Adds comprehensive memory instrumentation to the CUDA caching memory allocator.
# Counters
Added comprehensive instrumentation for the following stats:
- Allocation requests (`allocation`)
- Allocated memory (`allocated_bytes`)
- Reserved segments from cudaMalloc (`segment`)
- Reserved memory (`reserved_bytes`)
- Active memory blocks (`active`)
- Active memory (`active_bytes`)
- Inactive, non-releasable blocks (`inactive_split`)
- Inactive, non-releasable memory (`inactive_split_bytes`)
- Number of failed cudaMalloc calls that result in a cache flush and retry (`cuda_malloc_retries`)
- Number of OOMs (`num_ooms`)
Except for the last two, these stats are segmented between all memory, large blocks, and small blocks. Along with the current value of each stat, historical counts of allocs/frees as well as peak usage are tracked by the allocator.
# Snapshots
Added the capability to get a "memory snapshot" – that is, to generate a complete dump of the allocator block/segment state.
# Implementation: major changes
- Added `torch.cuda.memory_stats()` (and associated C++ changes) which returns all instrumented stats as a dictionary.
- Added `torch.cuda.snapshot()` (and associated C++ changes) which returns a complete dump of the allocator block/segment state as a list of segments.
- Added memory summary generator in `torch.cuda.memory_summary()` for ease of client access to the instrumentation stats. Potentially useful to dump when catching OOMs. Sample output here: https://pastebin.com/uKZjtupq
# Implementation: minor changes
- Add error-checking helper functions for Python dicts and lists in `torch/csrc/utils/`.
- Existing memory management functions in `torch.cuda` moved from `__init__.py` to `memory.py` and star-imported to the main CUDA module.
- Add various helper functions to `torch.cuda` to return individual items from `torch.cuda.memory_stats()`.
- `torch.cuda.reset_max_memory_cached()` and `torch.cuda.reset_max_memory_allocated()` are deprecated in favor of `reset_peak_stats`. It's a bit difficult to think of a case where only one of those stats should be reset, and IMO this makes the peak stats collectively more consistent.
- `torch.cuda.memory_cached()` and `torch.cuda.max_memory_cached()` are deprecated in favor of `*memory_reserved()`.
- Style (add access modifiers in the allocator class, random nit fixes, etc.)
# Testing
- Added consistency check for stats in `test_cuda.py`. This verifies that the data from `memory_stats()` is faithful to the data from `snapshot()`.
- Ran on various basic workflows (toy example, CIFAR)
# Performance
Running the following speed benchmark: https://pastebin.com/UNndQg50
- Before this PR: 45.98 microseconds per tensor creation
- After this PR: 46.65 microseconds per tensor creation
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27361
Differential Revision: D17758747
Pulled By: jma127
fbshipit-source-id: 5a84e82d696c40c505646b9a1b4e0c3bba38aeb6
Summary:
This is a modified version of https://github.com/pytorch/pytorch/pull/14705 since commit structure for that PR is quite messy.
1. Add `IterableDataset`.
3. So we have 2 data loader mods: `Iterable` and `Map`.
1. `Iterable` if the `dataset` is an instance of `IterableDataset`
2. `Map` o.w.
3. Add better support for non-batch loading (i.e., `batch_size=None` and `batch_sampler=None`). This is useful in doing things like bulk loading.
3. Refactor `DataLoaderIter` into two classes, `_SingleProcessDataLoaderIter` and `_MultiProcessingDataLoaderIter`. Rename some methods to be more generic, e.g., `get_batch` -> `get_data`.
4. Add `torch.utils.data.get_worker_info` which returns worker information in a worker proc (e.g., worker id, dataset obj copy, etc.) and can be used in `IterableDataset.__iter__` and `worker_init_fn` to do per-worker configuration.
5. Add `ChainDataset`, which is the analog of `ConcatDataset` for `IterableDataset`.
7. Import torch.utils.data in `torch/__init__.py`
9. data loader examples and documentations
10. Use `get_worker_info` to detect whether we are in a worker process in `default_collate`
Closes https://github.com/pytorch/pytorch/issues/17909, https://github.com/pytorch/pytorch/issues/18096, https://github.com/pytorch/pytorch/issues/19946, and some of https://github.com/pytorch/pytorch/issues/13023
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19228
Reviewed By: bddppq
Differential Revision: D15058152
fbshipit-source-id: 9e081a901a071d7e4502b88054a34b450ab5ddde
Summary:
goldsborough Modify the docs to match the changes made in #4999
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12158
Differential Revision: D10103964
Pulled By: SsnL
fbshipit-source-id: 1b8692da86aca1a52e8d2e6cea76a5ad1f71e058
