Modify the existing `softmax` operator in PyTorch, invoked by `torch.softmax`, to allow for reductions along the jagged dimension of a nested tensor. The function originally had a basic implementation for reducing along 1 non-ragged dimension. This diff, which uses the aten padding operator, enables PyTorch users to invoke `torch.softmax` on a nested tensor when reducing along the ragged dimension, e.g. `*` in a `(B, *, M)` nested tensor.
Write unit tests based on the `sum` and `mean` jagged operators to verify the accuracy of the ragged reduction implementation for `torch.softmax`. Add unit tests to verify error handling for unsupported features in `NestedTensor` `torch.softmax`.
Note that this implementation is limited to nested tensors with `ragged_idx == 1`, i.e. the ragged dimension is not transposed. In addition, the `softmax` operator is required to take in as input an integer for the reduction dimension `dim`, requiring new unit tests heavily inspired by the `sum` and `mean` jagged operator unit tests. `Softmax` also allows for reducing along the batch dimension.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131518
Approved by: https://github.com/davidberard98
Add support for transposed, non-contiguous `NestedTensor`s, where `ragged_idx > 1`, for the aten operators `sum` and `mean`. This diff enables reducing along the jagged dimension for non-contiguous `NestedTensor`s, transposed between non-batch dimensions as well as between a ragged and a non-batch dimension. For example, users can now reduce a `NestedTensor` of shape `(B, M, *, N)` along `*` or `(B, N, M, *)` along `*`.
Parametrize existing unit tests and add new unit tests verifying the accuracy of implementations on `NestedTensor`s that transpose between 2 non-batch dimensions as well as between a ragged and a non-batch dimension.
Differential Revision: [D59847927](https://our.internmc.facebook.com/intern/diff/D59847927/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131517
Approved by: https://github.com/davidberard98
Add example `NestedTensor`s with inner dimension of size `1` to `_get_example_tensor_lists` with `include_inner_dim_size_1=True`. This diff creates `NestedTensor`s of sizes `(B, *, 1)` and `(B, *, 5, 1)`, ensuring that the current implementations of jagged reductions for `sum` and `mean` hold for tensors of effective shape `(B, *)` and `(B, *, 5)`.
Differential Revision: [D59846023](https://our.internmc.facebook.com/intern/diff/D59846023/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131516
Approved by: https://github.com/davidberard98
Summary:
Modify the existing `mean` operator in PyTorch, invoked by `torch.mean`, to allow for reductions along the jagged dimension of a nested tensor. The function originally had a basic implementation for reducing along 1 non-ragged dimension. This diff enables PyTorch users to invoke `torch.mean` on a nested tensor when reducing along the ragged dimension, e.g. `*` in a `(B, *, M)` nested tensor.
Parametrize unit tests from `sum` to verify the accuracy of the ragged reduction implementation for `torch.mean`. Add unit tests and parametrize `sum` unit tests to verify error handling for unsupported features in `NestedTensor` `torch.mean`.
Test Plan:
Verify that the new unit test passes via the following command:
```
buck2 run mode/{opt,inplace} //caffe2/test:nested -- --regex test_mean
```
```
buck2 run mode/{opt,inplace} //caffe2/test:nested -- --regex test_jagged_op
```
Differential Revision: D59654668
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131132
Approved by: https://github.com/davidberard98, https://github.com/jbschlosser
Summary: Modify the existing `sum` operator in PyTorch, invoked by `torch.sum`, to allow for reductions along the ragged dimension of a nested tensor. This diff enables PyTorch users to invoke `torch.sum` on a nested tensor with `dim=1`, where `ragged_idx=1`.
Functions modified in `caffe2/torch/nested/_internal/ops.py`:
- `sum_dim_IntList()`: The function assumes that `ragged_idx=1`; in the case that `dim=1` as well, where `dim` is the dimension on which we reduce, this diff invokes the PyTorch benchmark found in D58423489. Specifically, this diff pads a nested tensor, e.g. of logical shape `(B, *, M)`, using [`torch.ops.aten._jagged_to_padded_dense_forward`](https://www.internalfb.com/code/fbsource/[92c2a067ab04e3eebc999254fed4ae2fbea6def3]/fbcode/deeplearning/fbgemm/fbgemm_gpu/fb/inductor_lowerings/elementwise_ops.py?lines=26), then reduces across the `*` dimension (`dim == 1`) to a `(B, M)` output tensor.
- `_wrap_jagged_dims()`: This diff adds special handling to allow for the case where `dim` contains `1` and not `0`, but to continue disallowing the case where `dim` contains `0` and not `1`. In this function's creation, I created a helper function, `_get_condition_for_invalid_jagged_reductions()`, which makes it clearer which conditions apply to which operators. Specifically, operators which are enabled with jagged reductions are specified at the top of the file in `SUPPORTED_JAGGED_REDUCTIONS` and have a different set of conditions that need to be tested, as reducing along `dim == 1` without `dim == 0` is now possible.
Functions modified in `caffe2/test/test_nestedtensor.py`:
- `test_sum_int_DimList()`: This diff adds special handling in the `sum` unit test to allow for the case where `dim` contains `1` and not `0`, but to continue disallowing the case where `dim` contains `0` and not `1`.
- `test_sum_int_DimList_ragged_dim_1()`: This diff adds a new unit test which verifies the accuracy and feasibility of reducing along the jagged dimension of a nested tensor.
Notes:
- This diff solely adds functionality for the case in which we reduce only along the ragged dimension. Cases in which we reduce along both the ragged and another dimension, like `dim == (1, 2)`, are not permitted, as this set of diffs focuses primarily on the former.
- The `sum` operator is the only operator which uses the function `_wrap_jagged_dims()`; all other operators use `_wrap_jagged_dim()`. I would like to later look into why this is the case and if we can consolidate this!
- I modified some of the comments in the `sum` function as well as the unit tests for more clarity.
Test Plan:
Verify that existing (`test_sum_int_DimList`) and new (`test_sum_int_DimList_ragged_dim_1`) unit tests pass via the following command:
```
buck2 run mode/{opt,inplace} //caffe2/test:nested -- --regex test_sum_int_DimList
```
Differential Revision: D59571209
Pull Request resolved: https://github.com/pytorch/pytorch/pull/130425
Approved by: https://github.com/davidberard98
This PR updates the public API for NJT construction `torch.nested.nested_tensor_from_jagged()` to accept values for min / max sequence length. It's useful to provide these ahead of time to avoid GPU -> CPU syncs from on-demand computation later on.
NB: The test changes are extensive because I reworked the existing `_validate_nt()` helper function used throughout our NJT construction tests to verify more (specifically: expected cached min / max seq len and contiguity).
API design question: should we additionally provide an option to compute these from `offsets` at construction time? I can think of three possible cases during construction:
1. Min / max seq len has already been obtained from *somewhere* (manual calculation, static values, etc.) and they should be used in the cache
2. Min / max seq len should be computed immediately at construction time for use in the cache (ideally, the caller wouldn't have to do this computation manually)
3. Min / max seq len are not needed at all (i.e. SDPA isn't ever called) and computation should be skipped
Pull Request resolved: https://github.com/pytorch/pytorch/pull/130175
Approved by: https://github.com/davidberard98, https://github.com/soulitzer
In this PR:
- Ensure that if a tensor not requiring grad is saved for backward unpacking does not trigger a detach (unless the user installs a saved tensor pack hook that returns a tensor requiring grad).
- Update non-reentrant checkpoint to also no longer detach for this case.
Alternatives:
- For custom autograd Function, you could directly save on ctx to work around this, but that would not work for when we switch to using custom ops.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127959
Approved by: https://github.com/YuqingJ
ghstack dependencies: #125795, #128545, #129262
This PR does 3 things:
1. Adds a copy-free strided->jagged layout conversion for NT
2. Adds a copy-free jagged->strided layout conversion for NT
3. Modifies and expands the .to() API to support the layout argument for the specific case of NT layout conversion.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115749
Approved by: https://github.com/jbschlosser
Idea: close over min / max sequence length in the main NJT view func (`_nested_view_from_jagged`) so that view replay during fake-ification propagates these correctly in torch.compile.
For dynamic shapes support for min / max sequence length, this PR uses a hack that stores the values in `(val, 0)` shaped tensors.
**NB: This PR changes SDPA to operate on real views instead of using `buffer_from_jagged()` / `ViewNestedFromBuffer`, which may impact the internal FIRST model. That is, it undoes the partial revert from #123215 alongside a fix to the problem that required the partial revert. We need to verify that there are no regressions there before landing.**
Differential Revision: [D55448636](https://our.internmc.facebook.com/intern/diff/D55448636)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122836
Approved by: https://github.com/soulitzer
Idea: close over min / max sequence length in the main NJT view func (`_nested_view_from_jagged`) so that view replay during fake-ification propagates these correctly in torch.compile.
For dynamic shapes support for min / max sequence length, this PR uses a hack that stores the values in `(val, 0)` shaped tensors.
**NB: This PR changes SDPA to operate on real views instead of using `buffer_from_jagged()` / `ViewNestedFromBuffer`, which may impact the internal FIRST model. That is, it undoes the partial revert from #123215 alongside a fix to the problem that required the partial revert. We need to verify that there are no regressions there before landing.**
Differential Revision: [D55448636](https://our.internmc.facebook.com/intern/diff/D55448636)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122836
Approved by: https://github.com/soulitzer
ghstack dependencies: #127007, #128057
This PR introduces naive CPU impls for:
* `_jagged_to_padded_dense_forward()`
* `_padded_dense_to_jagged_forward()`
On the CUDA side, these are backed by lifted FBGEMM kernels. We may want to revisit the CPU versions with higher-performance implementations at a later time.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127007
Approved by: https://github.com/davidberard98
Before: `softmax` definition uses `jagged_unary_pointwise()` (wrong)
After: `softmax` impl adjusts the `dim` arg to account for the difference in dimensionality between the outer NT and the NT's `_values`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119459
Approved by: https://github.com/soulitzer
Summary:
Add the following error checks for the `unbind` operator on `NestedTensor`s when `ragged_idx != 1`:
- The current implementation allows the creation of `NestedTensor` instances from the class definition with an `offsets` tensor that applies to a dimension other than the jagged dimension. This diff ensures that `unbind` fails when the `offsets` exceed the length of the jagged dimension.
Test Plan:
Added the following unit tests:
`test_unbind_with_lengths_ragged_idx_equals_2_bad_dim_cpu` verifies that `unbind` fails when there is a mismatch between the offsets and the jagged dimension, for `NestedTensor`s with `lengths`.
```
test_unbind_with_lengths_ragged_idx_equals_2_bad_dim_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
```
Reviewed By: davidberard98
Differential Revision: D57989082
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128058
Approved by: https://github.com/davidberard98
PyTorch can't depend on `fbgemm_gpu` as a dependency because `fbgemm_gpu` already has a dependency on PyTorch. So this PR copy / pastes kernels from `fbgemm_gpu`:
* `dense_to_jagged_forward()` as CUDA registration for new ATen op `_padded_dense_to_jagged_forward()`
* `jagged_to_padded_dense_forward()` as CUDA registration for new ATen op `_jagged_to_padded_dense_forward()`
CPU impls for these new ATen ops will be added in a follow-up PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125946
Approved by: https://github.com/davidberard98
Summary:
Extend coverage for the `NestedTensor` `unbind` operator to cases in which `ragged_idx != 1`.
Currently, the `unbind` operator in the `NestedTensor` class splits a tensor along the 0-th dimension, where the `ragged_idx` property, which controls the jagged dimension upon which `unbind` splits, is 1. This diff extends support for `ragged_idx != 1` in `NestedTensor`s, allowing `unbind` to split a tensor along a jagged dimension greater than 0 for `NestedTensor`s with and without the `lengths` property.
Test Plan:
Added the following unit tests:
`test_unbind_ragged_idx_equals_2_cpu`, `test_unbind_ragged_idx_equals_3_cpu`, and `test_unbind_ragged_idx_equals_last_dim_cpu` verify that `unbind` works for all jagged dimensions greater than 1, for `NestedTensor`s without `lengths`.
```
test_unbind_ragged_idx_equals_2_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
test_unbind_ragged_idx_equals_3_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
test_unbind_ragged_idx_equals_last_dim_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
```
`test_unbind_with_lengths_cpu` and `test_unbind_with_lengths_ragged_idx_equals_1_cpu` verify that `unbind` works when the jagged dimension is 1, for `NestedTensor`s with `lengths`.
```
test_unbind_with_lengths_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
test_unbind_with_lengths_ragged_idx_equals_1_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
```
`test_unbind_with_lengths_ragged_idx_equals_2_cpu` and `test_unbind_with_lengths_ragged_idx_equals_3_cpu` verify that `unbind` works when the jagged dimension is greater than 1, for `NestedTensor`s with `lengths`.
```
test_unbind_with_lengths_ragged_idx_equals_2_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
test_unbind_with_lengths_ragged_idx_equals_3_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
```
`test_unbind_with_lengths_ragged_idx_equals_0_cpu` verifies that `unbind` fails when the jagged dimension is 0 (the batch dimension), for `NestedTensor`s with `lengths`.
```
test_unbind_with_lengths_ragged_idx_equals_0_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
```
`test_unbind_with_lengths_ragged_idx_equals_2_bad_dim_cpu` verifies that `unbind` fails when there is a mismatch between the offsets and the jagged dimension, for `NestedTensor`s with `lengths`.
```
test_unbind_with_lengths_ragged_idx_equals_2_bad_dim_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
```
`test_unbind_with_wrong_lengths_cpu` verifies that `unbind` fails when the lengths exceed the limitations set by offsets, for `NestedTensor`s with `lengths`.
```
test_unbind_with_wrong_lengths_cpu (test_nestedtensor.TestNestedTensorSubclassCPU) ... ok
```
Differential Revision: D57942686
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127493
Approved by: https://github.com/davidberard98
Fixes#127097
**TL;DR**: dimensions marked with mark_dynamic can result in assertion failures if the marked-dynamic dimensions get specialized. In NJT, we don't care _that_ much that a dimension is marked as dynamic. So instead, mark with `maybe_mark_dynamic` which suggests that a dimension should be dynamic, but doesn't fail if the dimension gets specialized.
**Background**:
NJT marks the values tensor as dynamic:
49ad90349d/torch/nested/_internal/nested_tensor.py (L122)
It does this for two reasons:
1. **Conceptual**: We know that this dimension _should_ be dynamic; it's a nested tensor, so the sequence lengths will _probably_ vary between batches in the common case. Therefore, we should compile it as dynamic to prevent needing a recompile to trigger automatic dynamic shapes.
2. **Implementation detail**: Right now we run into issues with torch.compile / tensor_unflatten / other details when the dimensions are not marked as dynamic. We have some attempts to remove this (e.g. https://github.com/pytorch/pytorch/pull/126563) but while testing this I wasn't able to get all tests to pass, so there could be potential regressions here if we removed the mark_dynamic.
**Justification for this change**
1. **Conceptual**: AFAIK, we don't care enough about the dynamism of this dimension to error out if we specialize. We'd prefer that we don't have to recompile to get automatic dynamic shapes, but it's also better to not have this issue (and not to force the user to go hunt down all the other equivalent shapes to mark them as dynamic as well). This solution allows us to suggest the dynamism but not force it.
2. **Implementation detail**: This still marks the dimension as symbolic at the beginning of dynamo tracing, so we will (probably) avoid a lot of the issues we run into when we completely remove the `mark_dynamic` decorators.
Differential Revision: [D57933779](https://our.internmc.facebook.com/intern/diff/D57933779)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127453
Approved by: https://github.com/soulitzer, https://github.com/YuqingJ
PyTorch can't depend on `fbgemm_gpu` as a dependency because `fbgemm_gpu` already has a dependency on PyTorch. So this PR copy / pastes kernels from `fbgemm_gpu`:
* `dense_to_jagged_forward()` as CUDA registration for new ATen op `_padded_dense_to_jagged_forward()`
* `jagged_to_padded_dense_forward()` as CUDA registration for new ATen op `_jagged_to_padded_dense_forward()`
CPU impls for these new ATen ops will be added in a follow-up PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125946
Approved by: https://github.com/davidberard98
PyTorch can't depend on `fbgemm_gpu` as a dependency because `fbgemm_gpu` already has a dependency on PyTorch. So this PR copy / pastes kernels from `fbgemm_gpu`:
* `dense_to_jagged_forward()` as CUDA registration for new ATen op `_padded_dense_to_jagged_forward()`
* `jagged_to_padded_dense_forward()` as CUDA registration for new ATen op `_jagged_to_padded_dense_forward()`
CPU impls for these new ATen ops will be added in a follow-up PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125946
Approved by: https://github.com/davidberard98
Current tolerances fail on RTX 6000 (Ada) with `Mismatched elements: 2 / 144 (1.4%)`
```
AssertionError: Tensor-likes are not close!
Mismatched elements: 2 / 144 (1.4%)
Greatest absolute difference: 0.002197265625 at index (5, 0, 0) (up to 0.001 allowed)
Greatest relative difference: 0.08203125 at index (3, 0, 0) (up to 0.016 allowed)
To execute this test, run the following from the base repo dir:
python test/test_nestedtensor.py -k test_sdpa_with_packed_in_proj_cuda_bfloat16
This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
----------------------------------------------------------------------
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126356
Approved by: https://github.com/drisspg
Fixes#123698
This PR makes TensorImpl::has_symbolic_sizes_strides return false for NestedTensors.
1. It passes in the actual sizes when we call `_make_wrapper_subclass` - this is the change that makes the subclass register as `has_symbolic_sizes_strides() == True`
2. It adds a field to `_make_wrapper_subclass` where an explicit `numel` can be provided. This allows us to skip the numel computation for the storage, which previously fails due to arithmetic on NestedInts.
3. Implements `aten::numel` for NJT - this is separate from the overridden numel in `make_wrapper_subclass` for now. Note also that this means that we leave `dispatch_sizes_strides_policy="sizes"`, so that we call into the custom `numel` implementation (as well as `sizes` and `strides`), because `numel` cannot currently be computed from `sizes` for NJT.
Note also that this depends on #121361, because calling TensorImpl::set_sizes_and_strides() tries to clone the sizes into the tensor, which means that we need `clone` to be implemented on NestedInt.
Differential Revision: [D57225736](https://our.internmc.facebook.com/intern/diff/D57225736)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124687
Approved by: https://github.com/albanD
Fixes#121758
**TL;DR**: When profiling is turned on, the dispatcher will sometimes attach the autograd sequence number to the recorded profiler event. This PR expands the set of profiler events onto which we attach sequence numbers. Before, we'd only attach a sequence number if the current dispatch key was an Autograd dispatch key. Now, we attach a sequence number if the current dispatch key **set** contains Autograd.
**Context**:
The use case for this is torch.profiler for python subclasses.
Autograd attaches a "sequence number" to all ops that it encounters during the forward pass. Then, the corresponding sequence number can be associated with a backward kernel when backward is executed. This is used by the profiler to associate the forward ops to the backward ops; a forward op and a backward op with the same sequence number are "linked" in some post-processing step.
Prior to this PR, this profiler feature didn't work for python subclasses. The reason is that we don't collect profiler information for all the dispatches for a given kernel; we only dispatch the initial `call`, and not the subsequent `redispatch` invocations. Normally, an Autograd key (if we're running with autograd) is the highest dispatch key, so the initial `call` that we profile is an Autograd key, and we collect the sequence number. But when we're dealing with a python subclass, the first dispatch key is PythonTLSSnapshot, which eventually redispatches into Autograd. We don't record the Autograd sequence number in that case because we don't record redispatches.
To fix this, this PR collects a sequence number whenever the dispatch key **set** contains an Autograd key. That means we might sometimes collect multiple events with the same sequence number, or possibly attach sequence numbers when we won't actually use them? (e.g. maybe if the initial dispatch key handler removes Autograd for some reason). Although this might be a bit confusing for users looking directly at the sequence_nr directly, I think the main use case is for the profiler to create fwd-bwd links; and those should still be generated correctly in these cases.
Differential Revision: [D55724190](https://our.internmc.facebook.com/intern/diff/D55724190)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123304
Approved by: https://github.com/soulitzer
For internal purposes, this PR reverts the use of real views in SDPA -> autograd.Function "views" (i.e. `ViewBufferFromNested` and `ViewNestedFromBuffer`). This is a temporary fix to get the FIRST model launched and working.
**Note: this breaks some other Dynamo tests related to SDPA that rely on real views, but the breakage there isn't expected to be likely in a real-world scenario.**
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123215
Approved by: https://github.com/YuqingJ
For internal purposes, this PR reverts the use of real views in SDPA -> autograd.Function "views" (i.e. `ViewBufferFromNested` and `ViewNestedFromBuffer`). This is a temporary fix to get the FIRST model launched and working.
**Note: this breaks some other Dynamo tests related to SDPA that rely on real views, but the breakage there isn't expected to be likely in a real-world scenario.**
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123215
Approved by: https://github.com/YuqingJ
