Link to related RFC:
https://discourse.llvm.org/t/rfc-rename-torch-mlir-compile-apis-and-introduce-fx-based-analogs/76646
This commit updates the documentation, tests, CMake files, and API for
the proposed changes in the RFC. There is a new torch_mlir/fx.py for
user level APIs related to importing modules and a corresponding test
for this path can be found at test/python/fx_importer/basic_test.py.
---------
Co-authored-by: MaheshRavishankar <mravisha@amd.com>
If a tensor is initialized by a list with a single constant integer,
this folder turns it into a torch.vtensor.literal
---------
Co-authored-by: Dave Liddell <dliddell@xilinx.com>
So that the CumSum Op in OPT can get the constant that it requires to be lowered to TMTensor
---------
Co-authored-by: Rob Suderman <rob.suderman@gmail.com>
Co-authored-by: Xida Ren <xida.ren.dev@gmail.com>
We do not support average pool when `countIncludePad is set to false.
However if the input is unpadded then the setting of the boolean is
unneeded. Extended use by checking if padding is zero before rejecting
the lowering.
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.
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.
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
```
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 includes custom op matching for decomposed operations and fusing
dequantization into dense operations. As a validation we compare
to the dequant+mm torch implementation.
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>
Handle both `torch.dequantize` and `torch.quantize_per_tensor` including
the op based quantization parameter tracking. This includes adding
`qint32` to torch types as it was missing during the initial type
inclusion.
For testing we only have `torch.int8` and `torch.float` types on
function boundaries as the `qint8` types require passing the scale
and zero point quantization information which is not supported yet.
This PR updates the torch-to-tosa conversion with following changes:
- Support torch.none as min/max input argument for tosa.clamp op
- Support negative value as start index for tosa.slice op
- Add tosa.logical_or lowering support
e2e test:
python -m e2e_testing.main --config=tosa
LIT tests:
cmake --build build --target tools/torch-mlir/all
---------
Co-authored-by: Ze Zhang <ze.zhang@getcruise.com>
Adaptive pooling ops can only be decomposed into their non-adaptive
counterparts in trivial cases.
For example, the current decomposition for AtenAdaptiveAvgPool1dOp in
DecomposeComplexOps.cpp supports outSize = inSize (i.e., do literally
nothing), and outSize = 1 (i.e., do a batched average).
The reason adaptive pooling ops are difficult to lower to linalg is that
they are not constantly strided. They are computed by taking an input
tensor of shape (N, C, Hin), and an output size Hout, and computing the
output tensor at position (n,c, h) in the following way:
1. compute st(h) = (h*Hin)//Hout
2. compute en(h) = 1 + ((h+1)*Hin -1)//Hout
3. apply a computation (max or avg) to the slice: INPUT[n, c,
st(h):en(h)]
The provided sample implementation (for ConvertAtenAdaptiveAvgPool1dOp)
uses tensor.extract to access the input tensor inside the payload of a
linalg generic op. This is likely an unattractive use of linalg generic
ops, which is why I am asking for some more targeted feedback on the
validity of this approach before attempting to support the many other
adaptive pooling ops.
Specifically:
- Is the performance of this implementation bad enough to warrant
targeting different dialects entirely? e.g. TMtensor/linalg ext/ etc.
- If the provided implementation is of acceptable performance to the
community, then is it permissable to remove the Adaptive pooling
decompositions from DecomposeComplexOps.cpp? Based on the current
structure of the -torch-decompose-complex-ops pass, it does not seem
possible to only decompose the adaptive ops in special cases (it seems
to get stuck in an infinite loop on a match failure). I would be happy
to instead incorporate the case logic into the conversion directly, and
remove the decompositions once they are rendered completely obsolete.
As long as this approach is acceptable, I can clean up the
implementation with some helper functions, and quickly add support for
each of the remaining Adaptive pooling ops.
Adds a lowering to Linalg for reflection_pad1d. Based on ideas/code from draft PR
https://github.com/llvm/torch-mlir/pull/2693.
---------
Co-authored-by: Kumar Deepak <kumar@xilinx.com>
As noted in the plan when this work started, we need to produce an ORT
EP plugin for a downstream project, and this will necessitate a C-based
ONNX importer (as opposed to the existing Python one). Because this
comes with dependencies that we do not want to impart on various
projects, this is optional in torch-mlir. It is also factored so that it
can be used as standalone sources in downstreams that need it. Since it
only depends on public C APIs on the MLIR side, this will make build
coupling a lot better (since a C++ dep is not needed on the compiler and
it is trivial to dynamically load).
Our original plan was just to maintain this fork off to the side in our
ORT plugin, but once work started, it seemed better to write it clean
and contribute it upstream for anyone to use. We expect that for non-ORT
use, the Python importer will have better ergonomics for most folks.
I will follow-up with a test suite refactor so that we can drive the
Python or C importer.
This is a relatively mechanical port from Python to C, borrowing some
scaffolding from the old JitIR importer. It does attempt to lay some
groundwork for external data, which will need to be implemented on the
Python side as well.
This PR adds the `enable_ir_printing` option to `torch_mlir.compile`,
which can be used to print the IR for all intermediate passes.
When running the added test file via:
```shell
$ python test/python/compile.py 2> tiny.stderr
```
the file `tiny.stderr` is about 700 KB.
This replaces the lowering of aten.cat with tensor.concat, allowing more
efficient handling of concatenations in downstream flows. The refbackend
populates concat decomposition patterns that can be used to recover the
previous lowering.
This is part 1 of 2, which will also include upstreaming the FX
importer. I started with ONNX because it forces some project layout
updates and is more self contained/easier as a first step.
Deviating somewhat from the RFCs on project layout, I made the following
decisions:
* Locating the `onnx_importer.py` into `torch_mlir.extras` as Maks
already has opened up that namespace and it seemed to fit. Better to
have fewer things at that level.
* Setup the build so that the root project only contains MLIR Python and
pure Python deps (like the importers), but this can be augmented with
the `projects/` adding more depending on which features are enabled.
* The default build continues to build everything whereas in
`TORCH_MLIR_ENABLE_ONLY_MLIR_PYTHON_BINDINGS=1` mode, it builds a
`torch-mlir-core` wheel with the pure contents only.
`onnx_importer.py` and `importer_smoke_test.py` are almost verbatim
copies from SHARK-Turbine. I made some minor local alterations to adapt
to paths and generalize the way they interact with the outer project. I
expect I can copy these back to Turbine verbatim from here. I also
updated the license boilerplate (they have the same license but slightly
different project norms for the headers) but retained the correct
copyright.
Other updates:
* Added the ONNX importer unit test (which also can generate test data)
in lit, conditioned on the availability of the Python `onnx` package. In
a followup once I know everything is stable, I'll add another env var
that the CI can set to always enable this so we know conclusively if
tests pass.
* Moved the ONNX conversion readme to `docs/`.
* Renamed CMake option `TORCH_MLIR_ENABLE_ONLY_MLIR_PYTHON_BINDINGS` ->
`TORCH_MLIR_ENABLE_PYTORCH_EXTENSIONS` and inverted the sense. Made the
JitIR importer and LTC options `cmake_dependent_options` for robustness.
For easier tracking of issues, sort the TOSA passing list. It is still
significantly smaller then the XFAIL list would be.
Resolves#2620, at least until the xfail list gets smaller than the
passing list.
Signed-off-by: Eric Kunze <eric.kunze@arm.com>
`AtenStackOp` defines this folder for list operand containing single
element:
```
OpFoldResult AtenStackOp::fold(FoldAdaptor adaptor) {
auto list = getOperand(0).getDefiningOp<PrimListConstructOp>();
if (!list || !list->hasOneUse() || list.getElements().size() != 1)
return nullptr;
return list.getElements()[0];
}
```
However, unlike `AtenCatOp`, `AtenStackOp` cannot be folded away for
single element list operand because the result from a stack operation
contains an additional dimension (of size 1, like expand_shape).
This PR removes the `AtenStackOp::fold` method, and adds an e2e test for
single element list input case, which fails on current `main` as
follows:
```
Unexpected outcome summary: (linalg)
****** Failed tests - 1 tests
FAIL - "TensorsStackSingleElementListModule_basic"
@ trace item #0 - call to "forward"
@ output of call to "forward"
ERROR: shape (torch.Size([10, 32])) is not equal to golden shape (torch.Size([10, 1, 32]))
```
Thanks Chris Lalau Keraly for the bug report.
saienduri
Xida Ren (Cedar)
Dave Liddell
Rob Suderman
Ilija Kalinić
Stella Laurenzo
Yuanqiang Liu
James Newling
Quinn Dawkins
zjgarvey
Franz Haniel
John Wu
Ze Zhang
Sungsoon Cho
lonely eagle
lisaliu1
Frederik Harwath
Aart Bik
Vivek Khandelwal
kumardeepakamd
Rik Huijzer
JianzheXiao
Eric Kunze
Sambhav Jain