# Motivation
In backward of per-parameter sharding FSDP, each rank performs reduce scatter to sync gradients across ranks. A rank chunks each gradient tensor into `world_size` slices along the 0-th dimension and concatenate all slices along the 1-th dimension. Gradient tensors will be padded before concatenation when tensor.size(0) % world_size != 0.
### Example 1
Consider `world_size=3` and tensors A (2x4), B (3x3), C (1x2):
Input tensors:
```
AAAA BBB CC
AAAA BBB
BBB
```
Reduce-scatter-copy-in Output:
```
AAAABBBCC
AAAABBB00
0000BBB00
```
### Example 2
Consider `world_size=2` and tensors A (2x4), B (3x3), C(1x2), D(4x2):
Input tensors:
```
AAAA BBB CC DD
AAAA BBB 00 DD
BBB DD
000 DD
```
Reduce-scatter-copy-in first pad:
```
AAAA BBB CC DD
AAAA BBB 00 DD
BBB DD
000 DD
```
Then chunk and cat along dim as the output:
```
AAAABBBBBBCCDDDD
AAAABBB00000DDDD
```
The performance of reduce-scatter-copy-in is critical to per-parameter sharding FSDP. However, reduce-scatter-copy-in via composing existing ATen ops involves `cat` and irregular `pad`, leading redundant data copies and unsatisfactory performance.
# PR
We provide aten native support for reduce-scatter-copy-in, namely `_chunk_cat()`:
```
_chunk_cat(Tensor[] tensors, int dim, int num_chunks) -> Tensor
```
This PR includes the registration of `_chunk_cat` and `_chunk_cat.out`, OpInfo tests, and basic implementation composing existing ATen ops.
In the next PR, we will add the CUDA implementation. Comparing with baselines of composing existing ATen ops, `_chunk_cat()` CUDA implementation improves copy bandwidth from 498 GB/s to 966 GB/s on a production benchmark.
## Requirements on input
1. If input tensors have different ndims, dim should be non-negative and be less than the ndims of every input tensors. If all input tensors have the same ndims, we support both negative and non-negative dim.
2. For wrapped_dim, all tensors should have the same size for 0,...,wrapped_dim-1 dimensions. No requirements for (wrapped_dim, ...)-th dimension.
3. Expect positive num_chunks
4. Expect non-empty input tensor list and each input tensor should have at least 1 element
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121081
Approved by: https://github.com/albanD
We use the fact that we now propagate indexing properly to avoid having
to maintain two different implementations of the op. Doing this we also remove
a spurious guard on this op.
We move the ref into a decomp as we now use advanced indexing.
The only difference we did in the implementation is that we now use
advanced indexing rather than `torch.cat`.
We also remove it from core. Let's see how this goes.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119857
Approved by: https://github.com/peterbell10, https://github.com/larryliu0820
ghstack dependencies: #119863, #119864
Adding an `OpInfo` test for `split_with_sizes_copy` so we can use it to test [CUDA fast path for split_with_sizes_copy.out](https://github.com/pytorch/pytorch/pull/117203). Since the `OpInfo` test doesn't exist yet and introducing it requires modifications to the `CompositeExplicitAutograd` impl, adding the `OpInfo` test in a separate PR to establish a healthy baseline.
Changes made:
- Registered a batching rule for `split_with_sizes_copy`.
- Registered a decomposition for `split_with_sizes_copy`.
- Registered a DTensor prop rule for `split_with_sizes_copy`.
- Added required dtype and device checks to the composite impl.
- Added output resize to the composite impl.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118512
Approved by: https://github.com/albanD
Summary:
A follow up for #117097. In that PR I didn't add
`_scaled_dot_product_attention_for_cpu` into the core_aten_decomposition
table. This PR does that and also add a unit test.
Test Plan: python test/export/test_export.py -k
test_scaled_dot_product_attention
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117390
Approved by: https://github.com/drisspg
## Context
Add decompositions for `aten.max`, `aten.min`, and `aten.var_mean`. These operators follow a pattern of returning a tuple of outputs from two component operators:
```
aten.max(x) -> return aten.amax(x), aten.argmax(x)
aten.min(x) -> return aten.amin(x), aten.argmin(x)
aten.var_mean(x) -> return aten.var(x), aten.mean(x)
```
For `var_mean`, the `refs` implementation was doing something similar, so I changed it to call `torch.` ops instead like was done for other `refs` implementations previously. cc: @peterbell10 @lezcano
Note that Inductor lowers all these directly, so they are excluded from the Inductor decomp table.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110906
Approved by: https://github.com/manuelcandales
Summary:
## Context
Both `aten.sum` and `aten.squeeze` have a "most generic" variant in the form of `aten.sum.dim_IntList` and `aten.squeeze.dims` respectively. Add decompositions for other non generic variants of these operators to express them using the most generic variant.
Note that to register these decomps, the reference implementation under `_refs` had to be removed as registered decompositions. cc: @lezcano @peterbell10
Test Plan: Github CI + Meta Internal CI
Differential Revision: D49965952
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110645
Approved by: https://github.com/peterbell10, https://github.com/digantdesai, https://github.com/manuelcandales
## Context
Add existing decomps for `lift_fresh`, `split.Tensor`, and `unbind` to the core ATen decomposition table. Do not use them in inductor, since Inductor currently lowers these directly.
One note though is that `lift_fresh`'s decomposition has a note saying it's not correct under autograd. However, my understanding is that these decompositions are registered to the `"post_autograd"` decomposition table, meaning autograd wouldn't be a factor. Would like some confirmation that this premise is correct.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110102
Approved by: https://github.com/jansel
## Context
Introduce a core decomposition for `aten.floor_divide` into other `aten` ops, and add it to the core ATen decomposition table.
This replaces the decomposition of `floor_divide` that was used by Inductor. I noticed there was a note on that decomposition
```
# TorchInductor-only decomposition. It should not be taken to core.
# See https://github.com/pytorch/torchdynamo/pull/1120
```
but couldn't discern the reason why this is the case. cc: @lezcano
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110046
Approved by: https://github.com/peterbell10
- Extend `test_torch_dispatch_meta_outplace` to test torch ops that do not have an out parameter but have aten op overloads that have out parameters. Additionally, Python decompositions may register `OpOverloadPacket`'s so decompositions need to be tested to ensure all `OpOverloads` still function for the `Meta` key (e.g. if a python decomposition is registered for an aten op `aten.foo` with overloads `[default, out]`, the python function needs to support receiving out arguments)
- Add out parameter wrappers to python decomps for aten ops that have out overloads
CC. @ezyang @albanD @lezcano
Fixes#107713
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107707
Approved by: https://github.com/lezcano
Python decompositions wrapped by `out_wrapper` need to be unwrapped before compiling with TorchScript since:
- `out_wrapper` extends the decompositions signature with an out parameter, however this `out` parameter is not present in the source code of the original decomposition so the resulting `ScriptFunction` will not have an `out` parameter
- `out_wrapper` is in the `torch._prims_common.wrappers` module so its `globals()` are different to the globals of the decomposition to be wrapped. This may cause symbol resolution to fail with the TorchScript compiler since it is compiling the unwrapped decomps source code rather than the wrapper
The python decomposition for `aten.trace` is wrapped as an example, other decompositions are to be fixed in https://github.com/pytorch/pytorch/pull/107707
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109367
Approved by: https://github.com/lezcano
The "safety" aspect refers to the output not being registered as aliasing the
input, but after AOTAutograd I don't think this distinction matters. However,
we shouldn't use the same decomposition as the safe variant in case the backend
doesn't want to decompose split.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109668
Approved by: https://github.com/lezcano
ghstack dependencies: #109667
