Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/69299https://github.com/pytorch/pytorch/pull/68906 + https://github.com/pytorch/pytorch/pull/68749 plugged one correctness hole (non-blocking copies of offset pinned memory tensors) while introducing another (non-blocking copies of pinned memory tensors with a non-standard DataPtr context).
In this revision, we use both the tensor data pointer and context to attempt to identify the originating block in the pinned memory allocator.
Test Plan: New unit tests added to cover the missing case previously.
Reviewed By: yinghai
Differential Revision: D32787087
fbshipit-source-id: 0cb0d29d7c39a13f433eb1cd423dc0d2a303c955
(cherry picked from commit 297157b1a1)
Summary:
Also fixes the documentation failing to appear and adds a test to validate that op works with multiple devices properly.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/69640
Reviewed By: ngimel
Differential Revision: D32965391
Pulled By: mruberry
fbshipit-source-id: 4fe502809b353464da8edf62d92ca9863804f08e
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/68749
The logic for asynchronous copies (either HtoD or DtoH) using cudaMemcpyAsync relies on recording an event with the caching host allocator to notify it that a given allocation has been used on a stream - and thus it should wait for that stream to proceed before reusing the host memory.
This tracking is based on the allocator maintaining a map from storage allocation pointers to some state.
If we try to record an event for a pointer we don't understand, we will silently drop the event and ignore it (9554ebe44e/aten/src/ATen/cuda/CachingHostAllocator.cpp (L171-L175)).
Thus, if we use the data_ptr of a Tensor instead of the storage allocation, then reasonable code can lead to incorrectness due to missed events.
One way this can occur is simply by slicing a tensor into sub-tensors - which have different values of `data_ptr()` but share the same storage, for example:
```
image_batch = torch.randn(M, B, C, H, W).pin_memory()
for m in range(M):
sub_batch = image_batch[m].cuda(non_blocking=True)
# sub_batch.data_ptr() != image_batch.data_ptr() except for m == 0.
# however, sub_batch.storage().data_ptr() == image_batch.storage().data_ptr() always.
```
Therefore, we instead use the storage context pointer when recording events, as this is the same state that is tracked by the caching allocator itself. This is a correctness fix, although it's hard to determine how widespread this issue is.
Using the storage context also allows us to use a more efficient structure internally to the caching allocator, which will be sent in future diffs.
Test Plan: Test added which demonstrates the issue, although it's hard to demonstrate the race explicitly.
Reviewed By: ngimel
Differential Revision: D32588785
fbshipit-source-id: d87cc5e49ff8cbf59052c3c97da5b48dd1fe75cc
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:
Fixes https://github.com/pytorch/pytorch/issues/62533.
In very rare cases, the decorator for detecting memory leak is throwing assertion, even when the test is passing, and the memory is being freed with a tiny delay. The issue is not being reproduced in internal testing, but shows up sometimes in CI environment.
Reducing the severity of such detection to warning, so as not to fail the CI tests, as the actual test is not failing, rather only the check inside the decorator is failing.
Limiting the change to ROCM only for now.
cc jeffdaily sunway513 jithunnair-amd ROCmSupport
Pull Request resolved: https://github.com/pytorch/pytorch/pull/65973
Reviewed By: anjali411
Differential Revision: D31776154
Pulled By: malfet
fbshipit-source-id: 432199fca17669648463c4177c62adb553cacefd
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/66798
get_cycles_per_ms is copied and used in a few places, move it to common_utils so that it can be used as a shared util function
ghstack-source-id: 140790599
Test Plan: unit tests
Reviewed By: pritamdamania87
Differential Revision: D31706870
fbshipit-source-id: e8dccecb13862646a19aaadd7bad7c8f414fd4ab
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/62030
Remove dtype tracking from Python Storage interface, remove all the different `<type>Storage` classes except for `ByteStorage`, and update serialization accordingly, while maintaining as much FC/BC as possible
Fixes https://github.com/pytorch/pytorch/issues/47442
* **THE SERIALIZATION FORMAT IS FULLY FC/BC.** We worked very hard to make sure this is the case. We will probably want to break FC at some point to make the serialization structure of tensors make more sense, but not today.
* There is now only a single torch.ByteStorage class. Methods like `Tensor.set_` no longer check that the dtype of storage is appropriate.
* As we no longer know what dtype of a storage is, we've **removed** the size method from Storage, replacing it with nbytes. This is to help catch otherwise silent errors where you confuse number of elements with number of bytes.
* `Storage._new_shared` takes a `nbytes` kwarg and will reject previous positional only calls. `Storage._new_with_file` and `_set_from_file` require explicit element size arguments.
* It's no longer possible to convert storages to different types using the float/double/etc methods. Instead, do the conversion using a tensor.
* It's no longer possible to allocate a typed storage directly using FloatStorage/DoubleStorage/etc constructors. Instead, construct a tensor and extract its storage. The classes still exist but they are used purely for unpickling.
* The preexisting serialization format stores dtype with storage, and in fact this dtype is used to determine the dtype of the tensor overall.
To accommodate this case, we introduce a new TypedStorage concept that exists only during unpickling time which is used to temporarily store the dtype so we can construct a tensor. **If you overrode the handling of pickling/unpickling, you MUST add handling for TypedStorage** or your serialization code will degrade to standard file-based serialization.
Original pull request: https://github.com/pytorch/pytorch/pull/59671
Reviewed By: soulitzer, ngimel
Differential Revision: D29466819
Pulled By: ezyang
fbshipit-source-id: 4a14e5d3c2b08e06e558683d97f7378a3180b00e
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/64261
Note that this does not preserve byte-for-byte compatibility with
existing names.
Test Plan:
* Rely on CI to catch gross errors.
* Merge after release cut to catch subtle issues.
Reviewed By: albanD
Differential Revision: D30700647
Pulled By: dagitses
fbshipit-source-id: 7b02f34b8fae3041240cc78fbc6bcae498c3acd4
Summary:
Graphed workloads that try to capture a full backward pass must do warmup on a non-default stream. If warmup happens on the default stream, AccumulateGrad functions might tag themselves to run on the default stream, and therefore won't be capturable.
ngimel and I suspect some test_cuda.py tests run with the default stream as the ambient stream, which breaks `test_graph_grad_scaling` because `test_graph_grad_scaling` does warmup on the ambient stream _assuming_ the ambient stream is a non-default stream.
This PR explicitly sets a side stream for the warmup in `test_graph_grad_scaling`, which is what I should have done all along because it's what the new documentation recommends.
I pushed the PR branch straight to the main pytorch repo because we need to run ci-all on it, and I'm not sure what the requirements are these days.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/64339
Reviewed By: mruberry
Differential Revision: D30690711
Pulled By: ngimel
fbshipit-source-id: 91ad75f46a11f311e25bc468ea184e22acdcc25a
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/62664
Skipping a test for ROCm because of issue #62602
Test Plan: Imported from OSS
Reviewed By: soulitzer
Differential Revision: D30079534
Pulled By: NivekT
fbshipit-source-id: a9cf35e5d3a8d218edc9c5a704d1f9599d2f38a6
Summary:
Closes https://github.com/pytorch/pytorch/issues/59846.
https://github.com/pytorch/pytorch/issues/59846 is likely paranoia, and some of the test_streaming_backward_* in test_cuda.py already use gradient stealing (ie, they start with `.grad`s as None before backward). Regardless, this PR augments one of the tests to stress gradient stealing a bit more directly.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/60065
Reviewed By: mrshenli
Differential Revision: D29779518
Pulled By: ngimel
fbshipit-source-id: ccbf278543c3adebe5f4ba0365b1dace9a14da9b
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/59844.
Streaming backwards collects "leaf streams" for AccumulateGrad functions that stash or accumulate .grad attributes for autograd leaf tensors, and syncs those streams with some ambient stream(s) so later ops can safely consume the grads on the ambient stream(s).
But, currently, streaming backwards does not collect leaf streams for grads produced out-of-place (ie, not stashed onto a .grad attribute) by `torch.autograd.grad`, because these out-of-place grads are "captured" and returned before they reach an AccumulateGrad function. Some out-of-place grads might not even have an AccumulateGrad function to go to, because `torch.autograd.grad` can be told to make grads for non-leaf temporaries.[1]
The upshot is, when streaming backwards makes ops that produce out-of-place gradients run on side streams, no ambient stream is told to sync on these side streams, so `torch.autograd.grad` doesn't offer the same post-call safe-use guarantees for grads as the leaf accumulation of `torch.autograd.backward`.
This PR ensures `torch.autograd.grad` gives the same safe-use guarantees as `torch.autograd.backward` by also stashing leaf streams for grads created out-of-place.
I augmented a streaming backwards test to include a torch.autograd.grad attempt. The test fails on current master[2] and passes with the engine.cpp diffs.
I have no idea if this bug or its fix matter to distributed autograd. pritamdamania mrshenli should take a look before it's merged.
[1] example:
```python
leaf = torch.tensor(..., requires_grad=True)
tmp = leaf * 2
loss = tmp.sum()
torch.autograd.grad(loss, inputs=(tmp, leaf))
```
Technically, because `torch.autograd.grad` can be told to produce grads for non-leaf temporaries, these streams might NOT be "leaf streams". Maybe I should rename `leaf_streams`?
[2] the way the test currently fails is fun: it reports
```
AssertionError: False is not true : Tensors failed to compare as equal!With rtol=1.3e-06 and atol=1e-05, found 0 element(s) (out of 25) whose difference(s) exceeded the margin of error (including 0 nan comparisons). The greatest difference was 0.0 (5.0 vs. 5.0), which occurred at index (0, 0).
```
I suspect this [kafka trap](https://en.wiktionary.org/wiki/Kafkatrap) happens because assertEqual does a comparison test on the device, syncs on some bool result, sees failure and prints the tensors post-sync at which point is IS safe to access the values.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/60127
Reviewed By: mrshenli
Differential Revision: D29276581
Pulled By: albanD
fbshipit-source-id: a9f797e2fd76e2f884cce5a32ecf5d9b704c88ee
Summary:
Previous is https://github.com/pytorch/pytorch/issues/57781
We add now two CUDA bindings to avoid using ctypes to fix a windows issue.
However, we use ctypes to allocate the stream and create its pointer
(we can do this with a 0-dim tensor too if it feels better).
CC. ezyang rgommers ngimel mruberry
Pull Request resolved: https://github.com/pytorch/pytorch/pull/59527
Reviewed By: albanD
Differential Revision: D29053062
Pulled By: ezyang
fbshipit-source-id: 661e7e58de98b1bdb7a0871808cd41d91fe8f13f
Summary:
This is required in https://github.com/pytorch/pytorch/pull/57110#issuecomment-828357947
We need to provide means to synchronize on externally allocated streams for dlpack support in python array data api.
cc mruberry rgommers leofang asi1024 kmaehashi
Pull Request resolved: https://github.com/pytorch/pytorch/pull/57781
Reviewed By: mrshenli
Differential Revision: D28326365
Pulled By: ezyang
fbshipit-source-id: b67858c8033949951b49a3d319f649884dfd0a91
Summary:
Graphs tests are sometimes flaky in CI ([example](https://app.circleci.com/pipelines/github/pytorch/pytorch/328930/workflows/0311199b-a0be-4802-a286-cf1e73f96c70/jobs/13793451)) because when the GPU runs near its max memory capacity (which is not unusual during a long test), sometimes, to satisfy new allocations that don't match any existing unused blocks, the caching allocator may call `synchronize_and_free_events` to wait on block end-of-life events and cudaFree unused blocks, then re-cudaMalloc a new block. For ungraphed ops this isn't a problem, but synchronizing or calling cudaFree while capturing is illegal, so `synchronize_and_free_events` raises an error if called during capture.
The graphs tests themselves don't use much memory, so calling torch.cuda.empty_cache() at some point before their captures should ensure memory is available and the captures never need `synchronize_and_free_events`.
I was already calling empty_cache() near the beginning of several graphs tests. This PR extends it to the ones I forgot.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/59233
Reviewed By: mruberry
Differential Revision: D28816691
Pulled By: ngimel
fbshipit-source-id: 5cd83e48e43b1107daed5cfa2efff0fdb4f99dff
Summary:
This is based on https://github.com/pytorch/pytorch/issues/48224.
To make `foreach` more flexible, this PR pushes unsupported cases to slow path.
Also, this adds some tests to verify that
- `foreach` functions work with tensors of different dtypes and/or memory layouts in 7bd4b2c89f
- `foreach` functions work with tensors on different devices in a list, but are on the same device if the indices are the same: def4b9b5a1
Future plans:
1. Improve the coverage of unittests using `ops` decorator & updating `foreach_unary_op_db` and creating `foreach_(binary|pointwise|minmax)_db`.
2. Support broadcasting in slow path. Ref: https://github.com/pytorch/pytorch/pull/52448
3. Support type promotion in fast path. Ref https://github.com/pytorch/pytorch/pull/52449
CC: ngimel mcarilli ptrblck
Pull Request resolved: https://github.com/pytorch/pytorch/pull/56993
Reviewed By: zou3519
Differential Revision: D28630580
Pulled By: ngimel
fbshipit-source-id: e26ee74a39a591025e18c1ead48948cb7ec53c19
Summary:
Right now** there's a bug in libcuda.so that triggers sometimes when graphs with certain topologies are replayed back to back without a sync in between. Replays that hit this bug turn into spaghetti: kernels reordered ignoring dependencies, kernels elided, corrupted results. Currently, the only workaround I know that fixes all our repros is a manual sync between replays.
I'll remove the sync (or special case it based on cuda version) in a later PR, as soon as a fixed libcuda.so is available.
The only substantive change is the cudaDeviceSynchronize, other lines changed are de-indenting an unneeded scope.
** The bug is in current and semi-recent public versions of libcuda.so. We discovered the bug recently and we're not sure yet which public release was first affected. The version that ships with 11.3 is definitely affected, versions that shipped with 11.1 and earlier are likely not affected.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/57556
Reviewed By: mruberry
Differential Revision: D28343043
Pulled By: ngimel
fbshipit-source-id: 3b907241aebdb8ad47ae96a6314a8b02de7bfa77
Summary:
https://github.com/pytorch/pytorch/pull/56433 was reverted because the test perceived internal dropout state creation as a memory leak. This PR resubmits with the leak check skipped.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/57373
Reviewed By: anjali411
Differential Revision: D28152186
Pulled By: ezyang
fbshipit-source-id: 9a593fcdbbabbb09dc4e4221191663e94b697503
Summary:
I'd like the following pattern (a natural composition of Amp with full fwd+bwd capture) to work:
```python
# Create "static_input" with dummy data, run warmup iterations,
# call optimizer.zero_grad(set_to_none=True), then
g = torch.cuda._Graph()
s.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(s):
optimizer.zero_grad(set_to_none=True)
g.capture_begin()
with autocast():
out = model(static_input)
loss = loss_fn(out)
scaler.scale(loss).backward()
g.capture_end()
torch.cuda.current_stream().wait_stream(s)
# Training loop:
for b in data:
# optimizer.zero_grad() deliberately omitted, replay()'s baked-in backward will refill statically held .grads
static_input.copy_(b)
g.replay()
scaler.step(optimizer)
scaler.update()
```
Right now `GradScaler` can't work with this pattern because `update()` creates the scale tensor for the next iteration out of place. This PR changes `update()` to act in place on a long-lived scale tensor that stays static across iterations.
I'm not sure how this change affects XLA (see https://github.com/pytorch/pytorch/pull/48570), so we shouldn't merge without approval from ailzhang yaochengji.
Tagged bc-breaking because it's a change to the amp update utility function in native_functions.yaml. The function was never meant to be user-facing though.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/55562
Reviewed By: zou3519
Differential Revision: D28046159
Pulled By: ngimel
fbshipit-source-id: 02018c221609974546c562f691e20ab6ac611910
Summary:
Cudnn rnn calls that use use cudnn dropout maintain a "state" buffer across calls. [DropoutState](fe3f6f2da2/aten/src/ATen/native/cudnn/RNN.cpp (L1388-L1402))'s lock() and unlock() ensure the current call's use of the state buffer syncs with the end of the previous call's use of the state buffer (in case the previous call was on a different stream).
Telling a capturing stream to wait on an event recorded in a non-capturing stream is an error (1). Telling a non-capturing stream to wait on an event recorded during capture is also an error (2). So DropoutState's flow can error in either of two simple use cases:
```python
rnn = nn.LSTM(512, 512, 2, dropout=0.5).cuda()
out1 = rnn(in1)
# calling cudnn rnn with dropout in capture after calling it uncaptured triggers 1
capture_stream.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(capture_stream):
graph.capture_begin()
out2 = rnn(in2)
graph.capture_end()
torch.cuda.current_stream().wait_stream(capture_stream)
# calling cudnn rnn with dropout uncaptured after calling it in capture triggers 2
out3 = rnn(in3)
```
This PR fixes both cases by telling `DropoutState::lock()`: "if the most recent end-of-usage event was in a different capture state (ie, we crossed a capturing<->noncapturing border) or in a different capture, don't sync on it." While considering the fix I had two assumptions in mind:
- only one capture using the RNN can be underway at a time in this process
- no noncapturing ops in this process are issuing RNN calls while the capture using the RNN is underway.
That second assumption seems brittle if, for example, someone wants to capture an internal region of the forward method of a model wrapped with DataParallel: multiple threads could be issuing RNN calls with some currently capturing and some not. We should talk about whether that use case seems realistic.
(Bigger-picture thoughts: I don't know if forcing calls to serialize on using the shared state buffer is the best design. And if we want to do it that way, we might as well run all cudnn rnns with dropout on a dedicated side stream synced with the surrounding stream (capturing or not), in which case I don't think this PR's event-handling diffs would be needed.)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/56433
Reviewed By: heitorschueroff
Differential Revision: D27966444
Pulled By: ezyang
fbshipit-source-id: fe0df843c521e0d48d7f2c81a17aff84c5497e20
Summary:
Safely deallocating and repurposing memory used across streams relies on recording end-of-life events in all an allocation's usage streams beyond its original allocation stream. The events are later queried to see if all GPU work in those extra streams that could have used the allocation is done (from the CPU's perspective) before repurposing the allocation for use in its original stream.
The trouble is, calling EventQuery on an ordinary event recorded in a capturing stream is illegal. Calling EventQuery while capture is underway is also illegal. So when we call `tensor.record_stream` (or `c10::cuda::cudaCachingAllocator::recordStream`) on any tensor that's used or deleted in or around a capture, we often end up with a confusing error thrown from the cudaEventQuery in DeviceCachingAllocator::process_events().
This PR enables hopefully-safe deletion of tensors used across streams in or around capture with a conservative but simple approach: don't record or process end of life events for such tensors until the allocator's sure no captures are underway. You could whiteboard cases where this causes cross-stream-used allocations to be unavailable for reuse longer than absolutely necessary, but cross-stream-used allocations are uncommon, so for practical purposes this approach's impact on the memory footprint of captured sequences should be small.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/55860
Reviewed By: ejguan
Differential Revision: D27822557
Pulled By: ezyang
fbshipit-source-id: b2e18a19d83ed05bad67a8157a14a606ed14d04e
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52859
This reverts commit 92a4ee1cf6.
Added support for bfloat16 for CUDA 11 and removed fast-path for empty input tensors that was affecting autograd graph.
Test Plan: Imported from OSS
Reviewed By: H-Huang
Differential Revision: D27402390
Pulled By: heitorschueroff
fbshipit-source-id: 73c5ccf54f3da3d29eb63c9ed3601e2fe6951034
Summary:
**BC-breaking note**: This change throws errors for cases that used to silently pass. The old behavior can be obtained by setting `error_if_nonfinite=False`
Fixes https://github.com/pytorch/pytorch/issues/46849
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53843
Reviewed By: malfet
Differential Revision: D27291838
Pulled By: jbschlosser
fbshipit-source-id: 216d191b26e1b5919a44a3af5cde6f35baf825c4
Summary:
This reduces the memory usage of matmul significantly for expanded batch size.
This reduces the peak memory usage of
```
a = torch.rand(1, 1024, 1024, device="cuda")
b = torch.rand(1024, 1024, 1, device="cuda")
out = torch.matmul(a, b)
```
From 4GB to 16MB which is not too bad.
It also fixes the same problem when `b` is not batched.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54616
Reviewed By: ailzhang
Differential Revision: D27327056
Pulled By: albanD
fbshipit-source-id: 4bb5f4015aeab4174148512f3c5b8d1ffa97bf54
Summary:
Fixes https://github.com/pytorch/pytorch/issues/53511
torch.det does depend on torch.prod, which in turn depends on several other functions, and they also depend on torch.prod, so there is a circular relationship, hence this PR will enable complex backward support for several functions at once.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48125
Reviewed By: pbelevich
Differential Revision: D27188589
Pulled By: anjali411
fbshipit-source-id: bbb80f8ecb83a0c3bea2b917627d3cd3b84eb09a
Summary:
Resubmit of https://github.com/pytorch/pytorch/pull/51436.
Apparently some non-public windows builds run cuda tests on the default stream, so I changed a few capture tests to manually ensure all captures happen on non-default streams.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54038
Reviewed By: mruberry
Differential Revision: D27068649
Pulled By: ngimel
fbshipit-source-id: 4284475fa40ee38c0f8faff05a2faa310cf8a207
Summary:
Implements https://github.com/pytorch/pytorch/issues/51075#issuecomment-768884685 and additions discussed offline with ezyang ngimel . (Calling it "simple" is charitable but it's not too bad).
[High level strategy](https://github.com/pytorch/pytorch/pull/51436/files#diff-acc6337586bf9cdcf0a684380779300ec171897d05b8569bf439820dc8c93bd5R57-R82)
The current design aggregates stats from private pools with the ordinary pools, which may or may not be what we want.
Instead of adding PrivatePools as an internal feature of DeviceAllocator, I could inherit from DeviceAllocator (eg `DevicePrivateAllocator : public DeviceAllocator`) and create separate per-graph instances of the inherited class. I'm not sure if that would be better.
Graph bindings in Python are almost unchanged from https://github.com/pytorch/pytorch/pull/48875:
```python
# Same bindings as 48875, but now implicitly grabs a private mempool
graph1.capture_begin()
graph1.capture_end()
# pool=... is new. It hints that allocations during graph2's capture may share graph1's mempool
graph2.capture_begin(pool=graph1.pool())
graph2.capture_end()
# graph3 also implicitly creates its own mempool
graph3.capture_begin()
graph3.capture_end()
```
Test plan (other suggestions appreciated):
- [x] Stop maintaining manual references for all the tensors in my existing graphs+RNG tests. If private pools somehow give bad allocations, they should start failing intermittently. They run eager ops and eager allocations mixed with graph replays, so they may expose if eager ops and replays corrupt each other.
- [x] `test_graph_two_successive`: Capture successive graphs, with the second graph using the first graph's result. Try with and without sharing a pool. Check results, also check memory stats to confirm sharing a pool saves memory.
- [x] `test_graph_concurrent_replay`: Capture some graphs in separate private pools, replay them concurrently in different streams, check the results to make sure they don't corrupt each other's memory. Capture some graphs with a shared pool, replay them concurrently in different streams, check results, confirm they DO corrupt each other's memory.
- [x] `test_graph_three_successive`: A three-graph case, checking the safe and unsafe replay patterns in [Restrictions of the Strawman API](https://github.com/pytorch/pytorch/issues/51075)).
- [x] `test_graph_memory_stats_and_use_result_after_destroy_graph`: Comprehensively check torch.cuda.memory_stats() changes that result from graph capture and delete. Check that a tensor ref created during capture and held after graph delete stays valid until the tensor itself is deleted.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51436
Reviewed By: mruberry
Differential Revision: D26993790
Pulled By: ngimel
fbshipit-source-id: a992eaee1b8c23628e7b388a5a3c26e0f80e54da
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53276
- One of the tests had a syntax error (but the test
wasn't fine grained enough to catch this; any error
was a pass)
- Doesn't work on ROCm
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Differential Revision: D26820048
Test Plan: Imported from OSS
Reviewed By: mruberry
Pulled By: ezyang
fbshipit-source-id: b02c4252d10191c3b1b78f141d008084dc860c45
Summary:
Enabling four test cases in test_cuda.py for ROCm because they are passing.
Signed-off-by: Kyle Chen <kylechen@amd.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52739
Reviewed By: H-Huang
Differential Revision: D26706321
Pulled By: ngimel
fbshipit-source-id: 6907c548c4ac4e387f0eb7c646e8a01f0d036c8a
Summary:
Fixes #{[50510](https://github.com/pytorch/pytorch/issues/50510)}
Allows ```torch.nn.parallel.scatter_gather.gather``` to accept a list of NamedTuples as input and returns a NamedTuple whose elements are tensors. I added the author's fix using the ```is_namedtuple``` function.
While testing this fix, I encountered a deprecation warning instructing me to use ```'cpu'``` instead of ```-1``` to move the outputs to the CPU. However, doing this causes an assertion error in the ```_get_device_index``` function. I solved this by handling the CPU case in the affected ```forward``` function.
rohan-varma
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51104
Reviewed By: albanD
Differential Revision: D26395578
Pulled By: rohan-varma
fbshipit-source-id: 6e98c9ce1d9f1725973c18d24a6554c1bceae465
Summary:
These tests are failing for ROCm 4.0/4.0.1 release. Disable the tests until they are fixed.
- TestCuda.test_cudnn_multiple_threads_same_device
- TestCudaFuser.test_reduction
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51510
Reviewed By: H-Huang
Differential Revision: D26205179
Pulled By: seemethere
fbshipit-source-id: 0c3d29989d711deab8b5046b458c772a1543d8ed
Summary:
Building on top of the work of anjali411 (https://github.com/pytorch/pytorch/issues/46640)
Things added in this PR:
1. Modify backward and double-backward formulas
2. Add complex support for `new module tests` and criterion tests (and add complex tests for L1)
3. Modify some existing tests to support complex
Pull Request resolved: https://github.com/pytorch/pytorch/pull/49912
Reviewed By: zhangguanheng66
Differential Revision: D25853036
Pulled By: soulitzer
fbshipit-source-id: df619f1b71c450ab2818eb17804e0c55990aa8ad
Summary:
Add a new function, torch.cuda.set_per_process_memory_fraction(fraction, device), to torch.cuda. Related: https://github.com/pytorch/pytorch/issues/18626
The fraction (float type, from 0 to 1) is used to limit memory of cashing allocator on GPU device . One can set it on any visible GPU. The allowed memory equals total memory * fraction. It will raise an OOM error when try to apply GPU memory more than the allowed value. This function is similar to Tensorflow's per_process_gpu_memory_fraction
Note， this setting is just limit the cashing allocator in one process. If you are using multiprocess, you need to put this setting in to the subprocess to limit its GPU memory, because subprocess could have its own allocator.
## usage
In some cases, one needs to split a GPU device as two parts. Can set limitation before GPU memory using.
Eg. device: 0, each part takes half memory, the code as follows:
```
torch.cuda.set_per_process_memory_fraction(0.5, 0)
```
There is an example to show what it is.
```python
import torch
torch.cuda.set_per_process_memory_fraction(0.5, 0)
torch.cuda.empty_cache()
total_memory = torch.cuda.get_device_properties(0).total_memory
# less than 0.5 will be ok:
tmp_tensor = torch.empty(int(total_memory * 0.499), dtype=torch.int8, device='cuda')
del tmp_tensordel tmp_tensor
torch.cuda.empty_cache()
# this allocation will raise a OOM:
torch.empty(total_memory // 2, dtype=torch.int8, device='cuda')
"""
It raises an error as follows:
RuntimeError: CUDA out of memory. Tried to allocate 5.59 GiB (GPU 0; 11.17 GiB total capacity; 0 bytes already allocated; 10.91 GiB free; 5.59 GiB allowed; 0 bytes reserved in total by PyTorch)
"""
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48172
Reviewed By: bdhirsh
Differential Revision: D25275381
Pulled By: VitalyFedyunin
fbshipit-source-id: d8e7af31902c2eb795d416b57011cc8a22891b8f
Summary:
Fixes https://github.com/pytorch/pytorch/issues/48049
Root cause of the issue explained [here](https://github.com/pytorch/pytorch/issues/48049#issuecomment-736701769).
This PR implements albanD's suggestion to add the `!t.is_view()` check and disable autocast caching for views of tensors.
The added test checks for an increase in memory usage by comparing the initially allocated memory with the memory after 3 iterations using a single `nn.Linear` layer in a `no_grad` and `autocast` context.
After this PR the memory usage in the original issue doesn't grow anymore and yields:
```python
autocast: True
0: 0MB (peak 1165MB)
1: 0MB (peak 1264MB)
2: 0MB (peak 1265MB)
3: 0MB (peak 1265MB)
4: 0MB (peak 1265MB)
5: 0MB (peak 1265MB)
6: 0MB (peak 1265MB)
7: 0MB (peak 1265MB)
8: 0MB (peak 1265MB)
9: 0MB (peak 1265MB)
```
CC ngimel mcarilli
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48696
Reviewed By: bdhirsh
Differential Revision: D25276231
Pulled By: ngimel
fbshipit-source-id: e2571e9f166c0a6f6f569b0c28e8b9ca34132743
Summary:
Otherwise, this test will appear flaky for ROCm even though it is a generic PyTorch issue.
CC albanD
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48405
Reviewed By: mrshenli
Differential Revision: D25183473
Pulled By: ngimel
fbshipit-source-id: 0fa19b5497a713cc6c5d251598e57cc7068604be
Summary:
It is incorrect to assume that a newly recorded event will immediately query as False.
This test is flaky on ROCm due to this incorrect assumption.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/46857
Reviewed By: albanD
Differential Revision: D24565581
Pulled By: mrshenli
fbshipit-source-id: 0e9ba02cf52554957b29dbeaa5093696dc914b67
Summary:
This pull request enables the following tests on ROCm:
* TestCuda.test_tiny_half_norm_
* TestNNDeviceTypeCUDA.test_softmax_cuda_float16
* TestNNDeviceTypeCUDA.test_softmax_cuda_float32
* TestNNDeviceTypeCUDA.test_softmax_results_cuda_float16
* TestNNDeviceTypeCUDA.test_softmax_results_cuda_float32
The earlier failures, because of which the tests were skipped, were because of a precision issue for FP16 compute on MI25 hardware with ROCm 3.7 and older. The fix was delivered in the compiler in ROCm 3.8.
The pull request fixes https://github.com/pytorch/pytorch/issues/37493
cc: jeffdaily ezyang malfet mruberry
Pull Request resolved: https://github.com/pytorch/pytorch/pull/46363
Reviewed By: heitorschueroff
Differential Revision: D24325639
Pulled By: ezyang
fbshipit-source-id: a7dbb238cf38c04b6592baad40b4d71725a358c9
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/44220
Closes https://github.com/pytorch/pytorch/issues/44009
Currently if a dataloader returns objects created with a
collections.namedtuple, this will incorrectly be cast to a tuple. As a result, if we have data of these types, there can be runtime errors during the forward pass if the module is expecting a named tuple.
Fix this in
`scatter_gather.py` to resolve the issue reported in
https://github.com/pytorch/pytorch/issues/44009
ghstack-source-id: 113423287
Test Plan: CI
Reviewed By: colesbury
Differential Revision: D23536752
fbshipit-source-id: 3838e60162f29ebe424e83e474c4350ae838180b
Summary:
Amp gradient unscaling is a great use case for multi tensor apply (in fact it's the first case I wrote it for). This PR adds an MTA unscale+infcheck functor. Really excited to have it for `torch.cuda.amp`. izdeby your interface was clean and straightforward to use, great work!
Labeled as bc-breaking because the native_functions.yaml exposure of unscale+infcheck changes from [`_amp_non_finite_check_and_unscale_` to `_amp_foreach_non_finite_check_and_unscale_`]( https://github.com/pytorch/pytorch/pull/44778/files#diff-f1e4b2c15de770d978d0eb77b53a4077L6289-L6293).
The PR also modifies Unary/Binary/Pointwise Functors to
- do ops' internal math in FP32 for FP16 or bfloat16 inputs, which improves precision ([and throughput, on some architectures!](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#arithmetic-instructions)) and has no downside for the ops we care about.
- accept an instantiated op functor rather than an op functor template (`template<class> class Op`). This allows calling code to pass lambdas.
Open question: As written now, the PR has MTA Functors take care of pre- and post-casting FP16/bfloat16 inputs to FP32 before running the ops. However, alternatively, the pre- and post-math casting could be deferred/written into the ops themselves, which gives them a bit more control. I can easily rewrite it that way if you prefer.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/44778
Reviewed By: gchanan
Differential Revision: D23944102
Pulled By: izdeby
fbshipit-source-id: 22b25ccad5f69b413c77afe8733fa9cacc8e766d
Summary:
Modify contbuild to disable sanitizers, add option to run "cuda" test using TPX RE
(Note: this ignores all push blocking failures!)
Test Plan: CI
Reviewed By: walterddr, cspanda
Differential Revision: D23854578
fbshipit-source-id: 327d7cc3655c17034a6a7bc78f69967403290623
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/45159
By default, pybind11 binds void* to be capsules. After a lot of
Googling, I have concluded that this is not actually useful:
you can't actually create a capsule from Python land, and our
data_ptr() function returns an int, which means that the
function is effectively unusable. It didn't help that we had no
tests exercising it.
I've replaced the void* with uintptr_t, so that we now accept int
(and you can pass data_ptr() in directly). I'm not sure if we
should make these functions accept ctypes types; unfortunately,
pybind11 doesn't seem to have any easy way to do this.
Fixes#43006
Also added cudaHostUnregister which was requested.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Reviewed By: lw
Differential Revision: D23849731
Pulled By: ezyang
fbshipit-source-id: 8a79986f3aa9546abbd2a6a5828329ae90fd298f
Summary:
Fixes gh-42282
This adds a device-mismatch check to `addmm` on CPU and CUDA. Although it seems like the dispatcher is always selecting the CUDA version here if any of the inputs are on GPU. So in theory the CPU check is unnecessary, but probably better to err on the side of caution.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/43505
Reviewed By: mruberry
Differential Revision: D23331651
Pulled By: ngimel
fbshipit-source-id: 8eb2f64f13d87e3ca816bacec9d91fe285d83ea0
Summary:
Should close https://github.com/pytorch/pytorch/issues/36428.
The cudnn RNN API expects weights to occupy a flat buffer in memory with a particular layout. This PR implements a "speed of light" fix: [`_cudnn_rnn_cast_reflatten`](https://github.com/pytorch/pytorch/pull/42385/files#diff-9ef93b6a4fb5a06a37c562b83737ac6aR327) (the autocast wrapper assigned to `_cudnn_rnn`) copies weights to the right slices of a flat FP16 buffer with a single read/write per weight (as opposed to casting them to FP16 individually then reflattening the individual FP16 weights, which would require 2 read/writes per weight).
It isn't pretty but IMO it doesn't make rnn bindings much more tortuous than they already are.
The [test](https://github.com/pytorch/pytorch/pull/42385/files#diff-e68a7bc6ba14f212e5e7eb3727394b40R2683) tries a forward under autocast and a backward for the full cross product of RNN options and input/weight/hidden dtypes. As for all FP16list autocast tests, forward output and backward grads are checked against a control where inputs (including RNN module weights in this case) are precasted to FP16 on the python side.
Not sure who to ask for review, tagging ezyang and ngimel because Ed wrote this file (almost 2 years ago) and Natalia did the most recent major [surgery](https://github.com/pytorch/pytorch/pull/12600).
Side quests discovered:
- Should we update [persistent RNN heuristics](dbdd28207c/aten/src/ATen/native/cudnn/RNN.cpp (L584)) to include compute capability 8.0? Could be another PR but seems easy enough to include.
- Many (maybe all?!) the raw cudnn API calls in [RNN.cpp](https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/cudnn/RNN.cpp) are deprecated in cudnn 8. I don't mind taking the AI to update them since my mental cache is full of rnn stuff, but that would be a substantial separate PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/42385
Reviewed By: zhangguanheng66
Differential Revision: D23077782
Pulled By: ezyang
fbshipit-source-id: a2afb1bdab33ba0442879a703df13dc87f03ec2e
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/42139
A bunch of tests were failing with buck since we would output to
stdout and buck would fail parsing stdout in some cases.
Moving these print statements to stderr fixes this issue.
ghstack-source-id: 108606579
Test Plan: Run the offending unit tests.
Reviewed By: mrshenli
Differential Revision: D22779135
fbshipit-source-id: 789af3b16a03b68a6cb12377ed852e5b5091bbad
Summary:
In preparation for creating the new torch.fft namespace and NumPy-like fft functions, as well as supporting our goal of refactoring and reducing the size of test_torch.py, this PR creates a test suite for our spectral ops.
The existing spectral op tests from test_torch.py and test_cuda.py are moved to test_spectral_ops.py and updated to run under the device generic test framework.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/42157
Reviewed By: albanD
Differential Revision: D22811096
Pulled By: mruberry
fbshipit-source-id: e5c50f0016ea6bb8b093cd6df2dbcef6db9bb6b6
Summary:
Skipping the test test_streams as it is flaky on rocm.
cc: jeffdaily sunway513
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41697
Reviewed By: zhangguanheng66
Differential Revision: D22644600
Pulled By: malfet
fbshipit-source-id: b1b16d496e58a91c44c40d640851fd62a5d7393d
Summary:
The test asserts that the stream is "ready" but doesn't wait for the
event to be "executed" which makes it fail on some platforms where the
`query` call occurs "soon enough".
Fixes https://github.com/pytorch/pytorch/issues/38807
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41398
Reviewed By: zhangguanheng66
Differential Revision: D22540012
Pulled By: ezyang
fbshipit-source-id: 6f56d951e48133ce4f6a9a54534298b7d2877c80
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41330
`torch.cuda.check_error` is annotated as taking an `int` as argument but when running `torch.cuda.check_error(34)` one would get:
```
TypeError: cudaGetErrorString(): incompatible function arguments. The following argument types are supported:
1. (arg0: torch._C._cudart.cudaError) -> str
Invoked with: 34
```
Even if one explicitly casted the argument, running `torch.cuda.check_error(torch._C._cudart.cudaError(34))` would give:
```
AttributeError: 'str' object has no attribute 'decode'
```
This PR fixes both issues (thus allowing `check_error` to be called with a un-casted int) and adds a test.
ghstack-source-id: 107628709
Test Plan: Unit tests
Reviewed By: ezyang
Differential Revision: D22500549
fbshipit-source-id: 9170c1e466dd554d471e928b26eb472a712da9e1
Summary:
Should close https://github.com/pytorch/pytorch/issues/35810.
I decided to keep sparse handling on the Python side for clarity, although it could be moved to the C++ side (into `_amp_non_finite_check_and_unscale_`) without much trouble.
For non-fp16 sparse grads the logic is simple (call `_amp_non_finite_check_and_unscale_` on `grad._values()`) instead of `grad` itself. At least I hope it's that easy.
For fp16 sparse grads, it's tricker. Sparse tensors can be uncoalesced. From the [Note](https://pytorch.org/docs/master/sparse.html#torch.sparse.FloatTensor):
> Our sparse tensor format permits uncoalesced sparse tensors, where there may be duplicate coordinates in the indices; in this case, the interpretation is that the value at that index is the sum of all duplicate value entries.
An uncoalesced scaled fp16 grad may have values at duplicate coordinates that are all finite but large, such that adding them to make the coalesced version WOULD cause overflows.** If I checked `_values()` on the uncoalesced version, it might not report overflows, but I think it should.
So, if the grad is sparse, fp16, and uncoalesced, I still call `_amp_non_finite_check_and_unscale_` to unscale `grad._values()` in-place, but I also double-check the coalesced version by calling a second `_amp_non_finite_check_and_unscale_` on `grad.coalesce()._values()`. `coalesce()` is out-of-place, so this call doesn't redundantly affect `grad._values()`, but it does have the power to populate the same `found_inf` tensor. The `is_coalesced()` check and `coalesce()` probably aren't great for performance, but if someone needs a giant embedding table in FP16, they're better than nothing and memorywise, they'll only create a copy of nnz gradient values+indices, which is still way better than changing the whole table to FP32.
An `unscale` variant with liberty to create unscaled grads out-of-place, and replace `param.grad` instead of writing through it, could get away with just one `_amp_non_finite_check_and_unscale_`. It could say `coalesced = grad.coalesced()`, do only the stronger `_amp_non_finite_check_and_unscale_` on `coalesced._values()`, and set `param.grad = coalesced`. I could even avoid replacing `param.grad` itself by going one level deeper and setting `param.grad`'s indices and values to `coalesced`'s, but that seems brittle and still isn't truly "in place".
** you could whiteboard an uncoalesced fp32 grad with the same property, but fp32's range is big enough that I don't think it's realistic.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36786
Reviewed By: ezyang
Differential Revision: D22202832
Pulled By: ngimel
fbshipit-source-id: b70961a4b6fc3a4c1882f65e7f34874066435735
Summary:
Currently, a custom autograd function written with
```
torch.cuda.amp.custom_fwd(cast_inputs=dtype)
def forward(ctx, *args):
...
```
casts incoming floating-point CUDA tensors to `dtype` unconditionally, regardless of whether the function executes in an autocast-enabled region. I think I had the wrong idea there. Autocast-disabled regions should give the user control of input types. Also, `custom_fwd(cast_inputs=dtype)`-decorated functions' behavior should align with native fp32list/fp16list functions. C++-side casting wrappers have no effect when autocast is disabled, and `custom_fwd`'s casting should behave the same way.
The present PR changes `custom_fwd` so it only casts in autocast-enabled regions (also updates custom_fwd to ignore fp64 inputs, like the C++ wrappers).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36171
Differential Revision: D22179511
Pulled By: ngimel
fbshipit-source-id: 5a93d070179a43206066bce19da0a5a19ecaabbd
Summary:
https://github.com/pytorch/pytorch/pull/40129 fixed the error responsible for the first revert, but exposed another error in the same test.
This PR is intended as the "master copy" for merge, and it runs on full CI.
Two other PRs (restricted to run on a small subset of CI) supporting debugging DDP failures/hangs with multiple devices per process (`test_c10d.py:DistributedDataParallelTest.test_grad_layout_1devicemodule_2replicaperprocess`).
- https://github.com/pytorch/pytorch/pull/40290 tries the test with purely rowmajor contiguous params on an untouched master. In other words https://github.com/pytorch/pytorch/pull/40290 contains none of this PR's diffs aside from the test itself.
- https://github.com/pytorch/pytorch/pull/40178, for comparison, tries the test with this PR's diffs.
Both fail the same way, indicating failure is unrelated to this PR's other diffs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/40358
Differential Revision: D22165785
Pulled By: albanD
fbshipit-source-id: ac7cdd79af5c080ab74341671392dca8e717554e
Summary:
Currently, whether `AccumulateGrad` [steals](67cb018462/torch/csrc/autograd/functions/accumulate_grad.h (L42)) or [clones](67cb018462/torch/csrc/autograd/functions/accumulate_grad.h (L80)) an incoming gradient, the gradient ends up rowmajor contiguous, regardless of its param's layout. If the param's layout is channels last, or otherwise not rowmajor contigous, later kernels that apply gradients to params are forced into an uncoalesced memory access pattern for either the param or the gradient. This may not sound like a big deal but for any binary op on large tensors it's a >3X increase in gmem traffic => 3X slowdown.
The present PR changes `AccumulateGrad` to prefer, where possible, stashing gradients that match their params' layouts (["Gradient Layout Contract"](https://github.com/pytorch/pytorch/pull/34904/files#diff-ef1a56d24f66b280dcdb401502d6a796R29-R38)).
Allowing `AccumulateGrad` to stash non-rowmajor-contiguous grads means DDP allreduces and DP reduces must allow non-rowmajor-contiguous grads. This PR extends DDP and DP to allow gradients with non-rowmajor-contiguous strides as long as their layout is nonoverlapping and dense.
For good measure, I include changes that allow all five nccl primitives (allreduce, reduce, broadcast, allgather, reducescatter) to act on non-rowmajor-contiguous tensors (again as long as each input's layout is nonoverlapping and dense, and as long as all tensors participating in a given collective have the same layout). The primitive comm changes aren't necessary to enable the DDP changes, but I wasn't sure this would end up true until I had written both sets of changes. I think primitive comm enablement is reasonable to keep in the PR, especially since the code for it is simple.
Channels last params will be a major beneficiary of this PR, but I don't see it as channels-last-specific fix. The spirit is layout matching in general:
- Grads should be stashed with memory layouts matching their params.
- Src and dst tensors on opposite ends of collectives should have matching dense layouts.
This PR also updates autograd docs to describe potential BC-breaking changes below.
## BC notes
ngimel albanD gchanan
#### BC-breaking
In the common case where the user lets AccumulateGrad decide grad layouts, strides for grads of dense but non-rowmajor-contiguous params will change. Any user code that was accustomed to `view(-1)`ing these grads will break.
Also, the circumstances under which a grad can be stolen directly from the backward function that created it, as opposed to deep-copied by AccumulateGrad, have changed. In most cases we expect silent performance improvement, because we expect channels-last-aware backward kernels will create channels last gradients for channels last params. Now those can be stolen, whereas before this PR they were cloned and made rowmajor contiguous. IMO this is a mild BC breakage. Param backward hooks still see grads come in with whatever format the backward kernel gave them. The only BC breakage potential I see is if user code relies somehow on a grad in a hook having or not having the same deep memory as the eventual `param.grad`. Any such users hopefully know they're off the edge of the map and understand how to update their expectations.
#### BC escape hatches
At alband's recommendation, this PR's changes to AccumulateGrad do not alter the pre-PR code's decisions about whether grad is accumulated in or out of place. Accumulations of new grads onto an existing `.grad` attribute were (usually) in-place before this PR and remain in-place after this PR, keeping the existing `.grad`'s layout. After this PR, if the user wants to force accumulation into a grad with a particular layout, they can preset `param.grad` to a zeroed tensor with the desired strides or call `grad.contiguous(desired format)`. This likely won't be as performant as letting AccumulateGrad establish grad layouts by cloning or stealing grads with contract-compliant strides, but at least users have a control point.
One limitation (present before this PR and unchanged by this PR): Presetting `param.grad` does not ensure in-place accumulation all the time. For example, if `create_graph=True`, or if incoming `new_grad` is dense and existing `variable_grad` is sparse, accumulation occurs out of place, and the out-of-place result may not match the existing grad's strides.
----------------------------
I also noticed some potential DDP improvements that I considered out of scope but want to mention for visibility:
1. make sure Reducer's ops sync with AccumulateGrad streams
2. ~to reduce CPU overhead and incur fewer kernel launches, lazily create flat `contents` tensors by a single `cat` kernel only when a bucket is full, instead of `copy_`ing grads into `contents` individually as soon as they are received.~ PR includes a [minor change](https://github.com/pytorch/pytorch/pull/34904/files#diff-c269190a925a4b0df49eda8a8f6c5bd3R312-R315) to divide grads while copying them into flat buffers, instead of copying them in, then dividing separately. Without cat+div fusion, div-while-copying is the best we can do.
3. https://github.com/pytorch/pytorch/issues/38942
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34904
Differential Revision: D20496044
Pulled By: albanD
fbshipit-source-id: 248d680f4b1bf77b0a986451844ec6e254469217
Summary:
Benchmark with same build settings on same system.
gcc : version 7.5.0 (Ubuntu 7.5.0-3ubuntu1~18.04)
CUDA : 10.1
GPU : 1050ti
```python
import time
import torch
import numpy as np
for n, t in [(500_000, 10),
(1_000_000, 10)]:
for dtype in (torch.half, torch.float, torch.double):
# Input Setup
p = torch.from_numpy(np.random.rand(n)).to(dtype)
want = 1000
print(f'torch.multinomial(a) a.numel() == {n} for {t} times {dtype}')
start = time.time()
# Iterate
for _ in range(t):
torch.multinomial(p, want, replacement=False)
print(f'Took:', time.time() - start)
print('****' * 10)
for n, t in [(50_000, 100),
(100_000, 100)]:
for dtype in (torch.half, torch.float, torch.double):
# Input Setup
p = torch.rand(n, device='cuda', dtype=dtype)
want = 1000
print(f'torch.multinomial(a) a.numel() == {n} for {t} times {dtype}')
start = time.time()
# torch.cuda.synchronize()
# Iterate
for _ in range(t):
torch.multinomial(p, want, replacement=False)
# torch.cuda.synchronize()
print(f'CUDA Took:', time.time() - start)
```
Before:
```
torch.multinomial(a) a.numel() == 500000 for 10 times torch.float16
Took: 80.64455389976501
torch.multinomial(a) a.numel() == 500000 for 10 times torch.float32
Took: 3.7778031826019287
torch.multinomial(a) a.numel() == 500000 for 10 times torch.float64
Took: 5.045570611953735
torch.multinomial(a) a.numel() == 1000000 for 10 times torch.float16
Took: 161.53191947937012
torch.multinomial(a) a.numel() == 1000000 for 10 times torch.float32
Took: 7.640851736068726
torch.multinomial(a) a.numel() == 1000000 for 10 times torch.float64
Took: 10.399673461914062
****************************************
torch.multinomial(a) a.numel() == 50000 for 100 times torch.float16
CUDA Took: 4.873984098434448
torch.multinomial(a) a.numel() == 50000 for 100 times torch.float32
CUDA Took: 4.713594436645508
torch.multinomial(a) a.numel() == 50000 for 100 times torch.float64
CUDA Took: 11.167185068130493
torch.multinomial(a) a.numel() == 100000 for 100 times torch.float16
CUDA Took: 7.195427417755127
torch.multinomial(a) a.numel() == 100000 for 100 times torch.float32
CUDA Took: 7.669712066650391
torch.multinomial(a) a.numel() == 100000 for 100 times torch.float64
CUDA Took: 20.20938801765442
```
After:
```
torch.multinomial(a) a.numel() == 500000 for 10 times torch.float16
Took: 81.09321522712708
torch.multinomial(a) a.numel() == 500000 for 10 times torch.float32
Took: 0.06062650680541992
torch.multinomial(a) a.numel() == 500000 for 10 times torch.float64
Took: 0.0862889289855957
torch.multinomial(a) a.numel() == 1000000 for 10 times torch.float16
Took: 161.85304307937622
torch.multinomial(a) a.numel() == 1000000 for 10 times torch.float32
Took: 0.13271093368530273
torch.multinomial(a) a.numel() == 1000000 for 10 times torch.float64
Took: 0.17215657234191895
****************************************
torch.multinomial(a) a.numel() == 50000 for 100 times torch.float16
CUDA Took: 0.035035133361816406
torch.multinomial(a) a.numel() == 50000 for 100 times torch.float32
CUDA Took: 0.03631949424743652
torch.multinomial(a) a.numel() == 50000 for 100 times torch.float64
CUDA Took: 0.05507040023803711
torch.multinomial(a) a.numel() == 100000 for 100 times torch.float16
CUDA Took: 0.05105161666870117
torch.multinomial(a) a.numel() == 100000 for 100 times torch.float32
CUDA Took: 0.05449223518371582
torch.multinomial(a) a.numel() == 100000 for 100 times torch.float64
CUDA Took: 0.09161853790283203
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/39742
Differential Revision: D21976915
Pulled By: ngimel
fbshipit-source-id: 34431f814f31b6dfd6179a89f8e4fa574da7a306
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/39156
TensorList is now supported for boxing, so we can remove
unboxed only from it. I didn't check if there were other
operators that were incorrectly classified.
Fixes https://github.com/pytorch/pytorch/issues/38958
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Differential Revision: D21819821
Pulled By: ezyang
fbshipit-source-id: 6dcf91bc196554e1721d2c704f3bf524f069534b
Summary:
`_TestTorchMixin` is base class which is instantiated across multiple types.
It was inherited from `object` in order to hide it from unittest test discovery mechanism.
But this approach makes it almost impossible to use static code analyzer on the class.
This PR implements alternative approach by hiding base class into inner class, per https://stackoverflow.com/a/25695512
Change imported class access path in `test_cuda.py`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/39110
Test Plan:
run `test_torch.py --discover-tests` and `test_cuda.py --discover-tests` before and after change:
```
$ python test_torch.py --discover-tests|md5sum
2ca437bb5d65700763ce04cdacf6de3e -
$ python test_cuda.py --discover-tests|md5sum
b17df916fb0eeb6f0dd7222d7dae392c -
```
Differential Revision: D21759265
Pulled By: malfet
fbshipit-source-id: b01b06111469e551f7b78387449975e5248f6b9e
Summary:
This updates assertEqual and assertEqual-like functions to either require both or neither of atol and rtol be specified. This should improve clarity around handling precision in the test suite, and it allows us to remove the legacy positional atol argument from assertEqual. In addition, the "message" kwarg is replace with a kwarg-only "msg" argument whose name is consistent with unittest's assertEqual argument.
In the future we could make "msg" an optional third positional argument to be more consistent with unittest's assertEqual, but requiring it be specified should be clear, and we can easily update the signature to make "msg" an optional positional argument in the future, too.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/38872
Differential Revision: D21740237
Pulled By: mruberry
fbshipit-source-id: acbc027aa1d7877a49664d94db9a5fff91a07042
Summary:
This updates assertEqual and assertEqual-like functions to either require both or neither of atol and rtol be specified. This should improve clarity around handling precision in the test suite, and it allows us to remove the legacy positional atol argument from assertEqual. In addition, the "message" kwarg is replace with a kwarg-only "msg" argument whose name is consistent with unittest's assertEqual argument.
In the future we could make "msg" an optional third positional argument to be more consistent with unittest's assertEqual, but requiring it be specified should be clear, and we can easily update the signature to make "msg" an optional positional argument in the future, too.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/38872
Differential Revision: D21717199
Pulled By: mruberry
fbshipit-source-id: 9feb856f94eee911b44f6c7140a1d07c1b026d3a
Summary:
This pull request disables the unit tests that were observed to be failing once `test2` was enabled. These tests will be one by one looked at and fixed at the earliest, but until then disabling them to unblock `test2`
The pull request also disables fftPlanDestroy for rocFFT to avoid double-freeing FFT handles
cc: ezyang jeffdaily
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37427
Differential Revision: D21302909
Pulled By: ezyang
fbshipit-source-id: ecadda3778e65b7f4f97e24b932b96b9ce928616
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35615
Python 2 has reached end-of-life and is no longer supported by PyTorch.
Now we can clean up a lot of cruft that we put in place to support it.
These changes were all done manually, and I skipped anything that seemed
like it would take more than a few seconds, so I think it makes sense to
review it manually as well (though using side-by-side view and ignoring
whitespace change might be helpful).
Test Plan: CI
Differential Revision: D20842886
Pulled By: dreiss
fbshipit-source-id: 8cad4e87c45895e7ce3938a88e61157a79504aed
Summary:
The following code
```python
a = torch.randn(42,)
b = a.cuda(non_blocking=True)
```
will be **blocked** in the current master, and will **not be blocked** in pytorch 1.4 release. This can be verified by a `nvprof --print-api-trace python script.py` profiling. It is causing performance issue.
I isolated the problem, and jjsjann123 & ptrblck pointed out the fix. Thanks!
cc csarofeen ptrblck jjsjann123 VitalyFedyunin ngimel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35144
Differential Revision: D20601163
Pulled By: ngimel
fbshipit-source-id: edd2b1dabd8e615c106188f30ddb3e763bde7471
Summary:
Initial integration of eager autocasting, supporting out-of-place ops only for easier review.
Relevant issue/RFC: https://github.com/pytorch/pytorch/issues/25081
In-place ops and ops with user-supplied `out=...` can certainly be supported as well (my initial WIP https://github.com/pytorch/pytorch/pull/29552 handled many) but require substantially more complex special casing in the autocasting backend and tests. Support for these ops (much of which has already been written) will be broken into later PRs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32140
Differential Revision: D20346700
Pulled By: ezyang
fbshipit-source-id: 12d77b3917310186fbddf11c59b2794dc859131f
Summary:
This PR aims to improve the interoperability with [CuPy](https://github.com/cupy/cupy/pulls).
Instead of having two separate and conflicting memory pools. With this PR, CuPy can directly alloc memory from the PyTorch allocator by means of this proposal https://github.com/cupy/cupy/pull/3126
We would like to gather feedback to know if this approach makes sense for PyTorch, or other alternative designs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33860
Differential Revision: D20212788
Pulled By: ngimel
fbshipit-source-id: bc1e08a66da1992d26021147bf645dc65239581c
Summary:
Also, windows memory failures responsible for the earlier reversion have been fixed.
This PR (initially) contains 2 commits:
* a revert of the revert
* all changes to implement the original Apex scale update heuristic, squashed into a single commit for easier diff review
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33366
Differential Revision: D20099026
Pulled By: ngimel
fbshipit-source-id: 339b9b6bd5134bf055057492cd1eedb7e4461529
Summary:
Addresses https://github.com/pytorch/pytorch/issues/33300.
Calling .numpy() on a CUDA or non-strided (e.g. sparse) tensor segfaults in current PyTorch. This fixes the segfaults and throws the appropriate TypeError, as was intended.
Two tests, one in test_cuda.py and the other in test_sparse.py, are added to verify the behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33612
Differential Revision: D20038210
Pulled By: mruberry
fbshipit-source-id: 265531dacd37c392232fd3ec763489a62ef54795
Summary:
IIUC Python does not guarantee when an object is garbage collected. So it is possible that, some other test running before `TestCuda.test_memory_stats` creates object which is only garbage collected during `TestCuda.test_memory_stats`, causing mem stats to change and causing this test to fail. This kind of failure is very hard to debug (it took me and mcarilli and ptrblck quite a while to figure out what is happening), and it is the root cause of mcarilli's gradient scaling PR https://github.com/pytorch/pytorch/pull/26512 failing on Windows.
cc: csarofeen
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33575
Differential Revision: D20009260
Pulled By: ngimel
fbshipit-source-id: 62f2716aefac3aa6c7d1898aa8a78e6b8aa3075a
Summary:
Fixes https://github.com/pytorch/pytorch/issues/32863, (together with https://github.com/pytorch/pytorch/issues/33310 for the `TensorIterator` reductions)
This adds 64-bit indexed kernels for `THC_reduceDimIndex` and uses `THCTensor_canUse32BitIndexMath` to switch between the two at runtime.
I have a test for this locally but haven't included it here because `max` is much slower than `argmax`. To the point where the test takes several minutes to call max on just one `2**32` element tensor. That seems excessive, even for a slow test but I can push it if preferred.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33405
Differential Revision: D20010769
Pulled By: ezyang
fbshipit-source-id: a8a86f662598d5fade4d90448436418422c699a3
Summary:
This PR implements the gradient scaling API that mruberry, jjsjann123, ngimel, zdevito, gchanan and I have been discussing. Relevant issue/RFC: https://github.com/pytorch/pytorch/issues/25081.
Volume-wise, this PR is mostly documentation and tests. The Python API (found entirely in `torch/cuda/amp/amp_scaler.py`) is lightweight . The exposed functions are intended to make the implementation and control flow of gradient scaling convenient, intuitive, and performant.
The API is probably easiest to digest by looking at the documentation and examples. `docs/source/amp.rst` is the homepage for the Automatic Mixed Precision package. `docs/source/notes/amp_examples.rst` includes several examples demonstrating common but not-immediately-obvious use cases. Examples are backed by tests in `test_cuda.py` (and thankfully the tests pass :P).
Two small utility kernels have been added in `native/cuda/AmpKernels.cu` to improve performance and avoid host-device synchronizations wherever possible.
Existing optimizers, both in the wild and in Pytorch core, do not need to change to use the scaling API.
However, the API was also designed to establish a contract between user scripts and optimizers such that writers of _new_ custom optimizers have the control points they need to implement fast, optionally sync-free updates. User scripts that obey the scaling API can drop such custom optimizers in and reap performance benefits without having to change anything aside from the optimizer constructor itself. [I know what the contract with custom optimizers should be](35829f24ef/torch/cuda/amp/amp_scaler.py (L179-L184)), but I'm waiting for review on the rest of the API before I go about documenting it (it will be given a dedicated section in `docs/source/notes/amp_examples.rst`.
Currently, the gradient scaling examples do not include the auto-casting API as discussed in https://github.com/pytorch/pytorch/issues/25081. The gradient scaling API is intended to be orthogonal/modular relative to autocasting. Without auto-casting the gradient scaling API is fully use-_able_, but not terribly use-_ful_, so it's up to you guys whether you want to wait until auto-casting is ready before merging the scaling API as well.
### Todo
- [ ] How do I get c10 registered status for my two custom kernels? They're very simple.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26512
Differential Revision: D19859905
Pulled By: mruberry
fbshipit-source-id: bb8ae6966214718dfee11345db824389e4286923
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/30445
Create distributed and rpc directories under caffe/test for better management
of unit tests.
Differential Revision: D18702786
fbshipit-source-id: e9daeed0cfb846ef68806f6decfcb57c0e0e3606
Summary:
While putting finishing touches on the gradient scaling PR (https://github.com/pytorch/pytorch/pull/26512), I discovered my multi-GPU test (which uses `to()` to transfer tensors between devices) was intermittently failing with bad numerics. I knew it was going to be [a weird case from the start](https://www.imdb.com/title/tt8946378/quotes/qt4868203) and spent a week descending into madness. It turns out, for backward ops that create gradients on a different device from the device on whose stream the op is executed, the streaming backward synchronizations in [input_buffer.cpp](https://github.com/pytorch/pytorch/blob/master/torch/csrc/autograd/input_buffer.cpp#L46-L83) do not properly tell later ops to wait on the population/creation of those gradients. For example, a cross-device `to()` backward (CopyBackward Node) enqueues a cudaMemcpyAsync on the current stream of the source (incoming gradient's) device, then [syncs getCurrentCUDAStream on the destination device with the cudaMemcpyAsync](https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/cuda/Copy.cu#L76). However, `input_buffer.cpp` in such cases ([case (3)](https://github.com/pytorch/pytorch/blob/master/torch/csrc/autograd/input_buffer.cpp#L77-L81)) was not properly telling `opt_consumer_stream` to wait on the current stream of the destination device (`var`'s device).
Circumstances needed to repro in current master (see [my test](https://github.com/pytorch/pytorch/compare/master...mcarilli:backward_to_race_fix#diff-e68a7bc6ba14f212e5e7eb3727394b40R1901)):
- 2 devices, with non-default streams used for forward-pass ops on both devices (which is the default behavior in test_cuda.py)
- A `to()` that transfers a tensor requiring grad from one device to another
- A backward pass that routes back through to()'s backward (aka CopyBackward).
Under these circumstances, backward ops following CopyBackward on CopyBackward's destination device (aka the original forward-pass source device) race with the device-to-device transfer, and execute using partially-transferred data.
The present PR fixes the race condition and ensures that later ops wait on the CopyBackward transfer. This PR should also make streaming backward safe for other backward ops that span devices, as long as they play nice and populate any new gradients they create using the "current stream" of the device(s) on which they create those gradients.
There are a couple minor issues where I'm not sure of the best approach:
- Should we guard onto the var's device for the entire body of InputBuffer::add?
- I'm fairly sure we need to `recordStream` on `var` if the consumer stream is different from the stream on which (we expect) `var` was created, but calling `c10::cuda::CUDACachingAllocator::recordStream` in input_buffer.cpp might break CPU-only builds. I couldn't find a different API call to record streams that seemed CPU-build-agnostic. Could I wrap the call with a macro?
Thanks to mruberry for helpful suggestions and also the organization/naming of the stream pool and streaming backward code that allowed me to (just barely) wrap my head around the issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/31930
Differential Revision: D19517617
Pulled By: mruberry
fbshipit-source-id: 183d5460aefa5d27366b465b0473b80ec80fa044
Summary:
After several discussions, we agreed not to put any extra safety check for recordStream as either the check will cause failures in certain scenarios or there is no need to throw for user errors.
As a summary, it simply does what is described in https://github.com/pytorch/pytorch/issues/27405, check if a tensor is indeed allocated by a CUDACachingAllocator instance, if it is, then throw internal error if a block can not be retrieved.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/30870
Differential Revision: D18851669
Pulled By: yxia11
fbshipit-source-id: c2f01798cd24f1fd0f35db8764057d5d333dab95
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/30892
Fixes all outstanding lints and actually installs a properly configured
flake8
Test Plan: Imported from OSS
Differential Revision: D18862825
Pulled By: suo
fbshipit-source-id: 08e9083338a7309272e17bb803feaa42e348aa85
Summary:
Fixes https://github.com/pytorch/pytorch/issues/6962
The PR implements the handle pool mechanism for cublas as suggested by mcarilli in https://github.com/pytorch/pytorch/issues/6962#issuecomment-530563872.
~~I didn't add any unit test here yet because as mcarilli mentioned:~~
> ~~On my local machine, out of curiosity I also rewrote that test to use gemms instead of convolutions. The race condition seemed rarer, but the test did show that cublas use is not thread safe. I can share the script if you want.~~
~~Please share your script with me mcarilli. And if the race condition is rare, would it still be possible for the CI to detect it?~~
cc: colesbury
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29233
Differential Revision: D18372007
Pulled By: ezyang
fbshipit-source-id: 3492bf13410598e8452e89cf4e3e63e8df9c8c3d
Summary:
Per title. Also makes a few test_torch tests generic.
This PR removes ~half the floating_dtype decorators. Follow-up will remove the rest.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27599
Differential Revision: D17840056
Pulled By: mruberry
fbshipit-source-id: 428bb5498c452083e3608325e0b548b1d75baf2d
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:
Issue: https://github.com/pytorch/pytorch/issues/27366
The address of a view tensor might be shifted from the head of the storage.
```python
>>> x = torch.rand(10, 10, device=0, requires_grad=True)
>>> y = x[2:]
>>> hex(x.data_ptr())
'0x7f1b15c00000'
>>> hex(y.data_ptr())
'0x7f1b15c00050'
```
Currently, `Tensor.record_stream()` silently ignores shifted view tensors, because `CUDACachingAllocator` cannot find the block from the shifted address.
```c++
void recordStream(void* ptr, cuda::CUDAStream stream)
{
if (ptr) {
std::lock_guard<std::recursive_mutex> lock(mutex);
Block* block = find_allocated_block(ptr);
if (block) {
...
}
// 'block' is nullptr if 'ptr' is shifted.
}
}
```
So we cannot protect shifted view tensor which is used to compute or copy in an arbitrary stream against unexpected reallocation. Once we call `record_stream()` on a tensor, our intention is to protect the storage behind the tensor against reallocation until all works in the stream finish. This rule should be consistent regardless of the type of tensors including the view.
We can retrieve the head of the address from any types of tensors by `tensor.storage().data_ptr()`. Hence, I've thought it's better to pass to `recordStream()` rather than `tensor.data_ptr()` for consistent behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27371
Reviewed By: ezyang
Differential Revision: D17768558
Pulled By: albanD
fbshipit-source-id: 7705f52b0177625168edb6f71c07a029df471bc5
Summary:
This PR stop common_utils.py from setting the default tensor type when it's imported. See issue https://github.com/pytorch/pytorch/issues/27355. This is a frequent source of confusion for test writers.
Many tests relied on this setting (whether they knew it or not), and this PR also updates the test suite to pass without common_utils.py setting the default tensor type. Some larger test files now set the default floating dtype themselves, however. These test files are:
- test_autograd.py
- test_distributions.py
- test_jit.py
- test_nn.py
This is still a significant improvement from today, however. First, these files set the default floating dtype much more clearly than importing it from common_utils. Second, the rest of the test suite no longer sets this globally. Third, this PR is a springboard to updating those tests, too. In particular, as tests are made generic they can be moved aways from relying on this global setting.
Notable technical changes in this PR are:
- Significant updates to test_torch.py to make it pass without setting the default floating dtype globally.
- The default_floating_dtype decorator is now defined in common_utils, a couple versions of this operator were defined in test files previously.
- test_torch-specific parts of common_utils were refactored into test_torch.
- tensor creation methods in common_utils were updated to accept an optional dtype and device.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27444
Differential Revision: D17795235
Pulled By: mruberry
fbshipit-source-id: 7f77271c0c836e69f183ad9057a2c4b29f09d2e1
Summary:
- The tensor op tests generated in test_cuda.py are now generic and appear in test_torch,py
- Data previously held in auxiliary data structures and files, like test_cuda_ignores.txt, is inlined
Previously the tensor op tests used several auxiliary data structures, a file, and exception handling to filter the test suite. If a function wasn't implemented, for example, that exception would be caught. This let functions like trigamma, which isn't callable, appear to be tested. See https://github.com/pytorch/pytorch/issues/27230. Filtering from additional data stores is error prone, too. It requires developers understand what data stores are used and how they're used. The existing sources are also sometimes incorrect. The txt file claims that dist_ doesn't work on half tensors, for example, but the updated tests verify it does.
In addition to making these tests generic, this PR removes those auxiliary data structures and does not catch any exceptions. Exceptions are errors. (This also means that if something implemented breaks it will now report as an error. Previously the test suite would have reported a pass.) The test infrastructure was also simplified to not perform computations with CPU half tensors since they do not support many operations. This introduces a float<->half conversion quirk but eliminates awkward functions that would first convert cpu tensors to float, perform an operation, and convert them back.
With this change test_cuda.py is almost entirely CUDA-specific.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27210
Differential Revision: D17757907
Pulled By: mruberry
fbshipit-source-id: b3c191c379667b1a7d5361087bdf82f397f77f65
Summary:
- Makes more of test_cuda generic, including some serialization tests
- Updates some tests in test_torch to use latest extensibility points and patterns
Most remaining tests in test_cuda.py are either generated (to be moved in a follow-up PR) or deal with CUDA-specific features like streams, events, and querying CUDA devices.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27135
Differential Revision: D17696478
Pulled By: mruberry
fbshipit-source-id: 51ae424c8a72e725556a2f2bc92ad9a87244b3c0
Summary:
- Lets device generic classes be instantiated for all available device types EXCEPT those specified
- Creates TestDevicePrecision in test_torch.py, letting devices compare their results to the CPU's
- Moves 4 functions from test_cuda.py to TestDevicePrecision
- polygamma and digamma functions were cleaned up
The polygamma and digamma tests always ran with double tensors and will fail when using float tensors, despite former comments and code to the contrary. Notes were added to each function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26762
Differential Revision: D17677859
Pulled By: mruberry
fbshipit-source-id: 7cbe7d05ee0bc9b622c9127be36ced02f9c4506a
Summary:
Fixes https://github.com/pytorch/pytorch/issues/8817
This rewrites `argmax` and `argmin` to use `TensorIterator` as suggested by ngimel in https://github.com/pytorch/pytorch/issues/8817. To support this, the reduction operation is now passed the index along with the current element. I also had to change a few places where the input and output tensor `dtype`s were assumed to be the same.
Unfortunatley, this isn't enough to reimplement the variants of `min` and `max` that return indices. There are several places where multiple tensor outputs are assumed to all have the same `dtype` and so returning `pair<scalar_t, int64_t>` for `ops.project` isn't possible.
#### Performance Results
**Edit:** These timings are invalid, see below for a better perf comparison
Timings reported by [`argmax.py`](https://gist.github.com/SsnL/6898c240d22faa91da16fc41359756a2):
```
cuda : 0.1432
cpu : 26.976
numpy: 2.1350
```
So, the `TensorIterator` reductions are much faster on the GPU but significantly slower on the CPU. `htop` shows the cpu kernel using 4 cores for the cpu reduction so it's not clear what the issue is there.
Should I just revert to the old implementation on CPU or is it worth investigating further? I see that other `TensorIterator` cpu reductions are similarly faster in `numpy` e.g. `max`, `mean` `std`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26181
Differential Revision: D17631979
Pulled By: pbelevich
fbshipit-source-id: 58424818ef32cef031d436cb6191e9a6ca478581
Summary:
- Moves all ROCm-requiring test_torch tests to TestTorchDeviceType
- Moves test_stft and test_lu from test_cuda
- Moves many CUDA-only test_torch tests to TestTorchDeviceType
- Combines several test_torch CPU tests with their CUDA variants
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26435
Differential Revision: D17470469
Pulled By: mruberry
fbshipit-source-id: 90bb7fc09465c53eb2ab8da52eb2c2509775c16f
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25788
Previously, I thought that _lazy_init held the GIL throughout initialization, so
I could write the code in a single-threaded manner. This is not true; it
releases the GIL at various points, which make it possible for another thread to
race with initialization.
The correct fix is to add locking for the initialization section, so other
threads wait until the first thread finishes initializing before being let
in. There is some subtlety with how to handle lazy calls, which will call
_lazy_init reentrantly; this is handled using TLS that lets you know if you
are the initializing thread (and therefore reentrant calls are OK.)
Fixes#16559
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Differential Revision: D17366348
Pulled By: ezyang
fbshipit-source-id: 99b982709323e2370d03c127c46d87be97495916
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26252
Original commit changeset: 1375774f24c2
Testing to see if this is somehow the source of hangs on ROCm builds.
Test Plan: Change is to tests themselves. This diff is for testing the ROCm hang, however.
Differential Revision: D17390575
fbshipit-source-id: a6ffd5eb1df3971b99b6d42271a8d3d501ac79c6
Summary:
- Adds SkipCUDAIfRocm and skipCPUIfNoMkl decorators, ports corresponding tests
- Changes "SkipIf" input semantics for consistency
- Removes torchtest, which has been replaced with this new generic framework
- Refactors some common parts out of CUDA tests to TestTorchDeviceType
- Ensures all MAGMA tests run on default stream by putting the skipCUDANonDefaultStreamIf in the skipCUDAIfNoMagma decorator.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26244
Differential Revision: D17389060
Pulled By: mruberry
fbshipit-source-id: 1375774f24c2266049e6d4b899e7300ddf32eac8
Summary:
This PR moves many tests in test_torch.py to the generic device type framework. This means that many CUDA tests now run in test_torch.py and there is greater consistency in how tests for many device types are written.
One change is that all MAGMA tests are run on the default stream due to intermittent instability running MAGMA on the non-default stream. This is a known issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26232
Test Plan:
While this PR edits the tests itself, it was validated using two independent methods:
(1) The code was reviewed and it was verified that all deleted functions were actually moved.
(2) The output of the TestTorch CI was reviewed and test outputs were matched before and after this PR.
Differential Revision: D17386370
Pulled By: mruberry
fbshipit-source-id: 843d14911bbd52e8aac6861c0d9bc3d0d9418219
Summary:
This test can sometimes fail in CI.
I suspect this flakiness is because the test asks a CUDA stream to record an event, fails to synchronize the CPU with that stream, then checks if the event is recorded on the CPU. There is no guarantee this will have happened.
This one-line change preserves the intent of the test while ensuring the GPU has recorded the event before the CPU queries it.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26231
Differential Revision: D17382110
Pulled By: mruberry
fbshipit-source-id: 35b701f87f41c24b208aafde48bf10e1a54de059
Summary:
This PR addresses https://github.com/pytorch/pytorch/issues/24851 by...
1. lets device types easily register themselves for testing
2. lets tests be written to run on multiple devices and with multiple dtypes
3. provides a mechanism to instantiate those tests so they are discoverable and filterable by unittest and pytest
It refactors three tests from test_torch.py to demonstrate how to use it.
`test_diagonal` is the simplest example. Most tests just need to be modified to accept 'device' as an argument. The framework will then instantiate `test_diagonal_cpu` and `test_diagonal_cuda` (when CUDA is available) which call `test_diagonal` with the appropriate 'device' argument.
`test_neg` also has dtype variants. It accepts both 'device' and 'dtype' as arguments, and the dtypes it runs with are specified with the 'dtypes' decorator. Dtypes can be specified for all device types and particular device types. The framework instantiates tests like `test_neg_cpu_torch.float`.
`test_inverse` has device-specific dependencies. These dependencies are expressed with the sugary 'skipCUDAIfNoMagma' and 'skipCPUIfNoLapack' decorators. These decorators are device-specific so CPU testing is not skipped if Magma is not installed, and there conditions may be checked after or before the test case has been initialized. This means that skipCUDAIfNoMagma does not initialize CUDA. In fact, CUDA is only initialized if a CUDA test is run.
These instantiated tests may be run as usual and with pytest filtering it's easy to run one test on all device types, run all the tests for a particular device type, or run a device type and dtype combination.
See the note "Generic Device-Type Testing" for more detail.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25967
Differential Revision: D17381987
Pulled By: mruberry
fbshipit-source-id: 4a639641130f0a59d22da0efe0951b24b5bc4bfb
Summary:
cc: gchanan zou3519
I will look into why this is failing spuriously.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26108
Differential Revision: D17348399
Pulled By: zou3519
fbshipit-source-id: aed4ccfc3f106692d4e32acc029740309570b0c3
Summary:
Now that backward reuses forward streams calls to backward no longer need to be explicitly synced (in the great majority of cases). This is an opportunity to enable the _do_cuda_non_default_stream flag, which this PR does for test_cuda.py and test_distributions.py, where the flag was previously defined but set to false.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25989
Test Plan: Test changes the entire test suite, so the test suite is the test plan.
Differential Revision: D17329233
Pulled By: mruberry
fbshipit-source-id: 52f65b5ed53de26e35e6d022658d7fac22609f6a
Summary:
Changelog:
- De-duplicate the code in tests for torch.solve, torch.cholesky_solve, torch.triangular_solve
- Skip tests explicitly if requirements aren't met for e.g., if NumPy / SciPy aren't available in the environment
- Add generic helpers for these tests in test/common_utils.py
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25733
Test Plan:
- All tests should pass to confirm that the change is not erroneous
Clears one point specified in the discussion in https://github.com/pytorch/pytorch/issues/24333.
Differential Revision: D17315330
Pulled By: zou3519
fbshipit-source-id: c72a793e89af7e2cdb163521816d56747fd70a0e
Summary:
These unit tests pass after landing all the warp size awareness patches.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25963
Differential Revision: D17319124
Pulled By: bddppq
fbshipit-source-id: 22f5d5f1ca9c67e66a7ccf983b2d2f889a74e729
Summary:
This PR addresses issue https://github.com/pytorch/pytorch/issues/7601.
Currently models that use streams explicitly in forward have to do a lot of extra work to make backwards respect those streams. This PR extends the (recently added) input tracing (see TypeAndShape) to record the devices and streams of inputs. The autograd engine then uses this metadata to enact the expected stream parallelism without extra work from the user.
For example, a model with forward declared like (original example courtesy of ngimel):
```
def forward(self,x):
x0 = x.clone()
torch._C._cuda_setStream(self.stream1._cdata)
y0 = self.fc1(x0)
self.event1.record(stream = torch.cuda.current_stream())
torch._C._cuda_setStream(self.stream2._cdata)
y1 = self.fc2(x)
self.event2.record(stream = torch.cuda.current_stream())
self.stream2.wait_event(self.event1)
return y0 + y1
```
currently will backward on a single stream. With this change the kernels will go on the streams they are assigned in forward and both forward and backward will (for appropriate sizes) run the fc1 and fc2 kernels simultaneously.
The crux of this change is, as mentioned, an expansion of the TypeAndShape tracing and a relatively simple change to the autograd engine to use cuda events for stream synchronization. To make this efficient I also added a new AutoGPUAndStream class, exposed getting and setting streams on devices, and removed InputBuffer's AutoGPU (it's now redundant). While making these modifications I also fixed AutoGPU to check before setting the GPU when it's destroyed and to use THCudaCheck instead of its custom error handler. These changes mean that an often excessive cudaSetDevice() is not being called when inputs are added to a buffer.
In addition to allowing users to easily set and use streams that are respected in both forward and backward, this change may encourage modules to do the same and the expanded tracing might allow further optimizations in the autograd engine. (apaszke, for example, now after initial enumeration we know the number of devices that will be used by a graph task, which might help provide a sense of the "level of parallelism" we should expect.)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/8354
Test Plan: Two tests were added specifically for this behavior.
Differential Revision: D17275980
Pulled By: mruberry
fbshipit-source-id: 92bd50ac782ffa973b159fcbbadb7a083802e45d
Summary:
This best preserves accuracy, while erfinvf() should be used for half and float.
This is also consistent with the implementation before the migration: https://github.com/pytorch/pytorch/issues/24943
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25337
Differential Revision: D17102333
Pulled By: zou3519
fbshipit-source-id: 5178cff534cf5f10d86ab04d4b6c1779ffedf49e
Summary:
Improve handling of mixed-type tensor operations.
This PR affects the arithmetic (add, sub, mul, and div) operators implemented via TensorIterator (so dense but not sparse tensor ops).
For these operators, we will now promote to reasonable types where possible, following the rules defined in https://github.com/pytorch/pytorch/issues/9515, and error in cases where the cast would require floating point -> integral or non-boolean to boolean downcasts.
The details of the promotion rules are described here:
https://github.com/nairbv/pytorch/blob/promote_types_strict/docs/source/tensor_attributes.rst
Some specific backwards incompatible examples:
* now `int_tensor * float` will result in a float tensor, whereas previously the floating point operand was first cast to an int. Previously `torch.tensor(10) * 1.9` => `tensor(10)` because the 1.9 was downcast to `1`. Now the result will be the more intuitive `tensor(19)`
* Now `int_tensor *= float` will error, since the floating point result of this operation can't be cast into the in-place integral type result.
See more examples/detail in the original issue (https://github.com/pytorch/pytorch/issues/9515), in the above linked tensor_attributes.rst doc, or in the test_type_promotion.py tests added in this PR:
https://github.com/nairbv/pytorch/blob/promote_types_strict/test/test_type_promotion.py
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22273
Reviewed By: gchanan
Differential Revision: D16582230
Pulled By: nairbv
fbshipit-source-id: 4029cca891908cdbf4253e4513c617bba7306cb3
Summary:
Changelog:
- Iterate over mini batches of 262140 matrices (maximum)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/24438
Test Plan:
- Added slow tests to test the behavior in test_torch and test_cuda
Fixes https://github.com/pytorch/pytorch/issues/24403
Differential Revision: D17175603
Pulled By: soumith
fbshipit-source-id: 1abb0a1e92494cf43ef4ba9efb54a919cd18bfef
Summary:
Changelog:
- Enable broadcasting of RHS and LHS tensors for lu_solve. This means that you can now have RHS with size `3 x 2` and LHS with size `4 x 3 x 3` for instance
- Remove deprecated behavior of having 2D tensors for RHS. Now all tensors have to have a last dimension which equals the number of right hand sides
- Modified docs
Pull Request resolved: https://github.com/pytorch/pytorch/pull/24333
Test Plan: - Add tests for new behavior in test_torch.py with a port to test_cuda.py
Differential Revision: D17165463
Pulled By: zou3519
fbshipit-source-id: cda5d5496ddb29ed0182bab250b5d90f8f454aa6
Summary:
Fixing https://github.com/pytorch/pytorch/issues/24750
```
DEBUG = 0
OMP_NUM_THREADS = 1
import torch
base = torch.randn(1000000)
exp = torch.randn(1000000)
out = torch.empty_like(base)
timeit base.pow(0) +30x
old 6.26 ms ± 35.9 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
new 213 µs ± 3.38 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
timeit base.pow(1/3) +6x
old 56 ms ± 911 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 9.41 ms ± 237 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit base.pow(-1/3) +6x
old 57 ms ± 1.65 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 9.49 ms ± 293 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit base.pow(1/2) +6x
old 4.04 ms ± 14.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
new 620 µs ± 3.35 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
timeit base.pow(-1/2) +5x
old 6.56 ms ± 43 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
new 1.24 ms ± 19.3 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
timeit base.pow(1) no diff
old 322 µs ± 4.7 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
new 331 µs ± 7.26 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
timeit base.pow(-1) +3.5x
old 2.48 ms ± 15.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
new 717 µs ± 130 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
timeit base.pow(2) no diff
old 328 µs ± 7.42 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
new 324 µs ± 4.93 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
timeit base.pow(-2) +3.5x
old 2.45 ms ± 11.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
new 662 µs ± 3.83 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
timeit base.pow(3) +7x
old 2.39 ms ± 60.7 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
new 334 µs ± 7.26 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
timeit base.pow(-3) +9x
old 93.7 ms ± 5.27 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 10.3 ms ± 666 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit base.pow(123456.789) +5x
old 46.5 ms ± 418 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 9.68 ms ± 325 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit base.pow(-123456.789) +5x
old 46.5 ms ± 784 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 10 ms ± 541 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit base.pow(exp) +6x
old 60.6 ms ± 4 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 9.7 ms ± 379 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit torch.pow(0, exp) no diff
old 18.3 ms ± 859 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
new 21.2 ms ± 333 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
timeit torch.pow(1, exp) +30x
old 6.01 ms ± 81.6 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
new 203 µs ± 1.08 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
timeit torch.pow(-1, exp) +3x
old 30.8 ms ± 5.51 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 9.67 ms ± 441 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit torch.pow(42, exp) +8x
old 80.1 ms ± 1.57 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 9.51 ms ± 103 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit torch.pow(-42, exp) +2x
old 21.8 ms ± 4.37 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 9.5 ms ± 89.1 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit torch.pow(0, exp, out=out) no diff
old 20.2 ms ± 3.04 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 22.1 ms ± 648 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
timeit torch.pow(1, exp, out=out) +30x
old 6.7 ms ± 397 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
new 203 µs ± 4.64 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
timeit torch.pow(-1, exp, out=out) +3x
old 32.5 ms ± 3.61 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 9.4 ms ± 99.1 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit torch.pow(42, exp, out=out) +10x
old 91 ms ± 7.45 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 9.64 ms ± 291 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
timeit torch.pow(-42, exp, out=out) +2.5x
old 25.9 ms ± 5.03 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
new 10.1 ms ± 698 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
```
BC: enforce stronger shape requirements on the output tensor (out= keyword argument) and do not allow output tensor to be resized if it is also used as one of the inputs.
BC: enforce stronger integer tensor base power integer exponent requirement on CPU and CUDA: `Integers to negative integer powers are not allowed.`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23492
Differential Revision: D16731583
Pulled By: pbelevich
fbshipit-source-id: 4e5bf689357fe82a19371e42d48abbb7b4c1c3ca
Summary:
Fixes https://github.com/pytorch/pytorch/issues/8212
This fix is based on the idea that in-place ops(e.g. add_(...)) and out ops(e.g. tensor.add(..., out=...)) must check that the output tensor does not partially overlap with any of it's input tensors. Otherwise the result of such op is unexpected to the user. Since TensorIterator is a common backend for such ops and it's already used to check output self-overlapping, this fix is implemented in the same place.
MemOverlapStatus enum class is introduced to model two tensors overlapped state:
- TOO_HARD if at least one of them is not contiguous
- FULL if both are contiguous and share exactly the same memory array [data(), data() + numel() *itemsize()]
- PARTIAL is both are contiguous but underlying memory is shared partially, in other words memory arrays overlap but not identical.
- NO if both are contiguous but have independent non overlapping memory arrays
Performance test of clone/addcmul_/addcdiv_ with check_mem_overlaps:
a = torch.empty(10000000, device='cpu')
b = torch.randn(10000000, device='cpu')
timeit a.copy_(b)
master: 10.3 ms ± 429 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
branch: 10.2 ms ± 946 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
a = torch.empty(10000000, device='cuda')
b = torch.randn(10000000, device='cuda')
timeit a.copy_(b)
master: 373 µs ± 97.9 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each)
branch: 373 µs ± 120 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each)
a = torch.randn(1000000, device='cpu')
b = torch.randn(1000000, device='cpu')
c = torch.randn(1000000, device='cpu')
timeit a.addcmul_(b, c)
master: 2.02 ms ± 212 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
branch: 2.11 ms ± 200 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
a = torch.randn(1000000, device='cuda')
b = torch.randn(1000000, device='cuda')
c = torch.randn(1000000, device='cuda')
timeit a.addcmul_(b, c)
master: 72.6 µs ± 627 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
branch: 72.4 µs ± 18.1 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
a = torch.randn(1000000, device='cpu')
b = torch.randn(1000000, device='cpu')
c = torch.randn(1000000, device='cpu')
timeit a.addcdiv_(b, c)
master: 2.19 ms ± 583 µs per loop (mean ± std. dev. of 7 runs, 1000 loop each)
branch: 1.97 ms ± 125 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
a = torch.randn(1000000, device='cuda')
b = torch.randn(1000000, device='cuda')
c = torch.randn(1000000, device='cuda')
timeit a.addcdiv_(b, c)
master: 71.3 µs ± 1.98 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
branch: 71.7 µs ± 3.96 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
a = torch.empty(100, device='cpu')
b = torch.randn(100, device='cpu')
timeit a.copy_(b)
master: 12.1 µs ± 1.11 µs per loop (mean ± std. dev. of 7 runs, 100000 loops each)
branch: 11.1 µs ± 61.1 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
a = torch.empty(100, device='cuda')
b = torch.randn(100, device='cuda')
timeit a.copy_(b)
master: 20.9 µs ± 1.62 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
branch: 22.8 µs ± 2.63 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
a = torch.randn(100, device='cpu')
b = torch.randn(100, device='cpu')
c = torch.randn(100, device='cpu')
timeit a.addcmul_(b, c)
master: 24.1 µs ± 2.7 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
branch: 24 µs ± 91.6 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
a = torch.randn(100, device='cuda')
b = torch.randn(100, device='cuda')
c = torch.randn(100, device='cuda')
timeit a.addcmul_(b, c)
master: 34.5 µs ± 4.82 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
branch: 29.8 µs ± 496 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
a = torch.randn(100, device='cpu')
b = torch.randn(100, device='cpu')
c = torch.randn(100, device='cpu')
timeit a.addcdiv_(b, c)
master: 21.3 µs ± 210 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
branch: 23.8 µs ± 403 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
a = torch.randn(100, device='cuda')
b = torch.randn(100, device='cuda')
c = torch.randn(100, device='cuda')
timeit a.addcdiv_(b, c)
master: 30.3 µs ± 257 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
branch: 31.8 µs ± 214 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/24058
Differential Revision: D16767892
Pulled By: pbelevich
fbshipit-source-id: 0cdaaa471d003a2886b1736f8985842226b8493a
Summary:
CPU and CUDA testing code are largely the same.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23526
Reviewed By: ezyang
Differential Revision: D16586271
Pulled By: VitalyFedyunin
fbshipit-source-id: 91c70c05789120fde4718ce955de243087a8c993
Summary:
This is a similar issue as TestCuda.test_events_wait.
PyTorch test sets a policy() method to assertLeaksNoCudaTensors.
Whenever a test is run, assertLeaksNoCudaTensors is called,
which in turn calls CudaMemoryLeakCheck, which in turn calls
initialize_cuda_context_rng, where it executes torch.randn
on each device, where a kernel is launched on each device.
Since the kernel may not finish on device 0, the first assertion
self.assertTrue(s0.query()) fails.
The fix is to insert
torch.cuda.synchronize(d0)
torch.cuda.synchronize(d1)
at the beginning of the test so that previously launched kernels finish before the real
test begins.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23912
Differential Revision: D16688599
Pulled By: ezyang
fbshipit-source-id: 3de2b555e99f5bbd05727835b9d7c93a026a0519
Summary:
Changelog:
- Add batching for det / logdet / slogdet operations
- Update derivative computation to support batched inputs (and consequently batched outputs)
- Update docs
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22909
Test Plan:
- Add a `test_det_logdet_slogdet_batched` method in `test_torch.py` to test `torch.det`, `torch.logdet` and `torch.slogdet` on batched inputs. This relies on the correctness of `torch.det` on single matrices (tested by `test_det_logdet_slogdet`). A port of this test is added to `test_cuda.py`
- Add autograd tests for batched inputs
Differential Revision: D16580988
Pulled By: ezyang
fbshipit-source-id: b76c87212fbe621f42a847e3b809b5e60cfcdb7a
Summary:
Changelog:
- Rename `gels` to `lstsq`
- Fix all callsites
- Rename all tests
- Create a tentative alias for `lstsq` under the name `gels` and add a deprecation warning to not promote usage.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23460
Test Plan: - All tests should pass to confirm that the patch is correct
Differential Revision: D16547834
Pulled By: colesbury
fbshipit-source-id: b3bdb8f4c5d14c7716c3d9528e40324cc544e496
Summary:
PyTorch test sets a policy() method to assertLeaksNoCudaTensors.
Whenever a test is run, assertLeaksNoCudaTensors is called,
which in turn calls CudaMemoryLeakCheck, which in turn calls
initialize_cuda_context_rng, where it executes torch.randn
on each device, where a kernel is launched on each device.
Since the kernel may not finish on device 1, the assertion
self.assertTrue(s1.query()) fails.
The fix is to insert
torch.cuda.synchronize(d0)
torch.cuda.synchronize(d1)
at the beginning of the test so that previously launched kernels finish before the real
test begins.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23520
Differential Revision: D16547701
Pulled By: soumith
fbshipit-source-id: 42ad369f909d534e15555493d08e9bb99dd64b6a
Summary:
Rehash of https://github.com/pytorch/pytorch/issues/22322 .
Given that python 2.7 will be EOL'd on Jan 1, 2020 and we have models depending on python3.5+, we'd like to update the ROCm CI across the board to python3.6.
This PR adds the skip tests and some semantic changes for PyTorch.
Added pattern match skip for anything but the ROCm CI compared to #223222 for the python find step in the PyTorch build.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23088
Differential Revision: D16448261
Pulled By: bddppq
fbshipit-source-id: 69ece1a213418d9abf1444c496dce1c190ee07c8
Summary:
Given that python 2.7 will be EOL'd on Jan 1, 2020 and we have models depending on python3.5+, we'd like to update the ROCm CI across the board to python3.6.
This PR adds the skip tests and some semantic changes for PyTorch.
Open tasks/questions:
* RoiAlignTest.CheckCPUGPUEqual fails in the Caffe2 unit tests. Is this something expects / can be skipped?
* for testing, I've used update-alternatives on CentOS/Ubuntu to select python == python 3.6. Is this the preferred way?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22322
Differential Revision: D16199862
Pulled By: ezyang
fbshipit-source-id: 46ca6029a232f7d23f3fdb5efc33ae39a379fca8
Summary:
Changelog:
- Port SVD TH implementation to ATen/native/BatchLinearAlgebra.cpp
- Port SVD THC implementation to ATen/native/cuda/BatchLinearAlgebra.cu
- Allow batches of matrices as arguments to `torch.svd`
- Remove existing implementations in TH and THC
- Update doc string
- Update derivatives to support batching
- Modify nuclear norm implementation to use at::svd instead of _batch_svd
- Remove _batch_svd as it is redundant
Pull Request resolved: https://github.com/pytorch/pytorch/pull/21588
Test Plan:
- Add new test suite for SVD in test_torch.py with port to test_cuda.py
- Add tests in common_methods_invocations.py for derivative testing
Differential Revision: D16266115
Pulled By: nairbv
fbshipit-source-id: e89bb0dbd8f2d58bd758b7830d2389c477aa61fb
Summary:
Some of my qpth users have told me that updating to the latest version of PyTorch and replacing the btrifact/btrisolve calls with the LU ones wasn't working and I didn't believe them until I tried it myself :)
These updates have broken unpivoted LU factorizations/solves on CUDA. The LU factorization code used to return the identity permutation when pivoting wasn't used but now returns all zeros as the pivots. This PR reverts it back to return the identity permutation. I've not yet tested this code as I'm having some trouble compiling PyTorch with this and am hitting https://github.com/pytorch/pytorch/issues/21700 and am not sure how to disable that option.
Here's a MWE to reproduce the broken behavior, and my fix.
```python
torch.manual_seed(0)
n = 4
L = torch.randn(n,n)
A = L.mm(L.t()).unsqueeze(0)
b = torch.randn(1, n)
A_lu_cpu = torch.lu(A)
A_lu_cuda_nopivot = torch.lu(A.cuda(), pivot=False)
A_lu_cuda_pivot = torch.lu(A.cuda(), pivot=True)
print('A_lu_cuda_nopivot\n', A_lu_cuda_nopivot)
print('-----\nA_lu_cuda_pivot\n', A_lu_cuda_nopivot)
x_cpu = b.lu_solve(*A_lu_cpu)
x_cuda_nopivot = b.cuda().lu_solve(*A_lu_cuda_nopivot)
x_cuda_nopivot_fixed = b.cuda().lu_solve(
A_lu_cuda_nopivot[0], torch.arange(1, n+1, device='cuda:0').int())
x_cuda_pivot = b.cuda().lu_solve(*A_lu_cuda_pivot)
print(x_cpu, x_cuda_nopivot, x_cuda_nopivot_fixed, x_cuda_pivot)
```
Output:
```
A_lu_cuda_nopivot
(tensor([[[ 2.8465, -0.7560, 0.8716, -1.7337],
[-0.2656, 5.5724, -1.1316, 0.6678],
[ 0.3062, -0.2031, 1.4206, -0.5438],
[-0.6091, 0.1198, -0.3828, 1.5103]]], device='cuda:0'), tensor([[0, 0, 0, 0]], device='cuda:0', dtype=torch.int32))
-----
A_lu_cuda_pivot
(tensor([[[ 2.8465, -0.7560, 0.8716, -1.7337],
[-0.2656, 5.5724, -1.1316, 0.6678],
[ 0.3062, -0.2031, 1.4206, -0.5438],
[-0.6091, 0.1198, -0.3828, 1.5103]]], device='cuda:0'), tensor([[0, 0, 0, 0]], device='cuda:0', dtype=torch.int32))
(tensor([[-0.3121, -0.1673, -0.4450, -0.2483]]),
tensor([[-0.1661, -0.1875, -0.5694, -0.4772]], device='cuda:0'),
tensor([[-0.3121, -0.1673, -0.4450, -0.2483]], device='cuda:0'),
tensor([[-0.3121, -0.1673, -0.4450, -0.2483]], device='cuda:0'))
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22242
Differential Revision: D16049334
Pulled By: ezyang
fbshipit-source-id: 7eacae810d87ffbdf8e07159bbbc03866dd9979d
Summary:
Try to fix a sporadic failure on some CIs.
I've run this test hundreds of times on my machine (GeForce 1060, MAGMA) but I cannot reproduce this.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/21638
Differential Revision: D15827779
Pulled By: ezyang
fbshipit-source-id: 3586075e48907b3b84a101c560a34cc733514a02
Summary:
This PR covers two important points with respect to the QR decomposition:
- batching of input matrices (#7500)
- adding `some` as an option in `torch.qr` akin to NumPy's `mode` option (#10538)
Changelog:
- Enable batching for inputs to `torch.qr`
- Move QR decomposition implementation to ATen (CPU and CUDA)
- Remove existing implementations in TH/THC
- Add a `some` option to `torch.qr` that will enable users to switch between complete and reduced decomposition
- Modify doc strings
Pull Request resolved: https://github.com/pytorch/pytorch/pull/20689
Differential Revision: D15529230
Pulled By: soumith
fbshipit-source-id: 16af82b1d2db8a3a758fa8a5f798d83f5f950efb
Summary:
This #20919 without the changes to aten/src/THC/THCIntegerDivider.cuh
that broke the ROCm build.
cc bddppq
Original summary:
This fixes advanced indexing in cases where there's more than 2^31-1
bytes in the output. The `gpu_index_kernel` was missing the
`can_use_32bit_indexing`/`with_32bit_indexing` check.
This also adds a number of TORCH_INTERNAL_ASSERTS in Loops.cuh,
OffsetCalculator, and IntDivider that sizes are fit in a signed 32-bit
integer.
More comprehensive tests that require a 32 GB GPU are here:
https://gist.github.com/colesbury/e29387f5851521256dff562be07b981e
Pull Request resolved: https://github.com/pytorch/pytorch/pull/21019
Differential Revision: D15518477
Pulled By: colesbury
fbshipit-source-id: 4db5626fda76eb58250793e8aa7d4f2832db3a34
Summary:
This fixes advanced indexing in cases where there's more than 2^31-1
bytes in the output. The `gpu_index_kernel` was missing the
`can_use_32bit_indexing`/`with_32bit_indexing` check.
This also adds a number of TORCH_INTERNAL_ASSERTS in Loops.cuh,
OffsetCalculator, and IntDivider that sizes are fit in a signed 32-bit
integer.
More comprehensive tests that require a 32 GB GPU are here:
https://gist.github.com/colesbury/e29387f5851521256dff562be07b981eFixes#20888
Pull Request resolved: https://github.com/pytorch/pytorch/pull/20919
Differential Revision: D15501945
Pulled By: colesbury
fbshipit-source-id: e876e678e866d2efda8ee92c47a1d2d1310671f0
Summary:
This PR also moves Device::validate into the header file, which makes
statements like `Device d = kCPU` effectively free.
Device includes the device's index, so TensorIterator::compute_types
now implicitly checks that all CUDA inputs are on the same GPU.
Previously, this was done ad-hoc in places like TensorIterator::binary_op.
Note that zero-dim Tensor (scalars) are NOT required to be on the
same device as other inputs because they behave almost like Python numbers.
TensorIterator handles copying zero-dim Tensors to the common device.
Prior to this PR, TensorIterator would copy zero-dim Tensors between CPU
and GPU, but not between different GPUs (because Backend didn't encode
the GPU index). This removes that restriction.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/20690
Differential Revision: D15414826
Pulled By: colesbury
fbshipit-source-id: 1d0ad1f7d663252af36dd4590bcda418c2f7a09f
Summary:
Copy.cu goes from 308 to 190 lines of code. In general it uses, the same
copy strategy, using cudaMempcyAsync, a pointwise kernel, or a copy
using temporary buffers. The pointwise kernel has slightly improved
performance when broadcasting due to faster index calculation.
This deletes "`s_copy_`", "`_s_copy_from`", and "`_copy_same_type_`". The only
entry-point now is "`copy_`".
A mini-benchmark is here:
https://gist.github.com/colesbury/706de1d4e8260afe046020988410b992
Before:
https://gist.github.com/colesbury/ab454b6fe3791bff420d7bcf8c041f18
After:
https://gist.github.com/colesbury/9024d242b56ab09a9ec985fa6d1620bc
Results were measured on 2.2 GHz Broadwell; no-turbo; one thread;
compiled with GCC 7.3.0. (Results are slower than typical usage due to
turbo being off.)
The only significant differences is in the CUDA [1024] -> [1024, 1024]
broadcasting copy which is ~25% faster. I don't expect a noticeable
difference in real programs.
CPU copy overhead is a tiny bit (~200 ns) faster, but I don't expect
anyone to notice that.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/20685
Differential Revision: D15414819
Pulled By: colesbury
fbshipit-source-id: d3c6e04a5020470e3bef15b1fc09503cae5df440
Summary:
Add base support for torch.logspace. See #19220 for details.
SsnL can you feedback? Thanks a lot.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19542
Differential Revision: D15028484
Pulled By: soumith
fbshipit-source-id: fe5a58a203b279103abbc192c754c25d5031498e
Summary:
Changelog:
- Rename `potri` to `cholesky_inverse` to remain consistent with names of `cholesky` methods (`cholesky`, `cholesky_solve`)
- Fix all callsites
- Rename all tests
- Create a tentative alias for `cholesky_inverse` under the name `potri` and add a deprecation warning to not promote usage
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19498
Differential Revision: D15029901
Pulled By: ezyang
fbshipit-source-id: 2074286dc93d8744cdc9a45d54644fe57df3a57a
Summary:
This adds checks for `mul_`, `add_`, `sub_`, `div_`, the most common
binops. See #17935 for more details.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19317
Differential Revision: D14972399
Pulled By: zou3519
fbshipit-source-id: b9de331dbdb2544ee859ded725a5b5659bfd11d2
Summary:
Unit tests that hang on clock64() calls are now fixed.
test_gamma_gpu_sample is now fixed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19307
Differential Revision: D14953420
Pulled By: bddppq
fbshipit-source-id: efe807b54e047578415eb1b1e03f8ad44ea27c13
Summary:
The caching allocator tries to free all blocks on an out-of-memory
error. Previously, it did not free blocks that still had outstanding
stream uses. This change synchronizes on the outstanding events and
frees those blocks.
See #19219
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19222
Differential Revision: D14925071
Pulled By: colesbury
fbshipit-source-id: a2e9fe957ec11b00ea8e6c0468436c519667c558
Summary:
Enable multi-GPU tests that work with ROCm 2.2. Have been run three times on CI to ensure stability.
While there, remove skipIfRocm annotations for tests that depend on MAGMA. They still skip but now for the correct reason (no MAGMA) to improve our diagnostics.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19169
Differential Revision: D14924812
Pulled By: bddppq
fbshipit-source-id: 8b88f58bba58a08ddcd439e899a0abc6198fef64
Summary:
Changelog:
- Rename `btrisolve` to `lu_solve` to remain consistent with names of solve methods (`cholesky_solve`, `triangular_solve`, `solve`)
- Fix all callsites
- Rename all tests
- Create a tentative alias for `lu_solve` under the name `btrisolve` and add a deprecation warning to not promote usage
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18726
Differential Revision: D14726237
Pulled By: zou3519
fbshipit-source-id: bf25f6c79062183a4153015e0ec7ebab2c8b986b
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18598
ghimport-source-id: c74597e5e7437e94a43c163cee0639b20d0d0c6a
Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18598 Turn on F401: Unused import warning.**
This was requested by someone at Facebook; this lint is turned
on for Facebook by default. "Sure, why not."
I had to noqa a number of imports in __init__. Hypothetically
we're supposed to use __all__ in this case, but I was too lazy
to fix it. Left for future work.
Be careful! flake8-2 and flake8-3 behave differently with
respect to import resolution for # type: comments. flake8-3 will
report an import unused; flake8-2 will not. For now, I just
noqa'd all these sites.
All the changes were done by hand.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Differential Revision: D14687478
fbshipit-source-id: 30d532381e914091aadfa0d2a5a89404819663e3
Summary:
Changelog:
- Renames `btriunpack` to `lu_unpack` to remain consistent with the `lu` function interface.
- Rename all relevant tests, fix callsites
- Create a tentative alias for `lu_unpack` under the name `btriunpack` and add a deprecation warning to not promote usage.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18529
Differential Revision: D14683161
Pulled By: soumith
fbshipit-source-id: 994287eaa15c50fd74c2f1c7646edfc61e8099b1
Summary:
Changelog:
- Renames `btrifact` and `btrifact_with_info` to `lu`to remain consistent with other factorization methods (`qr` and `svd`).
- Now, we will only have one function and methods named `lu`, which performs `lu` decomposition. This function takes a get_infos kwarg, which when set to True includes a infos tensor in the tuple.
- Rename all tests, fix callsites
- Create a tentative alias for `lu` under the name `btrifact` and `btrifact_with_info`, and add a deprecation warning to not promote usage.
- Add the single batch version for `lu` so that users don't have to unsqueeze and squeeze for a single square matrix (see changes in determinant computation in `LinearAlgebra.cpp`)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18435
Differential Revision: D14680352
Pulled By: soumith
fbshipit-source-id: af58dfc11fa53d9e8e0318c720beaf5502978cd8
Summary:
Enable unit tests working with ROCm 2.3. In particular, these are unit tests where we skipped for double data types previously and some tests for multi-GPU setups.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18537
Differential Revision: D14651822
Pulled By: ezyang
fbshipit-source-id: 7dd575504ebe235a91489866c91000e9754b1235
Summary:
Changelog:
- Renames `trtrs` to `triangular_solve` to remain consistent with `cholesky_solve` and `solve`.
- Rename all tests, fix callsites
- Create a tentative alias for `triangular_solve` under the name `trtrs`, and add a deprecation warning to not promote usage.
- Move `isnan` to _torch_docs.py
- Remove unnecessary imports
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18213
Differential Revision: D14566902
Pulled By: ezyang
fbshipit-source-id: 544f57c29477df391bacd5de700bed1add456d3f
Summary:
- Remove single batch TH/THC implementations
- Remove `_batch_trtrs_lower` from `multivariate_normal`
- Add tests for batched behavior
- Modify trtrs_backward to accommodate for batched case
- Modify docs
In a future PR, this will be renamed to `triangular_solve`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18025
Differential Revision: D14523004
Pulled By: ifedan
fbshipit-source-id: 11c6a967d107f969b60e5a5c73ce6bb8099ebbe1
Summary:
Changelog:
- Renames `gesv` to `solve` to remain consistent with `cholesky_solve`.
- Rename all tests, fix callsites
- Create a tentative alias for `solve` under the name `gesv`, and add a deprecated warning to not promote usage.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18060
Differential Revision: D14503117
Pulled By: zou3519
fbshipit-source-id: 99c16d94e5970a19d7584b5915f051c030d49ff5
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17927
ghimport-source-id: 626d321e430b6b5c0ea3aa1eb9df8c1e2d058bf8
Stack:
* #17926 Implement at::has_internal_overlap helper function
* **#17927 Error out on in-place (unary) ops on tensors that have internal overlap**
On the way to #17935.
Works for CPU and CUDA on the following ops:
- abs_, acos_, asin_, atan_, ceil_, cos_, erf_, erfc_, exp_, expm1_
- floor_, log_, log10_, log1p_, log2_, round_, rsqrt_,
- sin_, sqrt_, tan_, tanh_, trunc_
This PR adds a check to see if the out/result tensor has internal
overlap. If it does, then we error out because the result **may** be
incorrect.
This is overly conservative; there are some cases where if the result is
the same as the input, the inplace operation is OK (such as floor_,
round_, and trunc_). However, the current code isn't organized in such a
way that this is easy to check, so enabling those will come in the future.
Reviewed By: ezyang
Differential Revision: D14438871
fbshipit-source-id: 15e12bf1fdb2ab7f74bb806e22bc74840bd6abd1
