Summary:
Several people have asked me about proper Amp usage with gradient accumulation. In particular, it's [unclear to people](https://github.com/NVIDIA/apex/issues/439#issuecomment-610351482) that you should only call `scaler.unscale_()` (if desired) and `scaler.update()` in iterations where you actually plan to step. This PR adds a minimal accumulation example.
I built the docs locally and it looks free from sphinx errors, at least.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36601
Differential Revision: D21082295
Pulled By: ngimel
fbshipit-source-id: b2faa6c02b9f7e1972618a0f1d5360a03f0450ac
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36678
Updated the docs to explicitly indicate that RRef control messages are
idempotent and retried upon failure.
ghstack-source-id: 102225791
Test Plan: build bot
Differential Revision: D20828041
fbshipit-source-id: ca4d71c65a453664c16c32134c47637a966b1a19
Summary:
Full details in task: https://our.intern.facebook.com/intern/tasks/?t=64776265
With pytroch 1.5+ we remove python2 support from PyTorch. All documentation under docs/ and on the pytorch.org website needs to remove Python 2 references.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36114
Differential Revision: D20901746
Pulled By: jlin27
fbshipit-source-id: 07f8dc8e6fab0b232e5048a63079cab0c433c85f
Summary:
Some more cleanup now that we no longer support python2 or 3.5 on master and eventually PyTorch 1.6 release.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35677
Differential Revision: D20838097
Pulled By: orionr
fbshipit-source-id: 95d553a1e8769f3baa395e0bc6d4ce7cd93236e9
Summary:
The original behavior of pytorch c10d only supports built-in c10d backends, such as
nccl/gloo/mpi. This patch is used to extend the c10d capability to support dynamically
loading 3rd party communication libraries which are derived from ProcessGroup base class.
related RFC is in: https://github.com/pytorch/pytorch/issues/27955
Through this way, user just need specify a 3rd party c10d backend name when invoking
torch.distributed.init_process_group(). The proposed logic will try to load corresponding
c10d backend cpp extension automatically. as for how to develop a new 3rd party c10d backend
through cpp extension, pls refer to test/cpp_extensions/cpp_c10d_extension.cpp
Pull Request resolved: https://github.com/pytorch/pytorch/pull/28068
Differential Revision: D19174838
Pulled By: agolynski
fbshipit-source-id: 3409a504a43ce7260e6f9d1207c00e87471fac62
Summary: This diff fixes the issues with current handling of debug information passed along the execution of the model. (For example, it is possible that multiple calls to the debug guard may override each other)
Test Plan: CI test/cpp/jit
Reviewed By: dzhulgakov
Differential Revision: D20602775
fbshipit-source-id: 4683957954028af81a1a0f1f12b243650230c9bb
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34710
Extending RecordFunction API to support new recording scopes (such as TorchScript functions), as well as giving more flexibility to set sampling rate.
Test Plan: unit test (test_misc.cpp/testRecordFunction)
Reviewed By: gdankel, dzhulgakov
Differential Revision: D20158523
fbshipit-source-id: a9e0819d21cc06f4952d92d43246587c36137582
Summary:
The current implementations of torch.real and torch.imag are not NumPy compatible. In particular:
- torch.real on a real tensor does not return the real tensor, like contiguous
- torch.real on a complex tensor does not return a real-valued view of the real part
- torch.imag on a complex tensor does not return a real-valued view of the imaginary part
- torch.Tensor.real and torch.Tensor.imag exist as methods, but in NumPy they are writable attributes
This PR makes the functions NumPy compatible by removing the method variants and out kwarg, restricting them to work on only real tensors, and updating the behavior of torch.real to return its input. New tests are added to test_torch.py to verify the behavior, a couple existing complex tests are skipped, and the documentation is updated to reflect the change.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35560
Differential Revision: D20714568
Pulled By: mruberry
fbshipit-source-id: 5dd092f45757b620c8426c829dd15ee997246a26
Summary:
## Motivation
This PR upgrades MKL-DNN from v0.20 to DNNL v1.2 and resolves https://github.com/pytorch/pytorch/issues/30300.
DNNL (Deep Neural Network Library) is the new brand of MKL-DNN, which improves performance, quality, and usability over the old version.
This PR focuses on the migration of all existing functionalities, including minor fixes, performance improvement and code clean up. It serves as the cornerstone of our future efforts to accommodate new features like OpenCL support, BF16 training, INT8 inference, etc. and to let the Pytorch community derive more benefits from the Intel Architecture.
<br>
## What's included?
Even DNNL has many breaking changes to the API, we managed to absorb most of them in ideep. This PR contains minimalist changes to the integration code in pytorch. Below is a summary of the changes:
<br>
**General:**
1. Replace op-level allocator with global-registered allocator
```
// before
ideep::sum::compute<AllocForMKLDNN>(scales, {x, y}, z);
// after
ideep::sum::compute(scales, {x, y}, z);
```
The allocator is now being registeted at `aten/src/ATen/native/mkldnn/IDeepRegistration.cpp`. Thereafter all tensors derived from the `cpu_engine` (by default) will use the c10 allocator.
```
RegisterEngineAllocator cpu_alloc(
ideep::engine::cpu_engine(),
[](size_t size) {
return c10::GetAllocator(c10::DeviceType::CPU)->raw_allocate(size);
},
[](void* p) {
c10::GetAllocator(c10::DeviceType::CPU)->raw_deallocate(p);
}
);
```
------
2. Simplify group convolution
We had such a scenario in convolution where ideep tensor shape mismatched aten tensor: when `groups > 1`, DNNL expects weights tensors to be 5-d with an extra group dimension, e.g. `goihw` instead of `oihw` in 2d conv case.
As shown below, a lot of extra checks came with this difference in shape before. Now we've completely hidden this difference in ideep and all tensors are going to align with pytorch's definition. So we could safely remove these checks from both aten and c2 integration code.
```
// aten/src/ATen/native/mkldnn/Conv.cpp
if (w.ndims() == x.ndims() + 1) {
AT_ASSERTM(
groups > 1,
"Only group _mkldnn_conv2d weights could have been reordered to 5d");
kernel_size[0] = w.get_dim(0) * w.get_dim(1);
std::copy_n(
w.get_dims().cbegin() + 2, x.ndims() - 1, kernel_size.begin() + 1);
} else {
std::copy_n(w.get_dims().cbegin(), x.ndims(), kernel_size.begin());
}
```
------
3. Enable DNNL built-in cache
Previously, we stored DNNL jitted kernels along with intermediate buffers inside ideep using an LRU cache. Now we are switching to the newly added DNNL built-in cache, and **no longer** caching buffers in order to reduce memory footprint.
This change will be mainly reflected in lower memory usage from memory profiling results. On the code side, we removed couple of lines of `op_key_` that depended on the ideep cache before.
------
4. Use 64-bit integer to denote dimensions
We changed the type of `ideep::dims` from `vector<int32_t>` to `vector<int64_t>`. This renders ideep dims no longer compatible with 32-bit dims used by caffe2. So we use something like `{stride_.begin(), stride_.end()}` to cast parameter `stride_` into a int64 vector.
<br>
**Misc changes in each commit:**
**Commit:** change build options
Some build options were slightly changed, mainly to avoid name collisions with other projects that include DNNL as a subproject. In addition, DNNL built-in cache is enabled by option `DNNL_ENABLE_PRIMITIVE_CACHE`.
Old | New
-- | --
WITH_EXAMPLE | MKLDNN_BUILD_EXAMPLES
WITH_TEST | MKLDNN_BUILD_TESTS
MKLDNN_THREADING | MKLDNN_CPU_RUNTIME
MKLDNN_USE_MKL | N/A (not use MKL anymore)
------
**Commit:** aten reintegration
- aten/src/ATen/native/mkldnn/BinaryOps.cpp
Implement binary ops using new operation `binary` provided by DNNL
- aten/src/ATen/native/mkldnn/Conv.cpp
Clean up group convolution checks
Simplify conv backward integration
- aten/src/ATen/native/mkldnn/MKLDNNConversions.cpp
Simplify prepacking convolution weights
- test/test_mkldnn.py
Fixed an issue in conv2d unit test: it didn't check conv results between mkldnn and aten implementation before. Instead, it compared the mkldnn with mkldnn as the default cpu path will also go into mkldnn. Now we use `torch.backends.mkldnn.flags` to fix this issue
- torch/utils/mkldnn.py
Prepack weight tensor on module `__init__` to achieve better performance significantly
------
**Commit:** caffe2 reintegration
- caffe2/ideep/ideep_utils.h
Clean up unused type definitions
- caffe2/ideep/operators/adam_op.cc & caffe2/ideep/operators/momentum_sgd_op.cc
Unify tensor initialization with `ideep::tensor::init`. Obsolete `ideep::tensor::reinit`
- caffe2/ideep/operators/conv_op.cc & caffe2/ideep/operators/quantization/int8_conv_op.cc
Clean up group convolution checks
Revamp convolution API
- caffe2/ideep/operators/conv_transpose_op.cc
Clean up group convolution checks
Clean up deconv workaround code
------
**Commit:** custom allocator
- Register c10 allocator as mentioned above
<br><br>
## Performance
We tested inference on some common models based on user scenarios, and most performance numbers are either better than or on par with DNNL 0.20.
ratio: new / old | Latency (batch=1 4T) | Throughput (batch=64 56T)
-- | -- | --
pytorch resnet18 | 121.4% | 99.7%
pytorch resnet50 | 123.1% | 106.9%
pytorch resnext101_32x8d | 116.3% | 100.1%
pytorch resnext50_32x4d | 141.9% | 104.4%
pytorch mobilenet_v2 | 163.0% | 105.8%
caffe2 alexnet | 303.0% | 99.2%
caffe2 googlenet-v3 | 101.1% | 99.2%
caffe2 inception-v1 | 102.2% | 101.7%
caffe2 mobilenet-v1 | 356.1% | 253.7%
caffe2 resnet101 | 100.4% | 99.8%
caffe2 resnet152 | 99.8% | 99.8%
caffe2 shufflenet | 141.1% | 69.0% †
caffe2 squeezenet | 98.5% | 99.2%
caffe2 vgg16 | 136.8% | 100.6%
caffe2 googlenet-v3 int8 | 100.0% | 100.7%
caffe2 mobilenet-v1 int8 | 779.2% | 943.0%
caffe2 resnet50 int8 | 99.5% | 95.5%
_Configuration:
Platform: Skylake 8180
Latency Test: 4 threads, warmup 30, iteration 500, batch size 1
Throughput Test: 56 threads, warmup 30, iteration 200, batch size 64_
† Shufflenet is one of the few models that require temp buffers during inference. The performance degradation is an expected issue since we no longer cache any buffer in the ideep. As for the solution, we suggest users opt for caching allocator like **jemalloc** as a drop-in replacement for system allocator in such heavy workloads.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32422
Test Plan:
Perf results: https://our.intern.facebook.com/intern/fblearner/details/177790608?tab=Experiment%20Results
10% improvement for ResNext with avx512, neutral on avx2
More results: https://fb.quip.com/ob10AL0bCDXW#NNNACAUoHJP
Reviewed By: yinghai
Differential Revision: D20381325
Pulled By: dzhulgakov
fbshipit-source-id: 803b906fd89ed8b723c5fcab55039efe3e4bcb77
Summary:
Adding ops to the list based on our discussion. :D
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35399
Differential Revision: D20651393
Pulled By: ailzhang
fbshipit-source-id: 8cf9026d10c0d74117953dbb68ebc2f537be956a
Summary:
I don't know why reduce_scatter collective operation is not documented so I add it to the document.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35274
Differential Revision: D20645850
Pulled By: mrshenli
fbshipit-source-id: 0a4458bff1a4e15a4593dd4dcc25e4e0f6e2265d
Summary:
Per title. See related https://github.com/pytorch/pytorch/pull/34570.
In PyTorch 1.7 the plan is for torch.div and Python's division operator to perform "true" division, like Python 3, JAX, and NumPy. To facilitate this change, this PR expands true_divide to be a method so it can cover all of torch.div's use cases.
New true_divide tests are added to test_torch.py, test_type_promotion.py, and test_sparse.py.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34794
Differential Revision: D20545507
Pulled By: mruberry
fbshipit-source-id: 55286f819716c8823d1930441a69008560ac2bd5
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34348
We need this function to do swap dequantize for prim::ListConstruct since
the output of prim::ListConstruct is a list of Tensors
Test Plan:
.
Imported from OSS
Differential Revision: D20504454
fbshipit-source-id: e6155e37da98e2219a6f79737cd46fe32a509c9f
Summary:
We should recommend DDP instead of DP. Hope we can also cherry-pick this for 1.5
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35063
Differential Revision: D20549621
Pulled By: ngimel
fbshipit-source-id: 86b1b2134664065cc6070ea4212895f993eaf543
Summary:
(Updated per review feedback)
`torch.floor_divide` is currently a function that can operate on two tensors or a tensor and a scalar (scalar x scalar floor division is handled natively by Python and the JIT has a builtin function for it). This PR updates it to:
- have an out variant: `floor_divide(x, y, out=z)`
- be a method on a tensor: `x.floor_divide(y)`
- have an in-place variant: `x.floor_divide_(y)`
- work with sparse tensors
Tests are added to test_sparse.py and test_torch.py for these new behaviors.
In addition, this PR:
- cleans up the existing sparse division and true_division code and improves their error message
- adds testing of sparse true_division to test_sparse.py
- extends existing floor_divide testing in test_torch to run on CUDA, too, not just the CPU
Unfortunately, making floor_divide a method requires breaking backwards compatibility, and floor_divide has been added to the BC whitelist since this is international. The BC issue is that the first parameter name to torch.floor_divide is changing from input to self. If you previously called torch.floor_divide with keyword arguments, e.g. torch.floor_divide(input=x, other=y), you will need to update to torch.floor_divide(self=x, other=y), or the more common torch.floor_divide(x, y).
The intent of this PR is to allow floor_divide to be substituted for division (torch.div, /) wherever division was previously used. In 1.6 we expect torch.div to perform true_division, and floor_divide is how users can continue to perform integer division with tensors.
There are two potential follow-up issues suggested by this PR:
- the test framework might benefit from additional tensor construction classes, like one to create dividends and divisors for multiple dtypes
- the test framework might benefit from a universal function test class. while methods have reasonable coverage as part of test_torch.py's TestTensorOp tests, function coverage is spotty. Universal functions are similar enough it should be possible to generate tests for them.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34552
Differential Revision: D20509850
Pulled By: mruberry
fbshipit-source-id: 2cd3c828aad67191c77f2ed8470411e246f604f8
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:
(Updated per review feedback)
`torch.floor_divide` is currently a function that can operate on two tensors or a tensor and a scalar (scalar x scalar floor division is handled natively by Python and the JIT has a builtin function for it). This PR updates it to:
- have an out variant: `floor_divide(x, y, out=z)`
- be a method on a tensor: `x.floor_divide(y)`
- have an in-place variant: `x.floor_divide_(y)`
- work with sparse tensors
Tests are added to test_sparse.py and test_torch.py for these new behaviors.
In addition, this PR:
- cleans up the existing sparse division and true_division code and improves their error message
- adds testing of sparse true_division to test_sparse.py
- extends existing floor_divide testing in test_torch to run on CUDA, too, not just the CPU
Unfortunately, making floor_divide a method requires breaking backwards compatibility, and floor_divide has been added to the BC whitelist since this is international. The BC issue is that the first parameter name to torch.floor_divide is changing from input to self. If you previously called torch.floor_divide with keyword arguments, e.g. torch.floor_divide(input=x, other=y), you will need to update to torch.floor_divide(self=x, other=y), or the more common torch.floor_divide(x, y).
The intent of this PR is to allow floor_divide to be substituted for division (torch.div, /) wherever division was previously used. In 1.6 we expect torch.div to perform true_division, and floor_divide is how users can continue to perform integer division with tensors.
There are two potential follow-up issues suggested by this PR:
- the test framework might benefit from additional tensor construction classes, like one to create dividends and divisors for multiple dtypes
- the test framework might benefit from a universal function test class. while methods have reasonable coverage as part of test_torch.py's TestTensorOp tests, function coverage is spotty. Universal functions are similar enough it should be possible to generate tests for them.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34552
Differential Revision: D20497453
Pulled By: mruberry
fbshipit-source-id: ac326f2007d8894f730d1278fef84d63bcb07b5d
Summary:
- Update API calls `backward` and `optim.step` now that we require `context_id`
- Add notes to clarify purpose of distributed autograd context (this was a source of confusion in some feedback)
- Add note that details why optimizer requires context_id
- Clearly specify that we don't have SMART mode yet
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34657
Differential Revision: D20427667
Pulled By: rohan-varma
fbshipit-source-id: 5f8a3539ccf648a78e9e9a0dfdfe389c678b1606
Summary:
This is a redo of https://github.com/pytorch/pytorch/pull/33791, which was reverted because it introduced a flaky test. The test was flaky and only flaky on Python3.5 because of dict order randomization.
I've fixed the issue with tests clobbering each other in b539fec and removed the override tests for `torch.nn.functional.tanh` and `torch.nn.functional.sigmoid`, which are deprecated and shouldn't be overridable in e0d7402. I also verified that no more test clobbering is happening.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34240
Differential Revision: D20252442
Pulled By: cpuhrsch
fbshipit-source-id: 069568e342a41c90e1dc76cbf85ba4aed47f24be
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34515
Once upon a time we thought this was necessary. In reality it is not, so
removing it.
For backcompat, our public interface (defined in `api/`) still has
typedefs to the old `script::` names.
There was only one collision: `Pass` as a `Stmt` and `Pass` as a graph
transform. I renamed one of them.
Test Plan: Imported from OSS
Differential Revision: D20353503
Pulled By: suo
fbshipit-source-id: 48bb911ce75120a8c9e0c6fb65262ef775dfba93
Summary:
This PR implements the following linear algebra algorithms for low-rank matrices:
- [x] Approximate `A` as `Q Q^H A` - using Algorithm 4.4 from [Halko et al, 2009](http://arxiv.org/abs/0909.4061).
+ exposed as `torch.lowrank.get_approximate_basis(A, q, niter=2, M=None) -> Q`
+ [x] dense matrices
+ [x] batches of dense matrices
+ [x] sparse matrices
+ [x] documentation
- [x] SVD - using Algorithm 5.1 from [Halko et al, 2009](http://arxiv.org/abs/0909.4061).
+ uses `torch.lowrank.get_approximate_basis`
+ exposed as `torch.svd_lowrank(A, q=6, niter=2, M=None) -> (U, S, V)`
+ [x] dense matrices
+ [x] batches of dense matrices
+ [x] sparse matrices
+ [x] documentation
- [x] PCA - using `torch.svd_lowrank`
+ uses `torch.svd_lowrank`
+ exposed as `torch.pca_lowrank(A, center=True, q=None, niter=2) -> (U, S, V)`
+ [x] dense matrices
+ [x] batches of dense matrices
+ [x] sparse matrices, uses non-centered sparse matrix algorithm
+ [x] documentation
- [x] generalized eigenvalue solver using the original LOBPCG algorithm [Knyazev, 2001](https://epubs.siam.org/doi/abs/10.1137/S1064827500366124)
+ exposed as `torch.lobpcg(A, B=None, k=1, method="basic", ...)`
+ [x] dense matrices
+ [x] batches of dense matrices
+ [x] sparse matrices
+ [x] documentation
- [x] generalized eigenvalue solver using robust LOBPCG with orthogonal basis selection [Stathopoulos, 2002](https://epubs.siam.org/doi/10.1137/S1064827500370883)
+ exposed as `torch.lobpcg(A, B=None, k=1, method="ortho", ...)`
+ [x] dense matrices
+ [x] batches of dense matrices
+ [x] sparse matrices
+ [x] documentation
- [x] generalized eigenvalue solver using the robust and efficient LOBPCG Algorithm 8 from [Duersch et al, 2018](https://epubs.siam.org/doi/abs/10.1137/17M1129830) that switches to orthogonal basis selection automatically
+ the "ortho" method improves iterations so rapidly that in the current test cases it does not make sense to use the basic iterations at all. If users will have matrices for which basic iterations could improve convergence then the `tracker` argument allows breaking the iteration process at user choice so that the user can switch to the orthogonal basis selection if needed. In conclusion, there is no need to implement Algorithm 8 at this point.
- [x] benchmarks
+ [x] `torch.svd` vs `torch.svd_lowrank`, see notebook [Low-rank SVD](https://github.com/Quansight/pearu-sandbox/blob/master/pytorch/Low-rank%20SVD.ipynb). In conclusion, the low-rank SVD is going to be useful only for large sparse matrices where the full-rank SVD will fail due to memory limitations.
+ [x] `torch.lobpcg` vs `scipy.sparse.linalg.lobpcg`, see notebook [LOBPCG - pytorch vs scipy](https://github.com/Quansight/pearu-sandbox/blob/master/pytorch/LOBPCG%20-%20pytorch%20vs%20scipy.ipynb). In conculsion, both implementations give the same results (up to numerical errors from different methods), scipy lobpcg implementation is generally faster.
+ [x] On very small tolerance cases, `torch.lobpcg` is more robust than `scipy.sparse.linalg.lobpcg` (see `test_lobpcg_scipy` results)
Resolves https://github.com/pytorch/pytorch/issues/8049.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29488
Differential Revision: D20193196
Pulled By: vincentqb
fbshipit-source-id: 78a4879912424595e6ea95a95e483a37487a907e
Summary:
See NumPy's division documentation here: https://numpy.org/doc/1.18/reference/generated/numpy.divide.html#numpy.divide.
True division is the same as PyTorch's default division except when both inputs are integer or bool tensors. In the latter case the inputs are (conceptually) cast to the default floating type before the division is performed.
The function is implemented for dense and sparse tensors and supports exporting to ONNX from PyTorch's eager mode or JIT traces. The function is inherently incompatible with exporting to ONNX via JIT script, and is another datapoint suggesting we should deprecate exporting scripted graphs to ONNX.
Tests are added for the type promotion, named tensor, and ONNX export behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34236
Reviewed By: houseroad
Differential Revision: D20334087
Pulled By: mruberry
fbshipit-source-id: 83d00d886f46f713215d7d9e02ffd043164c57f1
