This commit sets the PyTorch and TorchVision version to nightly release
2024-11-07.
This commit also updates the dtype check for the
`aten.fake_quantize_per_tensor_affine` and
`aten.fake_quantize_per_tensor_affine_cachemask` op since the op now
supports bfloat16 input.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
1. Adds case handling for `aten.slice.tensor` shape inference with
negative strides. This is not technically allowed by native pytorch, but
it is useful for ONNX ingest. We were getting some incorrect shapes for
these negative strided slice ops.
2. Adds scalarization support for ops seen in pytorch pad exports to
ONNX. These are typically `aten.view` `aten.transpose.int` and
`aten.slice.Tensor` with negative strides (and rank 2).
3. Allows view op `self` to be added to the worklist conditionally,
based on whether the view op actually occurs as a middle point in a
shape computation.
Attention often broadcasts a mask across the batch dimension as masking
is usually performed the same across attention heads. Added this
materialization to the mask dimensions optionally.
# Tracking
[Issue](https://github.com/nod-ai/SHARK-ModelDev/issues/848)
[TorchToLinalg Op
Support](https://github.com/nod-ai/SHARK-ModelDev/issues/347)
# Description
Aten_TrilinearOp is an implementation of a "trilinear einstein sum".
Essentially, just an einsum across 3 tensors.
There are a few inputs:
## Tensor Inputs
- i1, i2, i3 - The three input tensors for the _trilinear op.
## Expands
These inputs allow you to unsqueeze an input tensor at the specified
dims as a pre-processing step to make the shapes compatible for the rest
of the op:
- expand1: List[int], expand2: List[int], expand3: List[int]
## sumdim
- sumdim: List[int] - After applying element wise multiplication, the
values in sumdim denote where to collapse a dimension by summing over it
## unroll_dim
- unroll_dim: int - In the PyTorch implementation, this specifies a
dimension where you could slice the input tensors, multiply and sum
them, then concatenate the results in an output tensor. This complicates
the implementation significantly, but doesn't change the result, so I
opted against it. Along with that, a previously accepted path for
solving this involved reusing the AtenEinsumOp, which also would also
ignore this input.
# Solution
After trying a bunch of more complicated approaches for it, this op
actually ended up being quite simple: [See
_trilinear](https://dev-discuss.pytorch.org/t/defining-the-core-aten-opset/1464)
`_trilinear = (i1.unsqueeze(expand1) * i2.unsqueeze(expand2) *
i3.unsqueeze(expand3)).sum(sumdim)`
Wish I saw this earlier, but watcha gonna do: 🙃
## Not Reusing AtenEinsumOp
Frankly, I found multiple cases where valid inputs would have numerical
mismatches for EinsumOp, even when running tests against EinsumOp
directly. I think it has something to do with the singleton dimensions.
Will need to look into this further, but once I realized the simplified
approach, it appeared to be more reliable and much simpler.
Either way (credit to @zjgarvey), there are improvements to the einsum
op here. When I was originally trying to use the op, intermediate
tensors were being flattened properly, but then its 0th dimension was
being cast from a static dim to a dynamic dim due to integers not
folding correctly in the MLIR. Figured it's worth keeping these
improvements for future reusers of EinsumOp.
# The zero'd out dim "bug"
For some reason, if you specify a dimension in all `expands`,
```i.e.
[expand1=[0], expand2=[0], expand3=[0]],
[expand1=[1], expand2=[1], expand3=[1]]
```
The _trilinear op would specify `0` for that dimension in the output
shape, unless it was also included in `sumdim`. This goes against the
implementation of torch.einsum:
```
>>> a, b, c = [torch.rand(1, 3, 3, 3) for i in range(3)] # Simulate expand at dim=0 for all input tensors
>>> torch.einsum('abcd,abcd,abcd->abcd', a, b, c).shape
torch.Size([1, 3, 3, 3])
```
And is just straight up incorrect mathematically. I considered
"replacing" singleton dims with zeroed out dims, but that seemed like
carrying over a bug. Instead, I included a test for the case, verified
that the singleton dimensions were handled the way that torch.einsum
handles it, instead of torch._trilinear, and xfailed it with a note as
to why.
This commit sets the PyTorch and TorchVision version to nightly release
2024-10-29.
This commit also fixes the CI failure after this commit
54d9e24013
got merged. The issue was that the CI checks in the PR were run before
the previous roll pytorch update but the PR was actually merged after
the roll pytorch update. Hence, the failure was not caught before
merging the PR.
While exporting the fx_graph through fx_importer for `rrelu` and
`rrelu_with_noise` op for train mode, it decomposes the
`aten.rrelu_with_noise` op based on the PyTorch decomposition which is
the default behavior. However, the decomposition contains an input
mutation specifically here
9bbe4a67ad/torch/_decomp/decompositions.py (L325),
resulting in the runtime failure. This issue would probably be fixed by
https://github.com/pytorch/pytorch/pull/138503. Until then, the failing
tests are added to the xfail set.
Also, after the roll pytorch update following tests started passing for
fx_importer, and fx_importer_stablehlo config.
- "ElementwiseRreluTrainModule_basic"
- "ElementwiseRreluTrainStaticModule_basic"
- "ElementwiseRreluWithNoiseTrainModule_basic"
- "ElementwiseRreluWithNoiseTrainStaticModule_basic"
This commit also updates the dtype check for the `aten.linear` op since
the op now expects both the input tensors to have the same dtype.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
The fx tracer does not support tracing "real" quantized tensors
currently. A "real" quantized tensor here means a tensor that is created
using a method like `torch.quantize_per_tensor()` and carries the
quantization parameters (scale, zero_point, scheme) in the object.
However, it seems like the DQ-Q type fake quantizatation is now commonly
used as a high level representation of quantized operators and is only
lowered to native quantized ops (if available) in the respective
hardware backend. Quantization of floating point modules in PyTorch is
recently also performed as a graph transformation after
exporting/tracing the original module.
```python
# Examples of "real"/native quantization
tens = torch.randint(-127, 127, (1,), dtype=torch.int8)
torch._make_per_tensor_quantized_tensor(tens, 1, 0)
# tensor([90.], size=(1,), dtype=torch.qint8,
# quantization_scheme=torch.per_tensor_affine, scale=1.0, zero_point=0)
tens = torch.rand((1,))
torch.quantize_per_tensor(tens, 1, 0, torch.qint8)
# tensor([1.], size=(1,), dtype=torch.qint8,
# quantization_scheme=torch.per_tensor_affine, scale=1.0, zero_point=0)
# Example of DQ/Q quantization
import torch.ao.quantization.fx._decomposed
tens = torch.rand((1,))
torch.ops.quantized_decomposed.quantize_per_tensor.default(tens, 1, 0, -128, 127, torch.int8)
# tensor([1], dtype=torch.int8)
```
This means that a typical import flow for a quantized network
into/through torch-mlir would look like this:
`torch.export() -> quantization transformations on fx graph ->
fx_importer` Where the tensors in the graph are normal float/int tensors
and the quantization parameters are carried by the DQ/Q ops. These kinds
of graphs can be traced without issues.
Currently, our quantized convolution tests use the "real" quantized
tensors. This means that with the retirement of the `jit_ir_importer`,
these tests cannot be imported any longer. In summary, I see no reason
to stick to the "real" quantization in these tests, as both PyTorch 2.0
is using DQ/Q quantization and our linalg backend is also using it.
This patch updates our quantized convolution tests to use the DQ-Q
quantization with the ops from `torch.ops.quantized_decomposed`.
Note: For future reference, there seems to be an ongoing consolidation
of the ops for the DQ/Q scheme on the PyTorch side
(https://github.com/pytorch/ao/issues/986#issuecomment-2390296826).
This commit adds the torch-onnx-to-torch-backend pipeline which
converts the Torch Onnx IR to Torch Backend IR.
This commit also moves the `ScalarizeShapes` pass from the
`torch-backend-to-linalg-on-tensors-backend-pipeline` to the
`torch-onnx-to-torch-backend` pipeline since the primary goal of
this pass is to scalarize the shapes in the IR coming from the
Onnx models.
Set PyTorch and TorchVision version to nightly release 2024-10-15.
Tracker issue for the failing tests added to xfail_set in this PR.
Issue: https://github.com/llvm/torch-mlir/issues/3796
This commit disables the failing sparse tensor tests since they are not
maintained on day-to-day basis and blocks the roll PyTorch update for now.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
This was preventing dynamic dims in an ONNX model from being reified (causing the generation of `tensor.cast`s and preventing fusion in iree):
```mlir
%2 = torch.vtensor.literal(dense<[4, 256]> : tensor<2xsi64>) : !torch.vtensor<[2],si64>]
%7 = torch.prim.ListConstruct %int2 : (!torch.int) -> !torch.list<int>
%8 = torch.aten.reshape %2, %7 : !torch.vtensor<[2],si64>, !torch.list<int> -> !torch.vtensor<[2],si64>
//... chain of foldable ops linking %2 to the `shape` operand of a `torch.aten.broadcast_to ... -> !torch.vtensor<[?,?],si64>`
```
# Description
Implementation of the op for `torch.aten.unfold`: [TorchToLinalg Op
Support #347](https://github.com/nod-ai/SHARK-ModelDev/issues/849)
Documentation of op can be found here: [PyTorch
Docs](https://pytorch.org/docs/stable/generated/torch.Tensor.unfold.html)
For this op, we apply a sliding window of some `size` along a single
`dimension`, with `step` in between iterations.
`Declaration: aten::unfold(Tensor(a) self, int dimension, int size, int
step) -> Tensor(a)`
The resulting `unfolded` tensor modifies the shape of `dimension` to be
equal to the number of blocks that the sliding windows extracts/inserts,
with an additional dimension of `size` appended (the number of cols of
the output tensor directly translates from the size of the sliding
window).
So if we had a tensor of rank 3 (A x B x C), with dimension = 1, size =
2 and step = 2:
(A x B x C) |=> (A x (B - size) // step + 1 x C x size)
After extracting the window from the input tensor, we insert the (1 x
size) slice into the output tensor. We can make this simpler by mapping
the output indices from the input indices, like they do in the official
implementation:
[PyTorch
Code](https://github.com/pytorch/pytorch/blob/main/torch/_inductor/lowering.py#L1694)
This commit adds the support for the 1-d depthwise convolution as a
special case of 1-d group convolution.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
Addresses ~200 onnx model compile failures in
<https://github.com/nod-ai/SHARK-TestSuite> related to
<https://github.com/iree-org/iree/issues/18631>.
This change simplifies the result of the generated broadcast op
substantially, but reduces the case coverage slightly.
The case which will become unsupported:
- trying to actually broadcast a dynamic dim that is secretly 1.
When does this case appear in practical scenarios?
- for a model where onnx shape inference cannot figure out that a dim
should be 1.
Why do I think we should not support this case for now?
1. For all models with dynamic dim expand ops, the previous path
uniformly generates uglier linalg IR (making it harder for IREE to fuse
properly with other ops).
2. For models failing shape inference castastrophically enough to fail
to see a dim is statically 1, we can try to apply constant folding in
the onnx model before importing.
Leaving this as a draft PR, since it may be more appropriate to fix the
compilation failure in IREE rather than torch-mlir.
### Example of broadcast required in previous path:
```mlir
%300 = linalg.generic {indexing_maps = [#map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} outs(%299 : tensor<?x12x?x?xi1>) {
^bb0(%out: i1):
%306 = linalg.index 0 : index
%307 = linalg.index 3 : index
%308 = arith.index_cast %285 : i64 to index
%309 = arith.cmpi eq, %308, %c1 : index
%310 = arith.select %309, %c0, %306 : index
%311 = arith.index_cast %286 : i64 to index
%312 = arith.cmpi eq, %311, %c1 : index
%313 = arith.select %312, %c0, %307 : index
%extracted_79 = tensor.extract %reshape_78[%310, %c0, %c0, %313] : tensor<?x1x1x?xi1>
linalg.yield %extracted_79 : i1
} -> tensor<?x12x?x?xi1>
```
### Example of broadcast with simplified shape list:
```mlir
%409 = linalg.generic {indexing_maps = [#map15, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%reshape_135 : tensor<?x1x1x?xi1>) outs(%408 : tensor<?x12x?x?xi1>) {
^bb0(%in: i1, %out: i1):
linalg.yield %in : i1
} -> tensor<?x12x?x?xi1>
```
Current version does not work for a mixture of dynamic and static shaped
batch dimensions. Rework to grab the correct dynamic shapes.
---------
Co-authored-by: dan <danimal197@gmail.com>
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>