# Registers the derivative for mem efficient backward
- Use gradcheck to test correctness. The kernel is not implemented for fp64 so run checks with bumped tolerances in fp32
- I also made updates based off of Xformer main branch and flash-attention cutlass branch.
- This will enable the fused backward to be called for scaled dot product attention
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88856
Approved by: https://github.com/cpuhrsch
Fixes T135842750 (follow-up for #87377)
## Description
At present, having both `src_key_padding_mask` and `src_mask` at the same time is not supported on the fastpath in Transformer and Multi-Head Attention.
This PR enables using both masks on the fastpath on CPU and GPU: if both masks are passed, we merge them into a 4D mask in Python and change mask type to 2 before passing downstream.
Downstream processing in native code is not changed, as it already supports 4D mask. Indeed, it is done depending on the device:
- on CUDA, by `SoftMax.cu::masked_softmax_cuda`. When mask type is 2, it calls either `dispatch_softmax_forward` -> `softmax_warp_forward` or `at::softmax` (depending on the input size). In both cases 4D mask is supported.
- on CPU, by `SoftMax.cpp::masked_softmax_cpp`. It calls `hosted_softmax` which supports 4D mask.
## Tests
- Extended `test_mask_check_fastpath` to check that fast path is indeed taken in Transformer when two masks are passed
- Added `test_multihead_self_attn_two_masks_fast_path_mock` to check that fast path is taken in MHA when two masks are passed
- Added `test_multihead_self_attn_two_masks_fast_path` to check that fast and slow paths give the same result when two masks are passed in MHA
- `test_masked_softmax_mask_types` now covers mask type 2
- `test_transformerencoderlayer_fast_path` (CPU smoke test) is expanded to the case of both masks provided simultaneously
- `test_masked_softmax_devices_parity` checks that mask type 2 is accepted by CPU and CUDA paths
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88488
Approved by: https://github.com/mikekgfb
## Issues
Fixes https://github.com/pytorch/pytorch/issues/81129#issuecomment-1179435674
## Description
Passing a 2D attention mask `src_mask` into the fast path of `TransformerEncoderLayer` in CPU was causing an error and so was disabled in https://github.com/pytorch/pytorch/pull/81277. This PR unrolls this fix, enabling `src_mask` on the fast path:
- Either attention mask `src_mask` of shape `(L, L)` or padding mask `src_key_padding_mask` of shape `(B, L)` are now allowed on the CPU fast path. If softmax is applied along the last dimension (as in multi-head attention), these masks are processed without expanding them to 4D. Instead, when iterating through the input, `Softmax.cpp::host_softmax` converts the index to match the mask dimensions, depending on the type.
- If softmax is applied along the dimension other than the last, `Softmax.cpp::masked_softmax_cpu` expands masks to 4D, converting them to `mask_type=2`. Theoretically one could also add special optimized cases for `dim=0, 1, 2` and process them without mask expansion, but I don't know how often is that used
## Tests:
- `test_transformerencoderlayer_fast_path` is extended to cover both attention mask and padding mask
- `test_masked_softmax_mask_types_0_1` is added to ensure results from CPU softmax with attention and padding masks match the explicit slow calculation
- `test_masked_softmax_devices_parity` is added to ensure results from masked softmax on CPU and CUDA match
## Note
I had to replace `float` with `torch.get_default_dtype()` in a couple of tests for the following reason:
- `test_nn.py` [sets the default type to `torch.double`](https://github.com/pytorch/pytorch/blob/master/test/test_nn.py#L24-L26)
- If I execute `test_nn.py` and `test_transformers.py` in one `pytest` run, this default still holds for transformer tests
- Some tests in `test_transformers.py` which were previously following the slow path now switched to fast path, and hard-coded `float` started clashing with default `double`
Let me know if there is a better way around it - or maybe I'm not supposed to run tests with `pytest` like this
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87377
Approved by: https://github.com/mikekgfb, https://github.com/weiwangmeta, https://github.com/malfet
# Summary
Use the private _scaled_dot_product_attention to support _native_multiheaded_attention. _SDP provides access to fused kernels when certain conditions are meant enabling a speed up for MHA.
cc @cpuhrsch @jbschlosser @bhosmer @mikaylagawarecki
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87312
Approved by: https://github.com/cpuhrsch
# Summary
Add in a torch.backends.cuda flag and update context manager to pic between the three implementations of the scaled_dot_product_attention.
cc @cpuhrsch @jbschlosser @bhosmer @mikaylagawarecki
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87946
Approved by: https://github.com/cpuhrsch
# Summary
- This code creates the runtime dispatch system for choosing a performant fused SDP kernel. The only choice of fused kernel is flash_attention. It also creates python flags and a context manager that can be used to turn off and on behavior for dispatch.
- This also adds support for flash_attention with dense tensors.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85984
Approved by: https://github.com/cpuhrsch
# Summary
This exposes the _scaled_dot_product_attention function to python in the nn namespace. It is still underscored because the api for args, and kwargs is still in flux for the next few weeks and will eventually land as a prototype feature.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85044
Approved by: https://github.com/cpuhrsch
# Summary
This exposes the _scaled_dot_product_attention function to python in the nn namespace. It is still underscored because the api for args, and kwargs is still in flux for the next few weeks and will eventually land as a prototype feature.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85044
Approved by: https://github.com/cpuhrsch
Summary: Check that fastpath is taken, which type (sparsity fastpath or normal) for mask that is aligned and one that is not.
Test Plan: buck test caffe2/test:test_transformers
Differential Revision: D38259928
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82999
Approved by: https://github.com/jbschlosser
Adds an initial private API version of the SDP interface.
Signature:
```
_scaled_dot_product_attention(Tensor query, Tensor key, Tensor value, Tensor? attn_mask=None,
float dropout_p=0.0, bool need_attn_weights=True, bool is_causal=False) -> (Tensor, Tensor)
```
Returns a tuple of `(output, attn_weights)`.
Note the following:
* `need_attn_weights`: flag indicating that attention weights should be computed. This is useful to toggle off for flash attention as it does not materialize the weights by default, making it more expensive to return them.
* Boolean attention mask support only; `True` values within `attn_mask` indicate that the element should take part in attention (notably, this is reverse of MHA, which uses `True` to mask *out* values). Mask is optional.
* `is_causal`: Temporary flag indicating whether to use a causal attention weighting. If this is set to `True`, it takes precedent over any value passed in for `attn_mask`. Longer term, the `is_causal` flagging can be subsumed into the `attn_mask` arg via tensor subclassing (see e.g. [CausalTensor](https://github.com/facebookresearch/xformers/blob/sparse_cleanup/xformers/sparse/causal_tensor.py) in xFormers).
* Testing is currently done via reference with the existing Python impl of `F._scaled_dot_product_attention`.
* This PR does not yet drop-in the new SDP anywhere. A future PR can hook it up in BT or MHA.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81956
Approved by: https://github.com/drisspg, https://github.com/erichan1
Adds an initial private API version of the SDP interface.
Signature:
```
_scaled_dot_product_attention(Tensor query, Tensor key, Tensor value, Tensor? attn_mask=None,
float dropout_p=0.0, bool need_attn_weights=True, bool is_causal=False) -> (Tensor, Tensor)
```
Returns a tuple of `(output, attn_weights)`.
Note the following:
* `need_attn_weights`: flag indicating that attention weights should be computed. This is useful to toggle off for flash attention as it does not materialize the weights by default, making it more expensive to return them.
* Boolean attention mask support only; `True` values within `attn_mask` indicate that the element should take part in attention (notably, this is reverse of MHA, which uses `True` to mask *out* values). Mask is optional.
* `is_causal`: Temporary flag indicating whether to use a causal attention weighting. If this is set to `True`, it takes precedent over any value passed in for `attn_mask`. Longer term, the `is_causal` flagging can be subsumed into the `attn_mask` arg via tensor subclassing (see e.g. [CausalTensor](https://github.com/facebookresearch/xformers/blob/sparse_cleanup/xformers/sparse/causal_tensor.py) in xFormers).
* Testing is currently done via reference with the existing Python impl of `F._scaled_dot_product_attention`.
* This PR does not yet drop-in the new SDP anywhere. A future PR can hook it up in BT or MHA.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81956
Approved by: https://github.com/drisspg, https://github.com/erichan1
Summary:
Add test just to check if TransformerEncoder will crash when enumerating over params [with_no_grad, use_torchscript, training].
Motivation for this was that TransformerEncoder fast path (so with_no_grad=True) and use_torchscript=True would crash with the issue that NestedTensor doesn't have size. This was caused because the TransformerEncoder fast path generates a NestedTensor automatically as a perf optimization and torchscript attempts to find intermediate tensor sizes while it optimizes. But NestedTensor has not implemented a size method, so things fail.
This test goes together with this fix https://github.com/pytorch/pytorch/pull/79480
Test Plan:
```
buck build --show-output mode/opt -c fbcode.enable_gpu_sections=true -c fbcode.nvcc_arch=a100 mode/inplace //caffe2/test:transformers
./fbcode/buck-out/gen/caffe2/test/transformers#binary.par
```
Test runs and passes together with the changes from the PR above (I made another diff on top of this with those changes). Does not pass without the fix.
Reviewed By: mikekgfb
Differential Revision: D37222923
Pull Request resolved: https://github.com/pytorch/pytorch/pull/79796
Approved by: https://github.com/zrphercule
