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>
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`
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>
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.
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 static uneven divisible AdaptiveAvgPool2d means that although the
input size is not an integer multiple of ouput size, but the kernel and
stride size can also be fixed (not dynamic). The derivation logic of
kernel and stride size is consistent with
torch/_decomp/decomposations.py:adaptive_avg_pool2d as described in the
following:
1. Stride Size
Firstly , derive the start index in each reduce operation according to
the output size (`n`), `start_index = ([0, 1, ..., n - 1] * input_size)
// output_size`. For each index `k`, if `k * (input_size % output_size)
< output_size`, then the current and previous stride keeps the same as
`input_size // output_size`. So suppose `(n-1) * (input_size %
output_size) < output_size`, the stride in the whole AdaptiveAvgPool2d
process keeps static, as `input_size // output_size`.
2. Kernel Size
torch/_decomp/decomposations.py:adaptive_avg_pool2d calculates a static
kernel size when the input/output sizes satisfy either of the two
conditions, `input_size % output_size == 0` or `output_size %
(input_size % output_size) == 0`. Here if `input_size % output_size ==
0`, then the kernel size equals `input_size // output_size`, otherwise
`input_size // output_size + 1.`
Torch has all scalars represented as i64 and f64 types which results in
extraneous trunc-extf commands. We can rework this by elliding
widen-narrow cases away.
This PR adds a conversion in the TorchOnnxToTorch pass for the ONNX
Multinomial operation. It also adds a TorchToLinalg lowering for the
`aten.Multinomial` op and does a light refactor of some repeated code
that generates random floating point numbers in
`TorchToLinalg/Random.cpp`.
This patch adds a few misc pad op related changes:
1. Addresses issue <https://github.com/llvm/torch-mlir/issues/3457>
2. Addresses issue <https://github.com/llvm/torch-mlir/issues/3442>
3. Fixes the padding order for asymmetrically padded onnx.Conv ops
4. Enables passing quantization through those onnx.Conv op pre-paddings
5. Modifies the torch-to-linalg lowering of AtenReplicationPad2d op to
enable support for input rank != 4
Unfortunately, even with all of these changes, the e2e tests for the
ReplicationPad2d still fail the onnx config, since the torch export
procedure for rearranging the pad order is complicated enough that the
padding ints end up not being able to fold back to constants.
Register `aten.fake_quantize_per_channel_affine` and
`aten.fake_quantize_per_tensor_affine.tensor_qparams` ops
---------
Co-authored-by: Ze Zhang <ze.zhang@getcruise.com>
Fix the pad tensor rearrangement such that we change the representation
from [x1_begin, x2_begin, ..., x1_end, x2_end,...] to [xn_begin, xn_end,
...., x2_begin, x2_end, x1_begin, x1_end] where x1, x2 .. xn are the
dimensions of the pads tensor argument.
---------
Co-authored-by: zjgarvey <zjgarvey@gmail.com>
Co-authored-by: zjgarvey <47986913+zjgarvey@users.noreply.github.com>
Due to the custom operation parser, the print and parser were expecting
two different forms.
One having the dictionary before the value and the other after.
Following the format of the other constants ops, the constant.int will
follow the `value attr-dict` format. Updated the parser accordingly.
This bump triggered an upstream assert. Includes a WAR for #3506.
Also includes several things I needed to do to repro:
* When TORCH_MLIR_TEST_CONCURRENCY=1, test runs will be printed.
* Added TORCH_MLIR_TEST_VERBOSE=1 handling to enable verbose mode
(useful on CI).
---------
Co-authored-by: Stella Laurenzo <stellaraccident@gmail.com>
Before this PR, a statically shaped aten.convolution would generate
dynamically shaped linalg IR, and even `-canonicalize` would not be able
to fold it back into static shapes. This PR ensure that shape
calculations are folded on construction to directly generate statically
shaped linalg IR.
We achieve that by ensuring that `arith` ops involved in computing
shapes are created via `createOrFold`, so that later uses of
`getAsOpFoldResult` see constants instead of those ops.
For example
```
module {
func.func @forward(%arg0: !torch.vtensor<[32,336,112,112],f32>,
%arg1: !torch.vtensor<[336,168,3,3],f32>,
%arg2: !torch.vtensor<[336],f32>)
-> !torch.vtensor<[32,336,56,56],f32> {
%false = torch.constant.bool false
%int2 = torch.constant.int 2
%int1 = torch.constant.int 1
%0 = torch.prim.ListConstruct %int1, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
%1 = torch.prim.ListConstruct %int2, %int2 : (!torch.int, !torch.int) -> !torch.list<int>
%2 = torch.prim.ListConstruct : () -> !torch.list<int>
%3 = torch.aten.convolution %arg0, %arg1, %arg2, %1, %0, %0, %false, %2, %int2
: !torch.vtensor<[32,336,112,112],f32>, !torch.vtensor<[336,168,3,3],f32>, !torch.vtensor<[336],f32>, !torch.list<int>,
!torch.list<int>, !torch.list<int>, !torch.bool, !torch.list<int>, !torch.int
-> !torch.vtensor<[32,336,56,56],f32>
return %3 : !torch.vtensor<[32,336,56,56],f32>
}
}
```
would result in
```
[...]
%padded = tensor.pad %2 low[%14, %15, %16, %17] high[%14, %15, %16, %17] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst : f32
} : tensor<32x336x112x112xf32> to tensor<?x?x?x?xf32>
[...]
%45 = linalg.conv_2d_ngchw_gfchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>}
ins(%expanded, %expanded_37 : tensor<?x2x?x?x?xf32>, tensor<2x168x168x3x3xf32>)
outs(%expanded_44 : tensor<32x2x168x?x?xf32>) -> tensor<32x2x168x?x?xf32>
[...]
```
and with this PR all shapes are static.
This adds support for a few ops:
- torch.linalg_det
- torch._linalg_det (if the LU and pivot returns are unused)
- onnx.Det
An scf loop is used, since the row reduction algorithm applied here has
some loop-carried dependencies.
The current support being added here is very basic, and only works if no
permutations are required during row reduction, and assumes the matrices
are non-singular.
This adds a torchvision op to torch-mlir and a path from onnx.DeformConv
to torchvision.deform_conv2d.
I'm not implementing the torch->linalg lowering for the torchvision op
yet, but posting this PR to get feedback on some of the choices being
made here and to flesh out the onnx frontend a bit.
This adds an onnx->torch conversion for onnx.RoiAlign into
torchvision.roi_align or torchvision.roi_pool, and adds those two
torchvision ops to torch-mlir.
Add a new op with shape/dtypes and decompose into
`fake_quantize_per_tensor_affine` when the second result is unused.
The xfail_set change is on ONNX because torch cannot export this op to
ONNX.
1. truncates zero-points to i32
2. modifies the default accumulator type for i8 from i64 to i32.
3. now uses the input dtype to infer accumulator dtype.
This implements the Onnx.NegativeLogLikelihoodLoss op using the
signature provided
[here](https://onnx.ai/onnx/operators/onnx__NegativeLogLikelihoodLoss.html)
by replacing it with a `NLLLossForward` op.
Additionally, I included a helper function `get_loss_reduction_enum` to
convert from a string `reduction` parameter to the corresponding
intended integer value since this is an operation that will be reused
for any loss function module. This differs from `get_reduction_enum` in
`TorchUpstream.cpp` which handles the `reduce` parameter from
`scatter_reduce` type operations.
There is currently no int16 quantization support in torch. This patch
adds a new mlir type to correspond to the missing "torch.qint16" type,
and enables lowering of quantization-related onnx ops using int16 types.
In follow-up patches, custom quantization logic for ops like
aten.matmul/aten.mm/aten.convolution may need to be revisited to allow
support for qint16. The passes in FuseQuantizedOps.cpp may also need
slight modifications.
This addresses 7 of the model failures I'm seeing in the test suite. See
[Shark-Turbine issue
#566](https://github.com/nod-ai/SHARK-Turbine/issues/566).
Need the op ```linalg.conv_2d_ngchw_gfchw_q``` to be added upstream
before merging this. See [llvm-project PR #92136
](https://github.com/llvm/llvm-project/pull/92136).
A small additional expansion to operand quantization is included in this
patch to address a model failure that occurs when unblocking the
quantized group convolutions in one of these onnx models.
Updates:
- some unsupported modes are now going to report a match failure for
unsupported coordinate transformation modes.
- fixes a bug that was introduced in the last patch for resize (my
bad...)
- uses actual x and y coordinates for computing weights in bilinear
interpolation (rather than eps modified values)
- slightly simplifies the bilinear interpolation payload for readability
and performance
- passes coordinate transformation mode information from an onnx.Resize
op to the mode string for the aten._interpolate op. This allows us to
perform custom logic in the torch->linalg lowering to support
onnx.Resize options without losing the default behaviors of the
interpolate op.
This PR fixes the bugs for `Torch::AtenOneHotOp` by:
1) Using `Torch::kUnknownSize` as the default value for `numClasses` in
the pattern matching stage in `DecomposeAtenOneHotOp`
2) Adding `AtenIntScalarOp` to the patterns in `TorchToArith`
3) Handling both `int` and `float` types for `off` and `on` values in
`TorchOnnxToTorch` conversion
It also includes:
1) A new test in `TorchToArith/basic.mlir`, for `torch.aten.Int.Scalar`,
and
2) A new test in `decompose-complex-ops.mlir`, for `torch.aten.one_hot`
**Dependencies**
This PR is dependent on #3334.
I am trying to eliminate 'getWithLeastStaticInformation' in
DecomposeAtenTriuOp. Could you provide me with some suggestions?
@qingyunqu @zjgarvey
See issue https://github.com/llvm/torch-mlir/issues/3312
Discord Thread:
https://discord.com/channels/636084430946959380/1238330633328005243
## Context:
[This](https://github.com/llvm/torch-mlir/blob/main/python/torch_mlir/fx.py#L61)
was updated to support e2e tests for the TorchDynamo frontend in
Torch-MLIR, where we run FX decompositions and import the FX IR to
generate Torch dialect, followed by
`torch-function-to-torch-backend-pipeline`, skipping only the shape/type
refinement for now. However, we should be able to skip many of the torch
simplification passes, as depicted in the [frontend
roadmap](https://github.com/llvm/torch-mlir/blob/main/docs/images/roadmap_frontend.png).
Based on IREE's TorchDynamo
[pipeline](https://github.com/iree-org/iree/blob/main/compiler/plugins/input/Torch/InputConversion/Passes.cpp#L29),
the only two passes we seem to require are: `ReduceOpVariantsPass` and
`DecomposeComplexOpsPass`. This is inline with our findings as well
based on initial exploration.
This PR creates a dedicated frontend simplification pipeline for
TorchDynamo / FX Importer which calls only `ReduceOpVariantsPass` and
`DecomposeComplexOpsPass`. We rely on the e2e fx_importer tests to
ensure we're not regressing by removing many of the passes that were
historically needed for TorchScript.
One notable change here is that we do not call the
`LowerToBackendContractPass` anymore, which used to call
`TorchSimplificationPipeline` iteratively until VerifyBackendContract
was clean. Some of this was required for the shape/type refinement to
converge, which seems a non-issue for Dynamo frontend. Do we anticipate
this (the iterative invocation of TorchSimplificationPipeline followed
by VerifyBackendContract) to be worth retaining in the Dynamo frontend
pipeline? If so, I can make those changes, PLMK.
In constant folding progress, a new constant op will be created
according to the origin op's result type.
See the code in TorchDialect.cpp.
```cpp
Operation *TorchDialect::materializeConstant(OpBuilder &builder,
Attribute value, Type type,
Location loc) {
if (auto integerType = dyn_cast<Torch::IntType>(type))
return builder.create<Torch::ConstantIntOp>(loc, cast<IntegerAttr>(value));
if (auto floatType = dyn_cast<Torch::FloatType>(type))
return builder.create<Torch::ConstantFloatOp>(loc, cast<FloatAttr>(value));
if (auto numberType = dyn_cast<Torch::NumberType>(type)) {
if (auto floatValue = dyn_cast<mlir::FloatAttr>(value)) {
return builder.create<Torch::ConstantNumberOp>(loc, floatValue);
} else if (auto intValue = dyn_cast<mlir::IntegerAttr>(value)) {
return builder.create<Torch::ConstantNumberOp>(loc, intValue);
}
}
if (isa<Torch::BoolType>(type)) {
return builder.create<Torch::ConstantBoolOp>(loc, cast<IntegerAttr>(value));
}
if (isa<Torch::NoneType>(type))
return builder.create<ConstantNoneOp>(loc);
if (auto stringAttr = dyn_cast<StringAttr>(value))
return builder.create<ConstantStrOp>(loc, stringAttr);
if (auto elementsAttr = dyn_cast<ElementsAttr>(value)) {
// Only !torch.vtensor can be constant folded. !torch.tensor has
// non-trivial aliasing semantics which prevent deduplicating it.
assert(isa<ValueTensorType>(type) && "should be a vtensor type!");
return builder.create<ValueTensorLiteralOp>(loc, elementsAttr);
}
return nullptr;
}
```
So when the op has a tensor result type, it must be "ValueTensorType"
due to the **assert** statement. However, many fold methods in
TorchOps.cpp only have a judgment of "BaseTensorType".
After running the model tests in SHARK-TestSuite, I noticed a few model
failures due to half-fusion.
Notably, RDN_pytorch_vaiq_int8 had a depth=5 convolution chain with
multiple AtenViewOp's.
This change enables more customization with operand quantization, and
generalizes the patterns QuantizeOperands and QuantizeTransposeOperands
to QuantizeOperandsPastCommutingOps.
This allows for passing quantization through operations which are
functionally unaffected by quantization, such as view-like ops. The
purpose of this change is to address a myriad of quantization issues
seen in quantized onnx models that have some reshape-like operations
sandwiched in between a dequant and something like a matmul (whose other
operand is immediately quantizable).
rohan-tan-bhowmik
Srinath Avadhanula
Branko Trifkovic
zjgarvey
Ze Zhang
Vivek Khandelwal
lingzhiz1998
Xida Ren (Cedar)
Rob Suderman
pkapris-syrmia
Yuanqiang Liu
Felix Schneider
Chi_Liu
Rob Suderman
yyp0
Ivan Butygin
bosko-syrmia
Arham Khan
Xinyu Yang
Sagar Kulkarni
Scott Todd
Jiawei Wu
Phaneesh Barwaria
Christopher McGirr
Aart Bik
Matthias Gehre
Vinayak Dev
ptrifunovic98
Lei Zhang
Sambhav Jain
penguin_wwy
Angel Zhang
Peiming Liu
Aaron St George
NeverRaR
aldesilv
