To handle the conversion from raw bytes to `DenseElementsAttr` we need
to handle the endianness conversion during `torch-onnx-to-torch`.
Therefore when importing `onnx.Constant` it is better to represent using
the `onnx` constant operation so that only one location requires the
endianness correction.
`onnx` explicitly specifies that `raw_data` is stored in `little-endian`
layout. While converting
to `torch` we need to convert from a known endian format to an internal
format of consistent
layout. This means endianness must be correct during the import of
`onnx.Constant`.
---------
Co-authored-by: Xida Ren (Cedar) <cedar.ren@gmail.com>
Required some massaging of LTC to make it warning clean, and I had to
manually disable some warnings on the generated source files (which we
don't control).
The project is warning clean now.
The `-Werror` flag is disabled by default as we can't control everywhere
people will try to build/install. The CI enables it via
-DTORCH_MLIR_ENABLE_WERROR_FLAG=ON.
Note that we are waiting for actual FX traced graph support for sparse
tensors. For details see
https://github.com/pytorch/pytorch/issues/117188
Until then, however, we provide this clever importer that builds the FX
traced graph for for the dense case and then puts a sparse annotation
back on the parameters.
With import test.
This PR adds a check to the CI right after checking out the Torch-MLIR
repository to make sure that the changes in the PR don't require any
`git clang-format` modifications.
Linalg has quantized specific operations. We can lower to these
operations when there is a known zeropoint and scale operations. This
allows the `convolution` to occur with lower bitwidth's, improving the
overall performance.
We were seeing some assertion failures after some checks around folders
were tightened up in LLVM:
https://github.com/llvm/llvm-project/pull/75887 . This PR essentially
moves the logic that used to be applied at the LLVM level into the
folder, which seems to be the suggested fix.
I'm not sure if the IR that caused issues for us _should_ be valid?
```
%1 = torch.aten.detach %arg0 : !torch.tensor<[1],f32> -> !torch.tensor
```
A better fix might be to create a verifier ensuring the result of
`aten.detach` has the same type as its operand.
---------
Co-authored-by: aaron-stgeorge <aaron.stgeorge@getcruise.com>
This allows the following command to be used to ignore sweeping
formatting changes.
```
git blame --ignore-revs-file .git-blame-ignore-revs <file_of_interest>
```
After noticing a number of commits with unrelated formatting changes,
I think something was changed with clang-format at one point and we're
seeing a number of unrelated changes. Doing a refresh can help avoid
this.
The changes made here came from
```
find lib -iname *.h -o -iname *.cpp | xargs clang-format -i --style=llvm
find include -iname *.h -o -iname *.cpp | xargs clang-format -i --style=llvm
find projects -iname *.h -o -iname *.cpp | xargs clang-format -i --style=llvm
```
Torch does not have an equivalent matmul operation for integers. Instead
it sidechannels the information via its quantized types. For this
lowering we setup these sidechannels then invoke `torch.mm`.
Per the RFC and numerous conversations on Discord, this rebuilds the
torch-mlir CI and discontinues the infra and coupling to the binary
releases
(https://discourse.llvm.org/t/rfc-discontinuing-pytorch-1-binary-releases/76371).
I iterated on this to get latency back to about what it was with the old
(much larger and non-ephemeral) runners: About 4m - 4.5m for an
incremental change.
Behind the scenes changes:
* Uses a new runner pool operated by AMD. It is currently set to manual
scaling and has two runners (32-core, 64GiB RAM) while we get some
traction. We can either fiddle with some auto-scaling or use a schedule
to give it an increase during certain high traffic hours.
* Builds are now completely isolated and cannot have run-to-run
interference like we were getting before (i.e. lock file/permissions
stuff).
* The GHA runner is installed directly into a manylinux 2.28 container
with upgraded dev tools. This eliminates the need to do sub-invocations
of docker on Linux in order to run on the same OS that is used to build
wheels.
* While not using it now, this setup was cloned from another project
that posts the built artifacts to the job and fans out testing. Might be
useful here later.
* Uses a special git cache that lets us have ephemeral runners and still
check out the repo and deps (incl. llvm) in ~13s.
* Running in an Azure VM Scale Set.
In-repo changes:
* Disables (but does not yet delete):
* Old buildAndTest.yml jobs
* releaseSnapshotPackage.yml
* Adds a new `ci.yml` pipeline and scripts the steps in `build_tools/ci`
(by decomposing the existing `build_linux_packages.sh` for in-tree
builds and modularizing it a bit better).
* Test framework changes:
* Adds a `TORCH_MLIR_TEST_CONCURRENCY` env var that can be used to bound
the multiprocess concurrency. Ended up not using this in the final
version but is useful to have as a knob.
* Changes the default concurrency to `nproc * 0.8 + 1` vs `nproc * 1.1`.
We're running on systems with significantly less virtual memory and I
did a bit of fiddling to find a good tradeoff.
* Changed multiprocess mode to spawn instead of fork. Otherwise, I was
getting instability (as discussed on discord).
* Added MLIR configuration to disable multithreaded contexts globally
for the project. Constantly spawning `nproc * nproc` threads (more than
that actually) was OOM'ing.
* Added a test timeout of 5 minutes. If a multiprocess worker crashes,
the framework can get wedged indefinitely (and then will just be reaped
after multiple hours). We should fix this, but this at least keeps the
CI pool from wedging with stuck jobs.
Functional changes needing followup:
* No matter what I did, I couldn't get the LTC tests to work, and I'm
not 100% sure they were being run in the old setup as the scripts were a
bit twisty. I disabled them and left a comment.
* Dropped out-of-tree build variants. These were not providing much
signal and increase CI needs by 50%.
* Dropped MacOS and Windows builds. Now that we are "just a library" and
not building releases, there is less pressure to test these commit by
commit. Further, since we bump torch-mlir to known good commits on these
platforms, it has been a long time since either of these jobs have
provided much signal (and they take ~an hour+ to run). We can add them
back later post-submit if ever needed.
This preserves sparsity at the most obvious places of lowering TORCH
tensors to MLIR RankedTensorType tensors. Other places are marked for
audit. With some initial lowering tests.
This adds an encoding field to the torch type, using the interfaces for
printing, parsing, and verification. Note that although this change
prepares adding sparsity to the torch type (as illustrated by the round
trip and invalid tests), nothing in this change depends on the actual
contents of the encoding field!
This includes custom op matching for decomposed operations and fusing
dequantization into dense operations. As a validation we compare
to the dequant+mm torch implementation.
We can plumb the linear matmul into pytorch using its quantized types
with side channel information. To handle the final int8 operation we
dequantize and requantize.
This commit adds mapping from `onnx.pad` op to `torch.pad` op. Currently
it does not support `axes` parameter of `onnx.pad` op.
Signed-off-by: Gaurav Shukla <gaurav.shukla@amd.com>
The logic here is very similar to the conversion for AdaptiveAvgPool1d
#2661 with a few modifications:
1. buffVal = -inf instead of 0
2. the main linalg generic op accumulates a max, instead of a sum, to
the first output tensor
3. avg pooling requires dividing the sum pool by the kernel width, which
we stored as an auxilliary tensor (kSizeTensor). Here, the auxiliary
tensor will be recording the indices. Strangely enough, the only
signature available for this function is to return indices, and it
appears that they must be computed whether the user desires them or not.
See
[pytorch/torch/nn/functional.py](https://github.com/pytorch/pytorch/blob/main/torch/nn/functional.py#L1174).
Before writing other adaptive pooling conversions, the logic of this
decomposition should be rolled into a helper function that will work for
both max and avg pooling ops. Even the auxiliary tensor should likely be
automated. This code was written in a slightly more tedious way than
strictly necessary (often using loops to fill SmallVectors up to rank-2,
which is only two in this case), in order to more easily facilitate the
transition to a helper function.
convolution with [time,batch,channel] ordering, as opposed to the
default [batch, channel, time]. Currently implementing by transposing
the input and output, but may need to get its own implementation in the
future because this is supposed to be an op that gives a speedup. This
is used by fairseq
(https://github.com/facebookresearch/fairseq/issues/172).
(in case you were wondering like me, this is different from transposed
convolution. Transposed convolution has fractional strides).
---------
Co-authored-by: Xida Ren <xida.ren.dev@gmail.com>
Co-authored-by: Frederik Harwath <frederik.harwath@amd.com>
Fixes https://github.com/llvm/torch-mlir/issues/2764
In the case of OPT, there are ConstantOfShape ops whose input shape is
not static (that is, an initializer), but rather comes from a Constant
op. The importer can't handle such non-static input shapes.
The fix here is to create initializers for a subset of Constant ops
(ones with "value" attributes), so that their outputs can be used
statically. Additionally, there was no case for creating a splat of
int64, so I added that as well.
---------
Co-authored-by: Dave Liddell <dliddell@xilinx.com>
Rob Suderman
Ilija Kalinić
Stella Laurenzo
Aart Bik
Ramiro Leal-Cavazos
Yuanqiang Liu
Aaron St George
James Newling
Vivek Khandelwal
aldesilv
Quinn Dawkins
MaheshRavishankar
Phaneesh Barwaria
lonely eagle
Gaurav Shukla
zjgarvey
Xida Ren (Cedar)
Chi_Liu
Dave Liddell