Summary:
- expose a python call to set the allocator settings, it uses the same format as the value for PYTORCH_CUDA_ALLOCATOR
- keep the implementation contained within the cpp file to avoid increasing build times, only expose a function to call the setting
- make some of the Allocator Config methods public, now it looks more like a singleton
Test Plan: added the unit test
Differential Revision: D39487522
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84970
Approved by: https://github.com/zdevito
Fixes#84614
Prior to this PR CUDAGraph did not store the RNG seed, that is why `torch.cuda.manual_seed(new_seed)` would only reset the offset but not update the seed at all keeping whatever value was used during graph capture.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84967
Approved by: https://github.com/ngimel
Added ROCm support for the test_lazy_init unit test by including a condition on TEST_WITH_ROCM to switch CUDA_VISIBLE_DEVICES with HIP_VISIBLE_DEVICES.
This is needed because HIP_VISIBLE_DEVICES is set when running the single-GPU tests in CI: a47bc96fb7/.jenkins/pytorch/test.sh (L38), but this test sets CUDA_VISIBLE_DEVICES, which takes lower precedence than HIP_VISIBLE_DEVICES on ROCm.
**Testing Logs (to show behavior difference)**
12:40:41 Aug 30 11:40:41 CUDA_VISIBLE_DEVICES='0': 0
12:40:41 Aug 30 11:40:41 1
12:40:41 Aug 30 11:40:41 CUDA_VISIBLE_DEVICES='32': 32
12:40:41 Aug 30 11:40:41 1
12:40:41 Aug 30 11:40:41 HIP_VISIBLE_DEVICES='0': 0
12:40:41 Aug 30 11:40:41 1
12:40:41 Aug 30 11:40:41 HIP_VISIBLE_DEVICES='32': 32
12:40:41 Aug 30 11:40:41 0
**Passing UT**
Aug 30 17:03:15 test_lazy_init (main.TestCuda)
Aug 30 17:03:17 Validate that no CUDA calls are made during import torch call ... ok (2.471s)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84333
Approved by: https://github.com/jithunnair-amd, https://github.com/malfet
There there are conflicts between `torch.clear_autocast_cache()` and `cudaMallocAsync` from #82682.
Moreover, the use of autocast caching is not reasonable during training which is the main target of `make_graphed_callables`.
cc @eqy @ptrblck
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84289
Approved by: https://github.com/ngimel
This problem updates the the PR [#73040](https://github.com/pytorch/pytorch/pull/73040)
The compilation error in pyTorch with ROCm is successful with these changes when `NDEBUG` is enabled.
Solution:
For HIP we keep `__device__ __assert_fail()`
and for host side compilation we want to use the `__assert_fail()` from the glibc library.
Tested the code by compiling with below steps
```
python3 tools/amd_build/build_amd.py
python3 setup.py develop --cmake-only
cmake -DHIP_HIPCC_FLAGS_RELEASE="-DNDEBUG" build
cmake --build build
```
The UT test_fixed_cuda_assert_async is still skipped due performance overhead.
cc @jithunnair-amd
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81790
Approved by: https://github.com/shintaro-iwasaki, https://github.com/jeffdaily, https://github.com/malfet
Record stack trace information for each allocated segment in the allocator.
It takes around 1.5us to record 50 stack frames of context.
Since invoking a Pytorch operator is around 8us, this adds minimal overhead but we still leave it disabled by default so that we can test it more on real workloads first.
Stack information is kept both for allocated blocks and the last allocation used inactive blocks. We could potential keep around the _first_ allocation that caused the block to get allocated from cuda as well.
Potential Followups:
* stack frame entries are small (16 bytes), but the list of Frames is not compressed eventhough most frames will share some entries. So far this doesn't produce huge dumps (7MB for one real workload that uses all memory on the GPU), but it can be much smaller through compression.
* Code to format the information is slow (a few seconds) because it uses python and FlameGraph.pl
* Things allocated during the backward pass have no stack frames because they are run on another C++ thread.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82146
Approved by: https://github.com/albanD
### Description
Since the major changes for `_TypedStorage` and `_UntypedStorage` are now complete, they can be renamed to be public.
`TypedStorage._untyped()` is renamed to `TypedStorage.untyped()`.
Documentation for storages is improved as well.
### Issue
Fixes#82436
### Testing
N/A
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82438
Approved by: https://github.com/ezyang
cuDNN via the V8 API supports `bfloat16` on Ampere (`>= (8, 0)` but not older devices) which might be unexpected given current test settings. This PR fixes some dispatching to check the device capability before dispatching `bfloat16` convs and adjusts the expected failure conditions for the autocast test.
CC @xwang233 @ptrblck
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81139
Approved by: https://github.com/ngimel
Near term fix for https://github.com/pytorch/pytorch/issues/76368.
Q. Why does the user need to request `capturable=True` in the optimizer constructor? Why can't capture safety be completely automatic?
A. We need to set up capture-safe (device-side) state variables before capture. If we don't, and step() internally detects capture is underway, it's too late: the best we could do is create a device state variable and copy the current CPU value into it, which is not something we want baked into the graph.
Q. Ok, why not just do the capture-safe approach with device-side state variables all the time?
A. It incurs several more kernel launches per parameter, which could really add up and regress cpu overhead for ungraphed step()s. If the optimizer won't be captured, we should allow step() to stick with its current cpu-side state handling.
Q. But cuda RNG is a stateful thing that maintains its state on the cpu outside of capture and replay, and we capture it automatically. Why can't we do the same thing here?
A. The graph object can handle RNG generator increments because its capture_begin, capture_end, and replay() methods can see and access generator object. But the graph object has no explicit knowledge of or access to optimizer steps in its capture scope. We could let the user tell the graph object what optimizers will be stepped in its scope, ie something like
```python
graph.will_use_optimizer(opt)
graph.capture_begin()
...
```
but that seems clunkier than an optimizer constructor arg.
I'm open to other ideas, but right now I think constructor arg is necessary and the least bad approach.
Long term, https://github.com/pytorch/pytorch/issues/71274 is a better fix.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/77862
Approved by: https://github.com/ezyang
Resubmit of https://github.com/pytorch/pytorch/pull/77673, which was reverted due to Windows test failures: https://github.com/pytorch/pytorch/pull/77673#issuecomment-1130425845.
I suspect these failures happened because I don't explicitly set a side stream for graph capture in the new test.
Not setting a side stream explicitly is alright on Linux because cuda tests implicitly use a side stream.
I think Windows cuda tests implicitly use the default stream, breaking capture and leaving the backend in a bad state.
Other graphs tests explicitly set side streams and don't error in Windows builds, so i'm 95% sure doing the same for the new test will work.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/77789
Approved by: https://github.com/ezyang
In preparation of adopting future rocblas library options, it is necessary to track when the backward pass of training is executing. The scope-based helper class `BackwardPassGuard` is provided to toggle state.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/71881
Approved by: https://github.com/albanD
release_cached_blocks calls this:
```
void synchronize_and_free_events() {
TORCH_INTERNAL_ASSERT(captures_underway == 0);
```
Which means we can't call that function when we are capturing a cuda graph:
```
import torch
with torch.cuda.graph(torch.cuda.CUDAGraph()):
torch.zeros(2 ** 40, device="cuda")
```
results in:
```
RuntimeError: captures_underway == 0INTERNAL ASSERT FAILED at "/tmp/torch/c10/cuda/CUDACachingAllocator.cpp":1224, please report a bug to PyTorch.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76247
Approved by: https://github.com/ngimel
Summary:
Recent change (https://github.com/pytorch/pytorch/pull/69751) introduced the requirement of using `.coalesce()` explicitly in the tests. Unfortunately, not all tests are run in the current CI configuration and one test failure slipped through.
Fixes https://github.com/pytorch/pytorch/issues/74015.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/74027
Reviewed By: samdow
Differential Revision: D34858112
Pulled By: mruberry
fbshipit-source-id: 8904fac5e2b5335684a21f95a22646469478eb81
(cherry picked from commit 06d6e6d2a796af0e8444f4c57841a07ec4f67c9f)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73040
This patch fixes a compilation error in PyTorch with ROCm when `NDEBUG` is passed.
## Problem
Forward declaration of `__host__ __device__ __assert_fail()` is used in `c10/macros/Macros.h` for HIP compilation when `NDEBUG` is set However, HIP has `__device__ __assert_fail()` in `hip/amd_detail/amd_device_functions.h`, causing a function type error.
This issue does not appear in ROCm CI tests since it happens only when `NDEBUG` is passed.
## Solution
[EDIT] After the discussion on GitHub, we chose to entirely disable `CUDA_KERNEL_ASSERT()` for ROCm.
---
To solve this compilation error, this patch disables `CUDA_KERNEL_ASSERT()`, which uses `__assert_fail()` when
1. `c10/macros/Macros.h` is included for `*.hip` (precisely speaking, `__HIP__` or `__HIP_ARCH__` is defined), and
2. `NDEBUG` is passed.
Note that there's no impact on default compilation because, without a special compilation flag, those HIP files are compiled without `-NDEBUG`. And that's why this issue has not been found.
### Justification
[1] We cannot declare one host-and-device function for two separate host and device functions.
```
__device__ int func() {return 0};
__host__ int func() {return 0};
// Compile error (hipcc)
// __device__ __host__ int func();
```
[2] Forward declaration of a correct `__device__` only `__assert_fail()` for `__HIP__` causes the following error:
```
pytorch/c10/util/TypeCast.h:135:7: error: reference to __device__ function '__assert_fail' in __host__ __device__ function
ERROR_UNSUPPORTED_CAST
^
pytorch/c10/util/TypeCast.h:118:32: note: expanded from macro 'ERROR_UNSUPPORTED_CAST'
#define ERROR_UNSUPPORTED_CAST CUDA_KERNEL_ASSERT(false);
^
pytorch/c10/macros/Macros.h:392:5: note: expanded from macro 'CUDA_KERNEL_ASSERT'
__assert_fail(
```
[3] Maybe there's a way to properly define `__assert_fail()` for HIP + NDEBUG, but this might be too much. Please let me just disable it.
### Technical details
Error
```
pytorch/c10/macros/Macros.h:368:5: error: __host__ __device__ function '__assert_fail' cannot overload __device__ function '__assert_fail'
__assert_fail(
^
/opt/rocm/hip/include/hip/amd_detail/amd_device_functions.h:1173:6: note: previous declaration is here
void __assert_fail(const char *assertion,
```
CUDA definition (9.x) of `__assert_fail()`
```
#elif defined(__GNUC__)
extern __host__ __device__ __cudart_builtin__ void __assert_fail(
const char *, const char *, unsigned int, const char *)
__THROW;
```
ROCm definition (the latest version)
```
// 2b59661f3e/include/hip/amd_detail/amd_device_functions.h (L1172-L1177)
extern "C" __device__ __attribute__((noinline)) __attribute__((weak))
void __assert_fail(const char *assertion,
const char *file,
unsigned int line,
const char *function);
```
Test Plan:
CI + reproducer
```
python3 tools/amd_build/build_amd.py
python3 setup.py develop --cmake-only
cmake -DHIP_HIPCC_FLAGS_RELEASE="-DNDEBUG" build
cmake --build build
```
Reviewed By: xw285cornell
Differential Revision: D34310555
fbshipit-source-id: 7542288912590533ced3f20afd2e704b6551991b
(cherry picked from commit 9e52196e36820abe36bf6427cabc7389d3ea6cb5)
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
