[feature request] time-distributed layers for application of normal layers to sequence data #1927

erogol · 2017-06-28T12:59:25Z

As we move from images to videos, it seems imperative to feed sequential data into common image layers. However, the problem of dealing sequential data with such layers is not clears on Pytorch. I suggest here to devise something like -time-distributed layers for wrapping normal layers and aligning them for sequential inputs and outputs. I think in the course of time, it is going to be a more important pattern to have in Pytorch. Hope, I explain the problem clearly?

Lately, I solved that problem for 1D feat vectors as follows; I wait some comments to generalize it if necessary.

class TimeDistributed(nn.Module):
    def __init__(self, module):
        super(TimeDistributed, self).__init__()
        self.module = module

    def forward(self, x):
        if len(x.size()) <= 2:
            return self.module(x)
        t, n = x.size(0), x.size(1) 
        # merge batch and seq dimensions
        x_reshape = x.contiguous().view(t * n, x.size(2))
        y = self.module(x_reshape)
        # We have to reshape Y
        y = y.contiguous().view(t, n, y.size()[1])
        return y

pranv · 2017-06-28T18:25:12Z

I have thought about such an API a lot, both with respect to PyTorch and (earlier) with respect to Keras. Here are my thoughts:

What you've proposed in the code here is an abstraction of what seems to be the ubiquitous pattern used when implementing RNNs. The RNN/LSTM implementation always outputs a 3D tensor and if we want to apply softmax over it, we need to apply the same projection for each samples in the batch, at each time-step. There is a lot of utility for a wrapper that takes in the Linear layer and applies it over a 3D tensor. PyTorch's own examples have used this in a couple of places.

But I would argue that this is still a very limited use case and doesn't warrant a generic container. We can have a LinearOverTime layer or even extend the Linear layer to work with 3D inputs (implicitly flattening across one of the dimensions).

However, there are a lot of use cases where a collection of layers have to be applied repeatedly over an axis. For making such a container for this generic case, we have to consider 2 cases:

Stateful Recurrence: We have some state that is carried as we index along a particular axis. This cannot be parallelized, but the container can provide some utilities that manage the state transmissions.
Stateless Recurrence: In this case we can do something similar to what you've proposed -- flatten along some axis (user provided) and do the computation in parallel. (This is for speed only, as stateless recurrence with no state is a strict superset of this).

Common use cases for the first category are attention models, fancy memory augmented networks and novel RNN formulations (ex. Recurrent Highway Networks, Unitary Evolution RNNs etc.,). Common use cases for the second category are the time distributed (or axis distributed) versions of any cascade of layers.

Basically, I am imagining PyTorch's take on theano.scan: http://deeplearning.net/software/theano/library/scan.html . I think it'll be much more user friendly and powerful with PyTorch's dynamic paradigm.

If more people think that this is a good idea, I can work on it :)

8000

jekbradbury · 2017-06-28T23:14:38Z

Here is a generic TimeDistributed mixin (the name for the same idea in Torch7 was nn.Bottle):

class Bottle(nn.Module):
    def forward(self, input):
        if len(input.size()) <= 2:
            return super(Bottle, self).forward(input)
        size = input.size()[:2]
        out = super(Bottle, self).forward(input.view(size[0]*size[1], -1))
        return out.view(*size, -1)
#example:
class Linear(Bottle, nn.Linear):
    pass

That, or a wrapper like the one above, covers your case 2. For your case 1, we've felt so far that using a native Python for or while loop is the most straightforward and clear way to do this. That's how the decoder in OpenNMT-py is implemented, for instance. Basically, Theano and TensorFlow have to introduce a custom operator like scan or while_loop because they can't use native Python control flow -- but we can, and there's no reason not to.

jekbradbury · 2017-06-28T23:16:42Z

Meanwhile this #1935 will make TimeDistributed/Bottle unnecessary for Linear layers.

erogol · 2017-06-30T00:27:54Z

@jekbradbury but my feeling is (not verified) using only tensor operations without explicit loop would be faster. For example, assume that we have 12x8x3x224x224 input s.t. 8 is seq length. Looping means 8 times layer call, otherwise, if we just merge first two dimensions, this is only one pass on GPU (if memory allows you). And the complexity increases as we increase number of time-distributed layers. Correct me I am wrong.

jekbradbury · 2017-06-30T00:41:56Z

Yes, for the operations that can be parallelized across time or other dimensions you should use tensor operations (or implicit tensor operations in the new Linear). But for stateful recurrence you should use a native loop.

t-vi · 2018-06-07T08:59:03Z

Just to add something to this classic issue:

I'd like a functional wrapper, too.
Of course it should work with packed sequences (not hard to do either).
Finally the bikeshedding: Maybe "elementwise" is a more accurate description than time distributed? After all, this throws all timesteps in one giant batch rather than each timestep separately.

I've implemented this a few times now for things like embeddings and more obscure transformations, so I'd love to see this in torch.nn.rnn.utils and as a Module...

If we want this, I'd be very happy to submit a PR.

apaszke · 2018-06-07T09:08:48Z

Wouldn't it be simpler to just implement this as a decorator on the forward?

t-vi · 2018-06-07T09:18:51Z

The functional version is essentially a equivalent to a decorator.
But I'm not sure how I would go using the decorator for an existing class (I used it for Embedding and Bilinear, the latter with two args) without a wrapper or mixin class.
(So for packed sequences I use something like

def elementwise_apply(fn, *args):
    return torch.nn.utils.rnn.PackedSequence(fn(*[(arg.data if type(arg)==torch.nn.utils.rnn.PackedSequence else arg) for arg in args]), args[0].batch_sizes)

for tensors, I would use view(arg.size(0)*arg.size(1), *arg.shape[2:]) or so similar to the Bottle above and back).

…f33179 Summary: Previous import was 0e8d2bc5e51455c70ef790b9f65aa632ed9bc8a7 Included changes: - **[27d4b617](onnx/onnx@27d4b617)**: Adding RoIAlign operator (pytorch#1869) <Sam Pepose> - **[70c9026c](onnx/onnx@70c9026c)**: add ReverseSequence op (pytorch#1927) <Guoliang Hua> - **[ed2db02a](onnx/onnx@ed2db02a)**: README.md: Update badge style for build status (pytorch#1942) <Yulong Wang> - **[e36d3b54](onnx/onnx@e36d3b54)**: Enable python 3.7 in CI for Windows (pytorch#1943) <Raymond Yang> Differential Revision: D15077516 fbshipit-source-id: 8c02c113f1c7da8177a88759809970cd9f72ba00

…f33179 (pytorch#19718) Summary: Pull Request resolved: pytorch#19718 Previous import was 0e8d2bc5e51455c70ef790b9f65aa632ed9bc8a7 Included changes: - **[27d4b617](onnx/onnx@27d4b617)**: Adding RoIAlign operator (pytorch#1869) <Sam Pepose> - **[70c9026c](onnx/onnx@70c9026c)**: add ReverseSequence op (pytorch#1927) <Guoliang Hua> - **[ed2db02a](onnx/onnx@ed2db02a)**: README.md: Update badge style for build status (pytorch#1942) <Yulong Wang> - **[e36d3b54](onnx/onnx@e36d3b54)**: Enable python 3.7 in CI for Windows (pytorch#1943) <Raymond Yang> Differential Revision: D15077516 fbshipit-source-id: c8c6935381ff5a96ab9a4ee519685814f4ea6e59

ily-R · 2019-07-26T13:29:56Z

This worked for me : https://discuss.pytorch.org/t/timedistributed-cnn/51707/2?u=ilyes

gchanan · 2020-03-03T20:13:23Z

@t-vi can you clarify if you still think there is something useful to implement here and if so, what it is exactly? The current description is a bit unclear.

(assigning you just for the answer to this question).

t-vi · 2020-03-03T20:34:10Z

So what the feature request is about is an easy way to take a nn.Module (or a function) that works on some batch x something and apply it for each sequence element in on a seq_len x batch x something or batch x seq_len x something or a packed input.
(See the code above for the functional part.)
I'm not aware that we have that, it seems to come up now and then.

t-vi · 2020-03-03T20:39:37Z

Recent example from the forums: https://discuss.pytorch.org/t/timedistributed-cnn/51707/6
In particular, the code snippet offered in that thread is not the equivalent of Keras's TimeDistributed as it duplicates the layers instead of applying the same to all of them.

apaszke · 2020-03-04T17:59:28Z

Maybe we could just provide a function decorator that does this? Then you'd do:

class MyModule(nn.Module):
    ...

    @torch.utils.ndbatch(expected_dims=(2,)) # forward expects inputs to be [B, X, Y]
    def forward(self, input):
         return ...

ezyang · 2020-05-12T20:41:15Z

It seems like there are a few things going on here.

With respect to the very first post, there are a bunch of operators which don't support arbitrarily many batch dimensions. When an operator supports arbitrarily many batch dimensions, you don't need TimeDistributed, since things will automatically generalize even if you have a batch dimension and a sequence dimension. So one way to make life better here is to just start generalizing operators that only take a single batch dimension to start taking arbitrarily many batch dimensions. (As jekbradbury wrote above, this is indeed what happened Use torch.matmul in nn.Linear #1935) Sometimes you can't conveniently do this though, because that would introduce ambiguity (but generally you do know how many feature dimensions there are, since Modules need to know this up front to allocate parameters anyway)
This is all very well and good for modules contributed by PyTorch core, but PyTorch is bigger than the core library, and there is always a place for something like TimeDistributed for massaging third party code that has been implemented under the assumption of only one batch dimension. Various suggestions involving decorators have been put in this thread, but I think @zou3519's work on vmap will thoroughly put this to rest.

salimkhazem · 2022-09-13T13:10:33Z

Hello everyone, I do not know about your progress on this subject, but we (with @jeremyfix) wrote this version of TimeDistributed which can handle dense time tensors as well as PackedSequences. As other recurrent networks, it can also deal with batch first or time first ; Do you think it could be worth a PR ?

import torch
import torch.nn as nn
from torch.nn.utils.rnn import PackedSequence
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence


class TimeDistributed(nn.Module):
    """
    TimeDistributed for Pytorch which allows to apply a layer to every temporal slice of an input
    Args:
        Module: a Module instance
    PS : Input must be in the shape of (Seq_length, BS, )
    """

    def __init__(self, module, batch_first=False):
        if not isinstance(module, nn.Module):
            raise ValueError(
                "Please initialize `TimeDistributed` with a "
                f"`torch.nn.Module` instance. Received: {module.type}"
            )
        super(TimeDistributed, self).__init__()
        self.module = module
        self.batch_first = batch_first

    def forward(self, x):
        # Input shape is (default)
        #   - Seq_length, BS, *
        # or if batch_first
        #   - BS, Seq_length, *
        # It can also be provided as PackedSequence
        orig_x = x
        if isinstance(x, PackedSequence):
            x, lens_x = pad_packed_sequence(x, batch_first=self.batch_first)

        if self.batch_first:
            # BS, Seq_length, * -> Seq_length, BS, *
            x = x.transpose(0, 1)
        output = torch.stack([self.module(xt) for xt in x], dim=0)
        if self.batch_first:
            # Seq_length, BS, * -> BS, Seq_length, *
            output = output.transpose(0, 1)

        if isinstance(orig_x, PackedSequence):
            output = pack_padded_sequence(output, lens_x, batch_first=self.batch_first)
        return output

7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) [ghstack-poisoned]

7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) ghstack-source-id: 7ebfc36 Pull Request resolved: #87779

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include (TBD): Commits that's in this PR from the devel branch: ``` 7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) ``` [ghstack-poisoned]

7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) ghstack-source-id: f05ebc3 Pull Request resolved: #87779

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include (TBD): Commits that's in this PR from the devel branch: ``` 7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) ``` [ghstack-poisoned]

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: * codegen improvement: i. allow non-root trivial reductions, allow empty/no-op fusion ii. fixes vectorization checks and size calculation iii. bank conflict handle improvement iv. enables transpose scheduler * misc: i. CI tests failure fixes ii. cpp tests file clean up iii. trivial forwarding supports added in codegen runtime iv. added factory methods support in codegen Commits that's in this PR from the devel branch: ``` 7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D40869846](https://our.internmc.facebook.com/intern/diff/D40869846) [ghstack-poisoned]

7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) RUN_TORCHBENCH: nvfuser ghstack-source-id: 3b21376 Pull Request resolved: #87779

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: * codegen improvement: i. allow non-root trivial reductions, allow empty/no-op fusion ii. fixes vectorization checks and size calculation iii. bank conflict handle improvement iv. enables transpose scheduler * misc: i. CI tests failure fixes ii. cpp tests file clean up iii. trivial forwarding supports added in codegen runtime iv. added factory methods support in codegen Commits that's in this PR from the devel branch: ``` 7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D40869846](https://our.internmc.facebook.com/intern/diff/D40869846) [ghstack-poisoned]

7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) RUN_TORCHBENCH: nvfuser ghstack-source-id: 424af8e Pull Request resolved: #87779

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: * codegen improvement: i. allow non-root trivial reductions, allow empty/no-op fusion ii. fixes vectorization checks and size calculation iii. bank conflict handle improvement iv. enables transpose scheduler * misc: i. CI tests failure fixes ii. cpp tests file clean up iii. trivial forwarding supports added in codegen runtime iv. added factory methods support in codegen Commits that's in this PR from the devel branch: ``` 7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D40869846](https://our.internmc.facebook.com/intern/diff/D40869846) [ghstack-poisoned]

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: * codegen improvement: i. allow non-root trivial reductions, allow empty/no-op fusion ii. fixes vectorization checks and size calculation iii. bank conflict handle improvement iv. enables transpose scheduler * misc: i. CI tests failure fixes ii. cpp tests file clean up iii. trivial forwarding supports added in codegen runtime iv. added factory methods support in codegen Commits that's in this PR from the devel branch: ``` 7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D40869846](https://our.internmc.facebook.com/intern/diff/D40869846) Pull Request resolved: #87779 Approved by: https://github.com/davidberard98

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: * codegen improvement: i. allow non-root trivial reductions, allow empty/no-op fusion ii. fixes vectorization checks and size calculation iii. bank conflict handle improvement iv. enables transpose scheduler * misc: i. CI tests failure fixes ii. cpp tests file clean up iii. trivial forwarding supports added in codegen runtime iv. added factory methods support in codegen Commits that's in this PR from the devel branch: ``` 7117a7e patching nvfuser conv cudnn test numerics mismatch (pytorch#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (pytorch#2023) 6ac74d1 Fix sync map (pytorch#2047) f5bca33 Bank conflict checker improvements (pytorch#2032) d2ca7e3 Minor update on cp.async code generation. (pytorch#1901) d36cf61 Test file cleanup (pytorch#2040) 0b8e83f Allow non-root trivial reductions (pytorch#2037) a2dfe40 Fix vectorize size calculation (pytorch#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (pytorch#2025) 197221b removing ci workflow (pytorch#2034) 40e2703 Reduction rand like patch (pytorch#2031) bc77266 Add utility for checking bank conflict of shared memory (pytorch#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (pytorch#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (pytorch#2006)" (pytorch#2024) bca20c1 Cleanup trivial reduction workarounds (pytorch#2006) e4b6585 Trivial forwarding (pytorch#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (pytorch#1991) a4effa6 Enable output allocation cache (pytorch#2010) 35440b7 Patching bn inference (pytorch#2016) 0f9f0b4 Add matmul benchmark (pytorch#2007) 45045cd Enable tests previously disabled due to an aliasing bug (pytorch#2005) 967aa77 Contiguous indexing for View operations (pytorch#1990) a43cb20 Make inlining even more modular (pytorch#2004) dc45835 Test util cleanup (pytorch#2003) 3ca21eb More strict validation (pytorch#2000) a7a7d57 Fix build problem (pytorch#1999) fc235b0 Just fixes comments (pytorch#1998) 482386c cleanup (pytorch#1997) 4cbe0db Improve divisible split detection (pytorch#1970) 42ccc52 Minor build fix. (pytorch#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (pytorch#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (pytorch#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (pytorch#1994) 1d9858c Minor cleanup (pytorch#1992) f262d9c Add support for uniform RNG (pytorch#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (pytorch#1987) 634820c Add support for some empty fusion (pytorch#1981) eabe8d8 Segment self mapping fusions (pytorch#1954) e96aacf Enable Transpose operation (pytorch#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (pytorch#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (pytorch#1969) b1bd32c Minor fix (pytorch#1967) bd93578 Enable transpose scheduler (pytorch#1927) b7a206e Move scheduler vectorize utilities into their own file (pytorch#1959) d9420e4 View scheduling (pytorch#1928) c668e13 Upstream push ci fixes (pytorch#1965) c40202b Fix dump effective bandwidth (pytorch#1962) 93505bc WAR on index mapping when exact and permissive maps differ (pytorch#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (pytorch#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (pytorch#1946) f7bc341 Remove unused variables (pytorch#1955) df3393a Some cleanup (pytorch#1957) 7d1d7c8 TVDomainGuard factory (pytorch#1953) 357ba22 Fill allocation with nan on tests (pytorch#1956) 8eafc54 Fix detection of unmappable root domains (pytorch#1952) 90a51f2 Some indexing cleanups, Add eye support (pytorch#1940) ddc01e4 Exclude unsup 10000 ported data types (pytorch#1951) 992e17c test the groups the same order as they are merged (pytorch#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (pytorch#1941) ac4de38 Merge pull request pytorch#1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (pytorch#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (pytorch#1942) 89330aa Tensor factories must set the output shape as its input (pytorch#1939) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D40869846](https://our.internmc.facebook.com/intern/diff/D40869846) Pull Request resolved: pytorch#87779 Approved by: https://github.com/davidberard98

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: * codegen improvement: i. allow non-root trivial reductions, allow empty/no-op fusion ii. fixes vectorization checks and size calculation iii. bank conflict handle improvement iv. enables transpose scheduler * misc: i. CI tests failure fixes ii. cpp tests file clean up iii. trivial forwarding supports added in codegen runtime iv. added factory methods support in codegen Commits that's in this PR from the devel branch: ``` 7117a7e patching nvfuser conv cudnn test numerics mismatch (pytorch#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (pytorch#2023) 6ac74d1 Fix sync map (pytorch#2047) f5bca33 Bank conflict checker improvements (pytorch#2032) d2ca7e3 Minor update on cp.async code generation. (pytorch#1901) d36cf61 Test file cleanup (pytorch#2040) 0b8e83f Allow non-root trivial reductions (pytorch#2037) a2dfe40 Fix vectorize size calculation (pytorch#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (pytorch#2025) 197221b removing ci workflow (pytorch#2034) 40e2703 Reduction rand like patch (pytorch#2031) bc77266 Add utility for checking bank conflict of shared memory (pytorch#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (pytorch#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (pytorch#2006)" (pytorch#2024) bca20c1 Cleanup trivial reduction workarounds (pytorch#2006) e4b6585 Trivial forwarding (pytorch#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (pytorch#1991) a4effa6 Enable output allocation cache (pytorch#2010) 35440b7 Patching bn inference (pytorch#2016) 0f9f0b4 Add matmul benchmark (pytorch#2007) 45045cd Enable tests previously disabled due to an aliasing bug (pytorch#2005) 967aa77 Contiguous indexing for View operations (pytorch#1990) a43cb20 Make inlining even more modular (pytorch#2004) dc45835 Test util cleanup (pytorch#2003) 3ca21eb More strict validation (pytorch#2000) a7a7d57 Fix build problem (pytorch#1999) fc235b0 Just fixes comments (pytorch#1998) 482386c cleanup (pytorch#1997) 4cbe0db Improve divisible split detection (pytorch#1970) 42ccc52 Minor build fix. (pytorch#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (pytorch#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (pytorch#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (pytorch#1994) 1d9858c Minor cleanup (pytorch#1992) f262d9c Add support for uniform RNG (pytorch#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (pytorch#1987) 634820c Add support for some empty fusion (pytorch#1981) eabe8d8 Segment self mapping fusions (pytorch#1954) e96aacf Enable Transpose operation (pytorch#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (pytorch#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (pytorch#1969) b1bd32c Minor fix (pytorch#1967) bd93578 Enable transpose scheduler (pytorch#1927) b7a206e Move scheduler vectorize utilities into their own file (pytorch#1959) d9420e4 View scheduling (pytorch#1928) c668e13 Upstream push ci fixes (pytorch#1965) c40202b Fix dump effective bandwidth (pytorch#1962) 93505bc WAR on index mapping when exact and permissive maps differ (pytorch#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (pytorch#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (pytorch#1946) f7bc341 Remove unused variables (pytorch#1955) df3393a Some cleanup (pytorch#1957) 7d1d7c8 TVDomainGuard factory (pytorch#1953) 357ba22 Fill allocation with nan on tests (pytorch#1956) 8eafc54 Fix detection of unmappable root domains (pytorch#1952) 90a51f2 Some indexing cleanups, Add eye support (pytorch#1940) ddc01e4 Exclude unsupported data types (pytorch#1951) 992e17c test the groups the same order as they are merged (pytorch#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (pytorch#1941) ac4de38 Merge pull request pytorch#1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (pytorch#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (pytorch#1942) 89330aa Tensor factories must set the output shape as its input (pytorch#1939) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D40869846](https://our.internmc.facebook.com/intern/diff/D40869846) Pull Request resolved: pytorch#87779 Approved by: https://github.com/davidberard98

soumith added the triage review label Jul 13, 2017

houseroad mentioned this issue Apr 25, 2019

Automatic update of fbcode/onnx to 27d4b617e7097cda7d0d4c45ff2b09d248f33179 #19718

Closed

izdeby added needs research We need to decide whether or not this merits inclusion, based on research world and removed triage review labels Jul 29, 2019

colesbury added module: nn Related to torch.nn triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module labels Jul 29, 2019

gchanan assigned t-vi Mar 3, 2020

ezyang removed the needs research We need to decide whether or not this merits inclusion, based on research world label May 12, 2020

jjsjann123 pushed a commit to jjsjann123/pytorch that referenced this issue Oct 4, 2022

Enable transpose scheduler (pytorch#1927)

bd93578

jjsjann123 mentioned this issue Oct 26, 2022

[NVFuser] Upstream push 1026 #87779

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[feature request] time-distributed layers for application of normal layers to sequence data #1927

[feature request] time-distributed layers for application of normal layers to sequence data #1927

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

[feature request] time-distributed layers for application of normal layers to sequence data #1927

[feature request] time-distributed layers for application of normal layers to sequence data #1927

Comments

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!