Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18832
ghimport-source-id: fde4ad90541ba52dfa02bdd83466f17e6541e535
Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18833 [STACK] Cache device on TensorImpl; clean up TensorImpl constructors.
* **#18832 [STACK] Disallow changing the device of a tensor via set_.**
* #18831 [STACK] Stop swapping in Storages of the wrong device for Tensors.
This is necessary to cache the device on a TensorImpl.
Differential Revision: D14766231
fbshipit-source-id: bba61634b2d6252ac0697b96033c9eea680956e8
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18831
ghimport-source-id: 2741e0d70ebe2c2217572c3af54ddd9d2047e342
Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18833 [STACK] Cache device on TensorImpl; clean up TensorImpl constructors.
* #18832 [STACK] Disallow changing the device of a tensor via set_.
* **#18831 [STACK] Stop swapping in Storages of the wrong device for Tensors.**
This is necessary to support device caching, see https://github.com/pytorch/pytorch/pull/18751 and https://github.com/pytorch/pytorch/pull/18578.
In library code, we potentially swap in Storages with the wrong device when device_guard is False. This happens as follows with "view-like" operations.
1) We allocate a tensor on the 'wrong' device (because device_guard is false).
2) We swap out the 'wrong' storage with the 'right' storage using e.g. THCTensor_setStorage.
Instead, we can just construct the Tensor with the correct Storage from the beginning. This is what we do with 'view'.
Note there are two other "view-like" cases where this happens:
1) unfold
2) set_()
Because these aren't performance critical, I just added the device_guard instead of applying the above correction.
For completeness, this also includes a test that all `device_guard: false` functions behave properly under these conditions.
Reviewed By: dzhulgakov
Differential Revision: D14766232
fbshipit-source-id: 0865c3ddae3f415df5da7a9869b1ea9f210e81bc
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17991
changes:
-Breaks bc: Tensor::type() now returns DeprecatedTypeProperties& rather than Type&.
-Added DeprecatedTypeProperties, it serves as a temporary replacement for Type as the return value of Tensor::type(). This contributes to making Type just for dispatch purposes so that we can make it dtype agnostic.
-Tensor::dispatch_type() now returns Type& like Tensor::type() used to do.
-Changed callsites of Tensor::type() appropriately.
Reviewed By: ezyang
Differential Revision: D14443117
fbshipit-source-id: 239ccb7a09626279a71d1a37f8f82e7f57bf7d9e
Summary:
DLPack can have non-strided tensors, which is represented by a nullptr in the place of dl_tensor.strides.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18510
Differential Revision: D14647328
Pulled By: bwasti
fbshipit-source-id: 5364282810a5772cfc2319fc8133fe86fdd84dd1
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18838
It turns out that we don't have shape inference function of `Split` op at all. This diff adds that.
Reviewed By: bertmaher
Differential Revision: D14766871
fbshipit-source-id: 535cb4f24bdada603c76579e00e7a39aee93e19f
Summary:
Since parameter.data will create a new torch.Tensor each time, we get duplicate tensors when call _unique_state_dict now. Try to deduplicate it before creating new tensor.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18139
Reviewed By: dzhulgakov
Differential Revision: D14511262
Pulled By: houseroad
fbshipit-source-id: cb69795d0b6509721220650bbb19edeb3459a503
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17026
D14013931 was for FC. This diff is similar optimizations for Conv.
A subtle difference is that in FC, once we fold col_offset into bias during pre-processing step, we can treat everything as if A_zero_offset == 0 (symmetric quantization of A).
In Conv, we can't do this because padding still needs to use the original A_zero_offset.
From requantization point of view, once col_offset folded into bias, we can treat as if we're doing symmetric A quantization.
But, for steps involving padding like im2col, im2col fused with packing, and direct conv for depth-wise/group convolution we still need to pass the original A_zero_offset.
Reviewed By: jianyuh
Differential Revision: D14020276
fbshipit-source-id: c29caefd1127bbc6aff0e9d535939bb0c1ecb66c
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18826
ghimport-source-id: 7ffa3bc7ef7402a6d6eb6ba5849e197019d77bf8
Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18826 [jit] run cpp tests for non-cuda builds in test_jit.py**
We did all the work of nicely separating our cpp tests that don't require
CUDA, but they aren't run from test_jit.py if CUDA is missing.
Reviewed By: ZolotukhinM
Differential Revision: D14766287
fbshipit-source-id: 9326b3a5c90f6c20fc8cfaf1a1885a363b91f30a
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18379
ghimport-source-id: 9895ecc1ff7897e98853dc00675341f36726e7c7
Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18379 Enforce single parent for script submodules**
* #18378 Unify namespace of script::Module
* #18314 Add ability to specialize class types to ArgumentSpec
* #18226 Add Slot type to abstract the raw pointers being used for slots.
The assumption that a ScriptModule has a single parent is present in
our serialization format, and likely a few other places. It is not
enforced on creation of script module hierarchies though, meaning that
problems associated with (e.g. replicating a module twice in the output
format) will not be caught until much later in the development cycle.
This patch enforces the property when a submodule is registered.
It also removes NamedModule since it is no longer necessary in this regime.
This will also allow the easy discover of a modules fully-qualified name
without needing to traverse the Module hierarchy.
Differential Revision: D14603722
fbshipit-source-id: 63ab5d0cccf7d66c7833e0adf9023024ca9607cb
Summary:
Per our offline discussion, allow Tensors, ints, and floats to be casted to be bool when used in a conditional
Fix for https://github.com/pytorch/pytorch/issues/18381
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18755
Reviewed By: driazati
Differential Revision: D14752476
Pulled By: eellison
fbshipit-source-id: 149960c92afcf7e4cc4997bccc57f4e911118ff1
Summary:
Fix the layernorm formula when weight and bias passed in.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18233
Differential Revision: D14760375
Pulled By: wanchaol
fbshipit-source-id: d6bd3b137bc04c391aa5c24d021d1f811ba2a877
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18378
ghimport-source-id: 55c29bb436a2153d29ff2f4488d99d8863c187b1
Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18379 Enforce single parent for script submodules
* **#18378 Unify namespace of script::Module**
* #18314 Add ability to specialize class types to ArgumentSpec
* #18226 Add Slot type to abstract the raw pointers being used for slots.
This removes individual OrderedDicts in favor of a single unified
namespace for all things in a script::Module. This removes a whole
class of bugs where both a method and an parameter could get the
same name, for instance.
Since we no longer have to expose OrderedDict::Item objects, a lot of
downstream code can be simplified.
We no longer now double-store names (both in the key of the dictionary,
and in the object itself).
Differential Revision: D14603723
fbshipit-source-id: b5f7551b3074679623edd6ea70269830353b4d4c
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18648
ghimport-source-id: 1cf4a8fe91492621e02217f38cae5d7e0699fb05
Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18661 Step 7: remove _unique
* #18655 Step 6: Rename _unique2 to unique and add int? dim
* #18654 Step 5: remove _unque_dim in favor of unique_dim
* #18651 Step 4: add support for unique with dim=None
* #18650 Step 3: Add support for return_counts to torch.unique for dim not None
* #18649 Step 2: Rename _unique_dim2_temporary_will_remove_soon to unique_dim
* **#18648 Step 1: Secretly add return_counts to unique, and refactor unique_dim for performance**
`unique` is fragile, previously I tried to change it in #18391 and #17097, they all pass OSS tests but finally get reverted due to internal failure. My previous work of refactoring unique #18459 is based on #18391, and after #18391 get reverted, I could not work on #18459. To continue working on #18459, #18391, and #17097 without worrying about internal failures, I am suggesting the following steps for the improvements of `unique` and `unique_dim`. soumith Please take this and there is no need to put #18391 back.
The motivation is basically to move forward as much as possible without causing any internal failures. So I will try to divide it into steps and sort from low probability of internal failure to high probability. (I don't know what the internal failure is, so I have to guess). Let's merge these PR stack one by one until we enounter internal failure.
Step 1: Create two new ATen operators, `_unique2_temporary_will_remove_soon` and `_unique_dim2_temporary_will_remove_soon` and keep `_unique` and `_unique_dim` unchanged. The backend of these two functions and `_unique` and `_unique_dim` are all the same, the only difference is the temporary ones support `return_counts` but not the `_unique` and `_unique_dim`. Step one is mostly #18391 + #18459. The cuda8 errors has been fixed. At this point, there is no user visible API change, so no docs are updated. `torch.unique` does not support `return_counts` yet, and `return_counts` is tested through the newly added temporary operators. This step just added two new ATen operators, so there shouldn't be any internal failure.
Step 2: Rename `_unique_dim2_temporary_will_remove_soon` to `unique_dim`. This should cause no internal failure either, because no change to existing operators. The only thing to worry about is to delete `unique_dim` from python side because we don't want users to use it. At this point, C++ users now have `return_counts` support for `unique_dim`.
Step 3: Update the docs of `torch.unique` and use `unique_dim` inside `torch.unique` to support `return_counts` In the docs, we should say `torch.unique` with None dim support does not support `return_counts` yet. This might cause internal failure.
Step 4: Rename `_unique2_temporary_will_remove_soon` to `_unique2` and use `_unique2` inside `torch.unique` to support `return_counts`. Update the docs saying that `torch.unique` with None dim now support `return_counts`. This might cause internal failure.
Step 5: Remove `_unique_dim`. This might cause internal failure.
Step 6: Rename `_unique2` to `unique`, add optional `dim` argument to make it looks like the signature of Python's `torch.unique`. Inside `torch.unique`, use `unique` and get rid of `unique_dim`. Unbind `unique_dim` totally from Python at codegen. This is likely to cause internal fail.
Step 7: Remove `_unique`. This is very likely to cause internal failure.
This PR
======
This PR is for step 1. This create two new ATen operators, `_unique2_temporary_will_remove_soon` and `_unique_dim2_temporary_will_remove_soon` and implement `return_counts` inside them and do refactor for performance improvements.
Please review ngimel VitalyFedyunin. They are mostly copied from #18391 and #18459, so the review should be easy.
Below is a benchmark on a tensor of shape `torch.Size([15320, 2])`:
Before
---------
```python
print(torch.__version__)
%timeit a.unique(dim=0, sorted=True, return_inverse=False); torch.cuda.synchronize()
%timeit a.unique(dim=0, sorted=True, return_inverse=True); torch.cuda.synchronize()
```
```
1.0.1
192 µs ± 1.61 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
548 ms ± 3.39 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
```
```python
print(torch.__version__)
%timeit a.unique(sorted=True, return_inverse=False); torch.cuda.synchronize()
%timeit a.unique(sorted=True, return_inverse=True); torch.cuda.synchronize()
```
```
1.0.1
226 µs ± 929 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each)
302 µs ± 7.06 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
```
After
-------
```python
print(torch.__version__)
%timeit a.unique(dim=0, sorted=True, return_inverse=False); torch.cuda.synchronize()
%timeit a.unique(dim=0, sorted=True, return_inverse=True); torch.cuda.synchronize()
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted=True, return_inverse=False, return_counts=True); torch.cuda.synchronize()
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted=True, return_inverse=True, return_counts=True); torch.cuda.synchronize()
```
```
1.1.0a0+83ab8ac
190 µs ± 2.14 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
237 µs ± 1.23 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
219 µs ± 2.3 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
263 µs ± 1.15 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
```
```python
print(torch.__version__)
%timeit a.unique(sorted=True, return_inverse=False); torch.cuda.synchronize()
%timeit a.unique(sorted=True, return_inverse=True); torch.cuda.synchronize()
%timeit torch._unique2_temporary_will_remove_soon(a, sorted=True, return_inverse=False, return_counts=True); torch.cuda.synchronize()
%timeit torch._unique2_temporary_will_remove_soon(a, sorted=True, return_inverse=True, return_counts=True); torch.cuda.synchronize()
```
```
1.1.0a0+83ab8ac
232 µs ± 2.21 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
301 µs ± 1.65 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
264 µs ± 7.67 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
339 µs ± 9.2 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
```
Differential Revision: D14730905
fbshipit-source-id: 10026b4b98628a8565cc28a13317d29adf1225cc
Summary:
If the input `network` resides on multiple GPUs, `devices` must be a 2D list with `devices[0]` matching `network`'s devices. See #18591
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18687
Differential Revision: D14706162
Pulled By: mrshenli
fbshipit-source-id: dca630d3308f2dbcf8b75629c452d7a64092ba42
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18230
Implementing minimum qtensor API to unblock other workstreams in quantization
Changes:
- Added Quantizer which represents different quantization schemes
- Added qint8 as a data type for QTensor
- Added a new ScalarType QInt8
- Added QTensorImpl for QTensor
- Added following user facing APIs
- quantize_linear(scale, zero_point)
- dequantize()
- q_scale()
- q_zero_point()
Reviewed By: dzhulgakov
Differential Revision: D14524641
fbshipit-source-id: c1c0ae0978fb500d47cdb23fb15b747773429e6c
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18560
We have to import python protobuf here **before** we load cpp extension.
Otherwise it breaks under certain build conditions if cpp implementation of
protobuf is used. Presumably there's some registry in protobuf library and
python side has to initialize the dictionary first, before static
initialization in python extension does so. Otherwise, duplicated protobuf
descriptors will be created and it can lead to obscure errors like
Parameter to MergeFrom() must be instance of same class: expected caffe2.NetDef got caffe2.NetDef.
I think it also fixes https://github.com/facebookarchive/caffe2/issues/1573
Reviewed By: ezyang, iroot900
Differential Revision: D14622054
fbshipit-source-id: 2499eb88ecdee85ff8d845859048f7ae5da2a480
Summary:
to make test_operators.py more stable. in future, we will bump this up manually, and I think it's acceptable, since ir_version should be bumped too often.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18768
Reviewed By: zrphercule
Differential Revision: D14741514
Pulled By: houseroad
fbshipit-source-id: 0369dbc55424e345a113e49fc104a441ea290d58
Summary:
Introduce this check to see whether it will break any existing workflow
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18145
Reviewed By: dzhulgakov
Differential Revision: D14511711
Pulled By: houseroad
fbshipit-source-id: a7bb6ac84c9133fe94d3fe2f1a8566faed14a136
Summary:
The mkldnn-bridge is upgraded in this PR to support DNNLOWP operators.
Meanwhile, APIs have been updated in caffe2 to use latest version.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16308
Differential Revision: D14697018
Pulled By: yinghai
fbshipit-source-id: ca952589098accb08295fd5aa92924c61e74d69c
Summary:
Fixes : #6469
1. `ATen/native/native_functions.yml` had [dispatch](03e7953a98/aten/src/ATen/native/native_functions.yaml (L451-L455)) variants for for `embedding_dense_backward` , however `embedding_backward` explicitly made [call](03e7953a98/aten/src/ATen/native/Embedding.cpp (L35-L45)) to it, thus leading to error.
2. In case of CUDA type tensor, the function crashed used to crash on dereferencing of indices's data [pointer](03e7953a98/aten/src/ATen/native/Embedding.cpp (L93)).
Both have been solved and checked against (on CUDA and CPU)
1. As mentioned in the issue
```
import torch
class Test(torch.nn.Module):
def __init__(self):
super(Test,self).__init__()
self.embd = torch.nn.Embedding(1000, 100)
self.dense = torch.nn.Linear(100, 1)
def forward(self, inp):
inp = self.embd(inp)
return self.dense(inp)
test = Test()
inp = torch.tensor([0,1,2,1,1])
out = test(inp)
raw_loss = out.mean(dim=0)
loss_grad = torch.autograd.grad(outputs=raw_loss,
inputs=list(test.parameters()),
retain_graph=True, create_graph=True, only_inputs=True)
norm = sum([param.norm()**2 for param in loss_grad])
loss = raw_loss + norm
loss.backward(retain_graph=True)
print(test.embd.weight.grad)
```
2. Test Script
```
import torch
import time
start = time.time()
l = [1,1]*100
input = torch.tensor([[1,0],[1,0]],device='cpu')
embedding_matrix = torch.tensor([[1.0,3.0],[2.0,4]],requires_grad=True,device='cpu')
sq = embedding_matrix * embedding_matrix
emb = torch.nn.functional.embedding(input, sq,scale_grad_by_freq=False)
print('Embedding Matrix')
print(embedding_matrix)
print('-----------------')
sum_ = emb.sum()#prod.sum()
loss_grad, = torch.autograd.grad(outputs=sum_,inputs=embedding_matrix,create_graph=True)
print('Gradient')
print(loss_grad)
print('-----------------')
sum2_ = sum_ + loss_grad.sum()
print(sum2_)
sum2_.backward()
print(embedding_matrix.grad)
print(time.time() - start)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9078
Reviewed By: ezyang
Differential Revision: D14691901
Pulled By: soumith
fbshipit-source-id: 78e2612ba39080be564c876311671eb5a0119a0f
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18749
ghimport-source-id: 9026a037f5e11cdb9ccd386f4b6b5768b9c3259b
Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18751 Disallow changing the device of a tensor via set_.
* #18750 Use non-legacy constructors for tensor deserialization.
* **#18749 Add device and dtype to storage.**
The goal here is to fix our serialization, which currently depends on the legacy constructors. Having dtype and device on Storage allows us to use the non-legacy constructors.
This fits somewhat along our goal of removing Storage, my having Storage act like a Tensor.
Differential Revision: D14729516
fbshipit-source-id: bf4a3e8669ad4859931f4a3fa56df605cbc08dcb
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18750
ghimport-source-id: f1475cfb67841c41d9867d4429ba9125d5c7dd07
Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18751 Disallow changing the device of a tensor via set_.
* **#18750 Use non-legacy constructors for tensor deserialization.**
* #18749 Add device and dtype to storage.
Deserialization currently uses legacy constructors. This is bad because we need to maintain them, but there is a more immediate problem:
1) We are trying to implement device caching on TensorImpl to get rid of a virtual dispatch
2) This doesn't work if one is able to change the device of a Tensor underlying a Variable.
3) Deserialization does 2)
So the plan is to change deserialization, then enforce that we don't change the device out from underneath a Variable.
Differential Revision: D14729513
fbshipit-source-id: 090d6cdb375b94dc1bf4f554b2df243952b8cdc6
Summary:
It's not used and unfold's use of `device_guard: False` is scary.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18773
Differential Revision: D14736526
Pulled By: gchanan
fbshipit-source-id: 6281a284bee45fa5038783e4c1ed4d1ed7ca81ab
