**Background:** It's common to use `scalar_tensor()` in the input to `where()` to convert any scalars present to compatible tensors with matching options, *including layout*. This shows up in various places, notably including derivative formulas ([example](78491d6afc/tools/autograd/derivatives.yaml (L432-L434))). It causes problems for NJTs because they have `layout=torch.jagged` and it never makes sense to create a scalar tensor with this layout. Some of the breakage only seems to happen in CI for reasons I don't fully understand (see the revert of #140736 due to softshrink's derivative formula).
**This PR:**
* Allows non-contiguous NJT inputs to `where()` + adds tests for this
* Handles scalar tensor / dense tensor inputs for `condition` / `other` + adds tests for this
* Uses limited `broadcast_tensors()` / `broadcast_to()` support
* Improves `expand()` to work on non-contig NJTs
* Changes `scalar_tensor()` to use `torch.strided` instead of `torch.jagged` in both eager and torch.compile (i.e. meta registration)
* Changes backward formulas for `sinc`, `pow`, `special.i1`, and `special.i1e` to uses `scalar_tensor()` instead of e.g. `zeros({})`
**Alternative approach:** Update all problematic usages of `scalar_tensor()` to avoid ever passing `layout=torch.jagged`. This is an extensive change and includes `torch.where()` logic, a bunch of derivative formulas, and likely other places not yet discovered.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/141500
Approved by: https://github.com/malfet, https://github.com/cpuhrsch, https://github.com/soulitzer
# UPDATE:
This is take 3 of https://github.com/pytorch/pytorch/pull/131863 which was landed via co dev but not applying correclty
# Summary
Changes the stance of SDPA on what to do for fully masked out rows
## Current Behavior
Several PyTorch users have expressed frustration over this issue:
- https://github.com/pytorch/pytorch/issues/41508
- https://github.com/pytorch/pytorch/issues/103749
- https://github.com/pytorch/pytorch/issues/103963
These are significant issues with extensive discussion but no satisfactory resolution. The PyTorch team's consensus, as stated here:
https://github.com/pytorch/pytorch/issues/24816#issuecomment-524415617
Can be paraphrased as follows:
When passing in fully masked out rows, attention becomes ambiguous. We have two main options:
1. Uniformly attend to all values:
```python
scores[masked_out_rows] = 1 / len(row)
out[masked_out_rows] = 1 / len(row) * value
```
2. Decide that attention between no queries (masked) and no keys (masked) is meaningless:
```python
output[fully_masked_rows] = NaN
```
We went with option 2. Partially because it was easier to implement, but also people argued that users can slice the output to remove the NaNs:
``` Python
>fill_value = -float("inf")
>row0 = torch.randn(4)
>row1 = torch.tensor([(fill_value for _ in range(4)])
>matrix = torch.stack([row0, row1]).requires_grad_(True)
>out = torch.softmax(matrix, 1)
>out = out[0]
>print(out)
tensor([0.5377, 0.2729, 0.0692, 0.1201])
```
Cool, problem solved. But what happends when you call backwards..
```Python
>out.backward(torch.ones_like(out))
>print(matrix.grad)
tensor([[3.0957e-08, 1.4157e-08, 7.7802e-10, 1.3713e-08],
[ nan, nan, nan, nan]])
```
Those pesky NaNs are back!
## Why do we see NaNs today?
The core of the problem revolves around using softmax function in sdpa:
```python
> row = torch.tensor([(-float("inf")) for _ in range(4)])
> torch.softmax(row, 0)
tensor([nan, nan, nan, nan])
```
## Quick Aside: Masking in Attention
Attention itself doesn't have a concept of masking. The `sdpa` function has an argument called `attn_mask`, which would be more accurately named `attn_bias`. This is because we don't actually "mask" entries when computing attention. Instead, due to implementation details([performance](https://github.com/pytorch/pytorch/issues/25110#issuecomment-524519087)), we add a value to the masked-out query/key pairs.
We use a large negative number (typically -inf) to decrease the attention weight, as softmax assigns more weight to larger values.
## Alternative Approaches
If we use a very large negative number instead of -inf:
```python
> row = torch.tensor([(-1e6) for _ in range(4)])
> torch.softmax(row, 0)
tensor([0.2500, 0.2500, 0.2500, 0.2500])
```
However if users always remembered to "slice" out their outputs i.e.:
```Python
>fill_value = -1e6
>...
>out.backward(torch.ones_like(out))
>print(matrix.grad)
tensor([[-0.0563, -0.0564, 0.1613, -0.0486],
[ 0.0000, 0.0000, 0.0000, 0.0000]])
```
This would bring us back into a better state.
## A Third Option
We don't necessarily need to alter the behavior of softmax for -inf or very large negative numbers. The fundamental goal is to exclude certain query/key pairs from attention, regardless of the underlying implementation.
This PR implements the new semantic for masking w/ attention in fully masked-out rows:
```python
out[masked_out_rows] = 0
```
**Important Note**: This idea isn't entirely new. The [MaskedTensor](https://pytorch.org/tutorials/prototype/maskedtensor_overview#safe-softmax) prototype, a tensor subclass, was designed to handle such cases. However, it remains a prototype feature and hasn't gained widespread adoption.
## Details
This PR stack does 3 things:
1. Adds a PRIVATE _safe_softmax op
2. Updates semantic for flash_cpu fused kernel
3. Updates semantic for efficient_cuda fused kernel
_safe_softmax is not supposed to be used generically and is only meant to be used within the context of SDPA. Due to this fact instead of decomposing softmax and checking for -inf rows we instead "cheat" and use nan_to_num.
Why I think this is okay? (please find a counter point if avail)
There are multiple ways NaNs can emerge. For the fully masked out rows case nan_to_num works. But what if there were other NaNs, wouldn't this silently remove them?
The only case that this can happen is if the input itself had a NaN or an Inf
For example:
```Python
a = torch.ones([4], requires_grad=False, dtype=torch.float16)
a[1] = torch.finfo(torch.float16).max
print(a.softmax(-1))
```
Will return
`tensor([0., 1., 0., 0.], dtype=torch.float16)`
Where
```Python
a = torch.ones([4], requires_grad=False, dtype=torch.float16)
a[1] = float("inf")
a.softmax(-1)
```
returns:
`tensor([nan, nan, nan, nan], dtype=torch.float16)`
If we dont want to even allow for the possibility of "inf" or "NaN" attention scores to be converted to 0 then we can implemented it something like this
```Python
max = torch.max(a, dim=-1, keepdim=True)
exp = torch.exp(a - max.values)
denom = torch.sum(exp, dim=-1, keepdim=True)
softmax = exp / denom
softmax = torch.where(max.values == float('-inf'), 0.0, softmax)
```
however we would be paying for this in math performance.
## Why Now
I think one point that has substantially changed where PyTorch should lie on this argument is the fact that we have fused implementations for SDPA now. And these fused implementations allow us to easily and performantly support this new semantic.
Differential Revision: [D61418679](https://our.internmc.facebook.com/intern/diff/D61418679)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133882
Approved by: https://github.com/soulitzer
mvlgamma backward trips DEBUG=1 asserts when trying to construct an empty tensor with `layout=torch.jagged`. This happens due to passing `self.options()` to `arange()` in `mvlgamma_backward()`. Fix in this PR unconditionally constructs `arange()` with the strided layout.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/132422
Approved by: https://github.com/albanD
This PR fixes an issue presented when calling `aten.alias(int)` raises a TypeError.
```python
import torch
import torch.autograd.forward_ad as fwAD
def f(x):
return 4312491 * x
device = "cpu"
with torch._subclasses.fake_tensor.FakeTensorMode():
with fwAD.dual_level():
x = torch.randn(3, device=device)
y = torch.ones_like(x)
dual = fwAD.make_dual(x, y)
f(dual)
```
The test case above illustrates this bug.
1) `4312491` turns into a tensor that is a wrapped number
2) Forward mode AD calls `aten::alias` internally
3) The wrapped number (`4312491`) becomes a python integer
4) `aten.alias(int)` raises a `TypeError`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124774
Approved by: https://github.com/albanD, https://github.com/zou3519
Fixes https://github.com/pytorch/pytorch/issues/104505
I was originally going to ban all usages of as_strided + mutation in functionalization. But I'm pretty sure that as_strided + mutation is fine when we are calling as_strided on a base tensor.
So in this PR I added a slightly more conservative check: if we see an as_strided + mutation, where the input to an as_strided was **another** view op, then I error loudly in functionalization and link to the github issue above (in case anyone runs into this in the real world)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122502
Approved by: https://github.com/ezyang, https://github.com/albanD
Everything inside the `AT_DISPATCH` block is being compiled 5 times,
so it makes sense to limit it to the only line that uses `scalar_t` which is
the `numeric_limits` query.
Also a small optimization, instead of computing `grad.log()` and `(-grad).log()`
we can compute `grad.abs().log()` which is 2 pointwise ops instead of 3.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119397
Approved by: https://github.com/lezcano, https://github.com/albanD
masked_scatter_backward was previously implemented as a
CompositeExplicitAutograd, which involved a decomp that calls
masked_select, and masked_select in general produces data-dependent
shapes that inductor doesn't support. But masked_scatter_backward
reshapes the return value of masked_select such that the end result has
a static shape again.
I have converted masked_scatter_backward into an aten op to avoid this
issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109642
Approved by: https://github.com/ezyang
ghstack dependencies: #108170
Repeats #106429 for scatter_reduce so that the backward will pass for PT2. The .item() call is only needed to make double-backward work, which isn't supported anyway for PT2; so an easy fix is to just skip the .item() call if we know we won't need double-backward.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107353
Approved by: https://github.com/eellison
