Summary:
Original commit changeset: d6d62d0c96dd
Original Phabricator Diff: D84468451 and D84613184
D84468451 caused CUDA OutOfMemoryError in model.
Test Plan:
D84468451 was found through bisect. Also double checked on recent trunk 9866939225248c2adc307be7a804b26db0b9b555: f815887517
With this diff that backs out D84468451 and D84613184 : f816114560
Differential Revision: D85025378
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165910
Approved by: https://github.com/clee2000
This fixes AOTAutograd rms_norm not being bitwise equivalent to
eager, because it avoids a decomposition. You can force the
decomposition by having the decomposition in the dispatch table,
but if eager mode wouldn't have decomposed (because it went to the fused
one), we now default to preserving the fused call by default.
This largely reverts https://github.com/pytorch/pytorch/pull/103275/ for view ops. This means that in inference mode we could hit the wrong C++ kernel; if this occurs we should just SymInt'ify the C++ kernel.
Another neat side effect of this change is that Inductor's generated kernels for rms_norm now have rms_norm in their name.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164939
Approved by: https://github.com/bdhirsh
This fixes AOTAutograd rms_norm not being bitwise equivalent to
eager, because it avoids a decomposition. You can force the
decomposition by having the decomposition in the dispatch table,
but if eager mode wouldn't have decomposed (because it went to the fused
one), we now default to preserving the fused call by default.
This largely reverts https://github.com/pytorch/pytorch/pull/103275/ for view ops. This means that in inference mode we could hit the wrong C++ kernel; if this occurs we should just SymInt'ify the C++ kernel.
Another neat side effect of this change is that Inductor's generated kernels for rms_norm now have rms_norm in their name.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164939
Approved by: https://github.com/bdhirsh
This fixes AOTAutograd rms_norm not being bitwise equivalent to
eager, because it avoids a decomposition. You can force the
decomposition by having the decomposition in the dispatch table,
but if eager mode wouldn't have decomposed (because it went to the fused
one), we now default to preserving the fused call by default.
This largely reverts https://github.com/pytorch/pytorch/pull/103275/ for view ops. This means that in inference mode we could hit the wrong C++ kernel; if this occurs we should just SymInt'ify the C++ kernel.
Another neat side effect of this change is that Inductor's generated kernels for rms_norm now have rms_norm in their name.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164939
Approved by: https://github.com/bdhirsh
ghstack dependencies: #164573
For better tracking, we need to make maybe aliasing/mutating ops with proper tag. We need to special case native_batch_norm because it is not a CIA but has a wrong schema. I guess native_batch_norm will be removed at some point, so until then we just keep it around.
D60347117
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131990
Approved by: https://github.com/bdhirsh
In this PR, we implement lazy dictionary for export decomp behaviour for following reasons:
1. Custom op loading can happen after import time, as a result, the decomp table might not be able to pick up the decomp. Therefore we try to delay materialization as late as possible.
I intentionally seperated out the core_aten_decomp to not have any custom CIA ops in this PR to mitigate the risk of getting reverted but in the future, core_aten_decomp under torch/_decomp will exist as an alias to official export table (torch.export.default_decompositions)
Differential Revision: [D64140807](https://our.internmc.facebook.com/intern/diff/D64140807)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/137650
Approved by: https://github.com/justinchuby, https://github.com/bdhirsh
In this PR, we deprecate _preserve_ops feature in run_decomposition API. We can't kill this API completely because Executorch team depends on it. As the syncing between two repos is non-trivial, I just leave this argument as deprecated for now. In the next PR, i will immediately remove it.
After this PR, run_decompositions will only decompose what's inside the decomp table and preserve the rest by default. Note that this feature is only rolled out to OSS for now. Old code path is protected under IS_FBCODE flag.
Differential Revision: [D62163161](https://our.internmc.facebook.com/intern/diff/D62163161/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/135080
Approved by: https://github.com/justinchuby, https://github.com/avikchaudhuri, https://github.com/bdhirsh
When exporting a training model for Executorch (which requires all ops to be core aten) with cross entropy loss (`torch.nn.CrossEntropyLoss`), we ran into the following error from the fx verifier in `to_edge`:
```
torch._export.verifier.SpecViolationError: Operator torch._ops.aten.nll_loss2d_forward.default is not Aten Canonical.
```
The aten [implementation](https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/LossNLL.cpp#L624) of `torch.nn.CrossEntropyLoss` uses `nll_loss2d_forward` for inference and `nll_loss2d_backward` for training, so we need to add the decompositions for both (which already exist) to the list of core aten decompositions.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133534
Approved by: https://github.com/JacobSzwejbka
## Description
Create decomposition of _unsafe_index_put (non-core aten) that turns it into index_put (core aten)
## Testing
Phi3 mini + LoRA model successfully passed `to_edge` after failing due to a non-core aten `unsafe_index_put` getting introduced in a decomposition during joint graph calculations.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133365
Approved by: https://github.com/pianpwk
# 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.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131060
Approved by: https://github.com/jbschlosser
