Enabled mask and is_causal parameters for torch.aten.scaled_dot_product
attention + relevant comments + tests.
The tests added highlight the new capabilities introduced in this PR,
including:
Attention with F16 mask
Attention with Boolean mask
Causal attention with same Q K V shapes
Causal attention without Q K V shapes
Made sure that one cannot input both mask and is_causal.
As titled, create a new decomposition for `aten.fmod.Tensor` to
`aten.div`, `aten.trunc`, `aten.mul` and `aten.sub`. Note that we only
use `aten.trunc` for floating point operations. This further gets
decomposed to `aten.where` etc. by other existing decompositions.
This decomposition now makes TOSA pass for a simple model with
`aten.fmod` while it makes `stablehlo` fail. For now, we disallow this
decomposition for `stablehlo`
---------
Co-authored-by: Srinath Avadhanula <srinath.avadhanula@getcruise.com>
The lowering pattern for `aten.T` uses transposition implemented via
`linalg.generic`. For downstream passes it is advantageous to use named
ops wherever possible, so this patch changes the lowering to use
`linalg.transpose` instead.
Addresses an issue in <https://github.com/llvm/torch-mlir/issues/3651>
where some unflatten ops generated from onnx models weren't propagating
static shape information. It may be necessary to add further
optimizations for the more general case when some static information is
present in the unflatten (or possibly reshape/view) op's `sizes` list,
but not reflected in the output shape. These ops will only successfully
infer shapes if the `sizes` list is gotten from a list of constant ints
(with possibly one -1). A common example where this fails is when some
of the `sizes` are determined from `aten.size.int` ops on dynamic
tensors, and other `sizes` are known statically.
This PR includes:
- a canonicalizer for `aten.unflatten.int` which converts to
`aten.unsqueeze` when it is expanding one dim to two, and one of the new
dims is statically 1.
- an improvement to the folder for `aten.__or__.bool` which does not
rely on *both* operands being static.
This PR add `floordiv` to the `PY_BUILTIN_TO_TORCH_OP`. For
`aten.mul.int` and `aten.floordiv.int` ops, we add new Canonicalization
Patterns as follow:
```
%1 = torch.aten.mul.int %input, %const-5
%2 = torch.aten.mul.int %1, %const-6
```
Will be replaced by
`torch.aten.mul.int %input, %const-30`
And
```
%1 = torch.aten.mul.int %input, %const-5
%2 = torch.aten.floordiv.int %1, %const-5
```
Will directly return `%input`
This PR also relaxes the `float` type constraint in TorchToTosa for the
`AtenRsubScalarOp` conversion.
To test:
`cmake --build build --target check-torch-mlir-all`
Supports the result with dynamic shape and scalar indices like
```
func.func @test_gather_scalar(%arg0: !torch.vtensor<[3,4,5],f32>, %arg1: !torch.vtensor<[], si64>) -> !torch.vtensor<[?,?],f32> attributes {torch.onnx_meta.opset_version = 13 : si64} {
%0 = torch.operator "onnx.Gather"(%arg0, %arg1) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[3,4,5],f32>, !torch.vtensor<[], si64>) -> !torch.vtensor<[?,?],f32>
return %0 : !torch.vtensor<[?,?],f32>
}
```
`Torch::AtenSqueezeOp` is referring to the result shape, so it will
failed on lowering if the result shape is dynamic.
The current implementation uses a `linalg.generic` to broadcast the bias
tensor for the lowering of convolutions. This is suboptimal for later
pattern matching. This patch changes it to use the respective named op,
`linalg.broadcast`, instead.
The `axis` attribute is optionally available. Added support by computing
the pad based on the axis values.
---------
Signed-off-by: Rob Suderman <rob.suderman@gmail.com>
- This PR adds new (and equivalent) more tensorized impl of
MelWeightMatrix which lowers all the way to linalg.
- [Ref Pytorch
Impl](https://gist.github.com/PhaneeshB/4e6dfcded3007b1b686fbe28f07a67cd)
- Thanks to @rsuderman for pointing out the difficulties [earlier
impl](#3503) posed during lowering to linalg and also for providing a
better numpy impl 🙏
This commit adds the shape info for the tensors created during the
decomposition of GroupNorm op.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
Set PyTorch and TorchVision version to nightly release 2024-08-18.
This commit also updates the `scaled_dot_product_attention` op.
A new attribute `enable_gqa` has been added. As of now, only the
default value for the same is supported.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
Discovered in https://github.com/llvm/torch-mlir/issues/3104
Most likely when building with stablehlo, while waiting for it missing
dependency was generated to location shared with another dependency.
This commit extends the OnnxToTorch lowering for BatchNormalization op
for supporting the case when training=True.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
The `layout` attribute was not considered for the `onnx.RNN` operation.
Added support for the attribute to transpose the inputs / outputs of the
RNN when valid.
The einsum lowering was missing the behavior for duplicate indices in
the equation. This amounts to a diagonalization along duplicate pairs of
indices in the equation.
Closes#3575
The PyTorch remainder operator is meant to compute the Python modulus
operator entrywise:
https://pytorch.org/docs/stable/generated/torch.remainder.html#torch.remainder
In python the modulus operator is meant to always return a result with
the same sign as the divisor:
https://docs.python.org/3/reference/expressions.html#binary-arithmetic-operations
In other words, torch.aten.remainder should return a Python-style
modulus instead of a C-style modulus. However the remainder operator was
simply translated into arith.ModSI or arith.ModF, which both effectively
compute the C-style modulus. Now the lowering has been modified so that
the modulus operator works properly with negative numbers, both in the
dividend, and the divisor.
This patch adds basic support for lowering graphs with per-channel
quantization. Per-channel quantized ops have to be excluded from
`FuseQuantizedOps` for now but can be used in QDQ quantized form.
Using this patch, we're able to import and execute (on the linalg
backend) graphs with per-channel quantization applied using the "new"
PyTorch 2.0 Export Quantization.
The saga of aligning onnx and torch padding conventions continues.
```python
onnx_pads = [low_x, low_y, low_z, high_x, high_y, high_z]
torch_pads = [low_z, high_z, low_y, high_y, low_x, high_x]
```
So not only is the lexicographical ordering hierarchy swapped (low/high
x spatial-dim -> spatial-dim x low/high) but the ordering in the the
spatial-dim specification is also reversed.
This patch properly reverses the pad ordering (and actually uses the
`shuffledPadding` to pad).
rohan-tan-bhowmik
Srinath Avadhanula
Felix Schneider
Branko Trifkovic
justin-ngo-arm
zjgarvey
Ze Zhang
Vivek Khandelwal
jinchen
lingzhiz1998
Xida Ren (Cedar)
Rob Suderman
Phaneesh Barwaria
Ian Wood
yyp0
Yevhenii Havrylko
pkapris-syrmia
Yuanqiang Liu
aldesilv
Marius Brehler