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`
Previously we only had full suite timeouts, making it impossible to
identify
which specific tests were hanging. This patch adds:
1. Per-test timeout support in the test framework
2. A default 600s timeout for all tests
3. A deliberately slow test to verify the timeout mechanism works
The timeout is implemented using Python's signal module. Tests that
exceed
their timeout are marked as failures with an appropriate error message.
This should help catch and isolate problematic tests that enter infinite
loops, without needing to re-run the entire suite multiple times.
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.
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 adds the `generate-runtime-verification` pass into the linalg
refbackend, and moves all tests that now abort at runtime into the crash
set, sorted by their respective errors.
I have fixed on set of errors found that way, which are mismatches
between the static dimensions we cast to and the actual dynamic
dimensions. This was caused by wrong annotations on the test cases, like
in
https://github.com/llvm/torch-mlir/pull/3615/files#diff-48bfbf41fcad5fa01b49197d251114f84a2b8de4f1d87ab938a061aedd1419b1R1931
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.`
- Adds support for lowering depthwise + quantized convolution ops to
linalg::DepthwiseConv2DNhwcHwcQOp
- Changed the variable name for groupSize (which is really C/G) to the
more appropriate numGroups (G).
- Discovered in e2e testing that linalg does not accept (Cin = groups &&
Cout = K*groups for K>1) as a "depthwise" conv, so this also updates the
case-checking to reflect this issue.
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`.
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>
This commit adds the support for new data types: uint4, and int4 and
uint8 tensor protos. Also, it moves some tests from failing to crashing.
Fixes https://github.com/llvm/torch-mlir/issues/3507
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
Addresses an issue with onnx.Gather lowering to linalg:
<https://github.com/nod-ai/SHARK-Turbine/issues/242>
The builder for tensor.expand_shape, without an explicitly provided
output shape, fails to infer an output shape in the case of multiple
dynamic reassociation dims. I tried adding the output shape explicitly
for tensor.expand_shape, but ran into compilation issues later on (see
<https://github.com/iree-org/iree/issues/17760>).
This PR adds support by lowering this op to tensor.reshape when multiple
dynamic reassociation dims are provided.
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>
The `index_put` operation, `input[indices] = values`, allows for the
values to be any shape that is broadcastable to the slice
`input[indices]`. This commit adds broadcasting support to the Linalg
lowering of `IndexPutHackedTwinOp`.
Fixes: #3465
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.
Resolves#3384.
Many ONNX operators are defined by functions and therefore could be
expanded into simpler ONNX operations during importing, avoiding the
need for tools downstream to support these operators directly.
This commit adds this capability to onnx_importer.py. When importing a
node, the schema for the node's operator is retrieved. If the schema
provides a function for the operator, a specialized version for the
node's types and attributes will be created and imported as an MLIR
function with private visibility. An MLIR function call will then be
emitted, instead of a normal operator node. Caching is used to avoid
generating redundant functions within the same module.
In order to avoid a disruptive change to the importer output for a
large number of operators that already have TorchOnnxToTorch support,
an allowlist strategy is used by default. With this commit, only one
operator is allowlisted for expansion, MeanVarianceNormalization.
However, many other operators can be correctly expanded by the current
code, so hopefully the allowlist can be gradually extended. It is
possible to disable the allowlist in the configuration, in which case
all functions are expanded (useful for testing).
Tools downstream of the importer may now need to do inlining when
consuming the output of the importer, e.g.:
cat imported.mlir | torch-mlir-opt --inline --convert-onnx-to-torch
Explanations for subtle code changes:
- Looking up the correct schema and function for an operator requires
knowing the opset version. NodeImporter retrieves this from the
opset imports on the ModelProto retained by the GraphInfo. Previously,
the model_proto field on GraphInfo was None when importing a subgraph
in import_regions, but this conflicts with the new need for opset
version info. Since the apparent purpose of setting it to None was to
control how GraphInfo generates its input map, a new flag is added to
GraphInfo (is_subgraph) to control this behavior, so that the actual
ModelProto can now be provided without breaking this. This also turned
out to be useful for getting the Config via ModelInfo via GraphInfo.
- Some operators' functions are context-dependent, which means the
function definition depends on the types of the inputs. Therefore node
importing now needs to look up the types of a node's inputs, not just
its outputs as was the case previously. Consequently the operand to
find_type_proto_for_name() may now be a graph input or initializer in
some cases, so it has to be updated.
Issues was found here https://github.com/nod-ai/SHARK-Turbine/issues/643
- [ONNX] Fix padding attributes for onnx.AveragePool
- [Linalg] Add countIncludePad false support for AtenAvgPool1/2dOp
- [Linalg] Add an avg_pool2d countIncludePad False e2e tests
- [Linalg] Fix conflict with AtenAvgPool3dOp
- [Linalg] Fix e2e crash with AtenAvgPool1dOp
- [Linalg] Add dynamic dim support for AtenAvgPool2dOp
- [Linalg] Fix AvgPool2dDivisorOverrideModule crash
This commit also adds the Torch declaration for aten.max_unpool2d and
aten.max_unpool3d op. The TorchToLinalg lowering for the same will be
added in a follow-up commit.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
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.
Yuanqiang Liu
zjgarvey
yyp0
Branko Trifkovic
penguin_wwy
rohan-tan-bhowmik
Srinath Avadhanula
justin-ngo-arm
Ze Zhang
Vivek Khandelwal
lingzhiz1998
Rob Suderman
Xida Ren (Cedar)
pkapris-syrmia
Matthias Gehre
Felix Schneider
Suraj Sudhir
Ivan Butygin
Vinayak Dev
bosko-syrmia
Arham Khan
Xinyu Yang
Gaurav Shukla
Sagar Kulkarni
Jiawei Wu
Phaneesh Barwaria
Aart Bik
Ramiro Leal-Cavazos
ptrifunovic98
Andrea 🦈
Wu Yuan
Chi_Liu
Sambhav Jain