This patch replaces all MHLO operations with their StableHLO
counterparts and adds a validation pass to ensure that no MHLO operations
remain before translating all Stablehlo operations to the MHLO dialect
for further lowering to the Linalg dialect.
This patch also updates all lit tests so that they refer to the
`convert-torch-to-stablehlo` pass and so that they check for StableHLO
operations.
Previously, torchvision had not released WHL files for Python v3.8,
causing failures in torch-mlir python package builds, so we had disabled
building for Python v3.8.
Now that the WHL files are back, this patch re-enables v3.8 builds.
- Use v3 of actions/checkout, since the version we use (v2) uses
Node.js 12, which is deprecated by GitHub.
- Source the PowerShell venv sctipt (instead of the bash sript) since
the calling script is a PowerShell script. Without this, the build
doesn't use venv at all.
- Make the build dependencies in whl-requirements.txt (used by
setup.py) match those in requirements.txt. To that end, this patch
creates a build-requirements.txt that is referenced by
requirements.txt and whl-requirements.txt.
* [custom op] Generalize shape library logic to work with dtypes
This commit generalizes the shape library logic, so that dtype rules
for ops can also be expressed using the same mechanism. In other
words, each op can now have a shape function and a dtype function
specified in Python that is imported during lowering to calculate the
shapes and dtypes throught a program. For more information about how
to specify a dtype function, see the updated
`docs/adding_a_shape_and_dtype_function.md`.
For those not familiar with how the shape library works, the file
`docs/calculations_lib.md` provides an overview.
This was an experimental attempt at rolling out own op-by-op executor
with `__torch_dispatch__`, but it proved difficult to make it robust.
Op-by-op execution is very easy to implement robustly now with the
PyTorch 2.0 stack, so we don't need eager_mode.
Downstream users were using eager_mode to implement lockstep numerical
accuracy debuggers. We implemented the same functionality with
TorchDynamo in https://github.com/llvm/torch-mlir/pull/1681 so now there
is not much reason to continue maintaining it.
This more accurately reflects what it is. The previous name was
conflating the use of RefBackend (which `linalg`, `tosa`, and `mhlo`
configs all use) with the use of the linalg backend (e.g. TorchToLinalg).
This conflation was artifically giving the linalg backend a "privileged"
position, which we want to avoid. We still keep it as the default
backend, and it remains the most complete, but at least there's not
artificial boosting.
This adds a basic e2e Config for TorchDynamo using
Linalg-on-Tensors/RefBackend.
But TorchDynamo is pretty orthogonal to
various other pieces, so it should compose nicely with variations like:
- Switching out all the backends (Linalg-on-Tensors, TOSA, MHLO)
- PyTorch functionalization and decompositions
- Taking the example inputs and compiling with all dynamic or all static
shapes without duplicating tests.
This adds it to the CI, but there are still a lot of XFAIL's.
This also adds a helper `from torch_mlir.dynamo import
make_simple_dynamo_backend` which simplifies some of the steps for
making a Torch-MLIR-based TorchDynamo backend. We include "simple" in
the name because we are going to be exploring various things next from
the long-term roadmap.
The next steps are:
- Burn down all the XFAIL's.
- Start working on the pieces from the [long-term roadmap](https://github.com/llvm/torch-mlir/blob/main/docs/long_term_roadmap.md).
- Add functionalization/decompositions into the TorchDynamo flow and
remove reliance on the current Torch-MLIR "frontend".
- Write a pure-Python direct FX->MLIR importer.
- Hook up the new PyTorch symbolic shape stuff.
- Explore PrimTorch decompositions for simplifying backends.
We want each build to be reproducible regardless of prior builds and
prior package installations, but pip, by default, uses cached packages
from previous invocations of `pip install`. As a result, the incorrect
dependencies downloaded in the RollPyTorch workflow in the main
repository cannot be reproduced in private forks of the repository. To
resolve this problem, this patch adds a `--no-cache-dir` flag to pip, so
that it fetches and inspects each requested package independent or prior
installations.
Before this patch, the update_shape_lib.sh and update_torch_ods.sh
scripts only worked on in-tree builds, which implied that the
RollPyTorch action was forced to run the longer-running in-tree build.
As a result of this patch, we should be able to run through the basic
checks in the RollPyTorch action faster, while running the full suite of
tests off the critical path.
The key change in this patch is that the update scripts now look for the
directory that is most recently modified between in-tree or out-of-tree
build directories. The change also correctly handles the case when only
one of the two directories exists.
This patch makes a few small, but key, changes to enable ccache on
Windows. First, it replaces the hendrikmuhs/ccache-action action with
command line invocations to the ccache binary, since the action has two
bugs, one of which causes CI to refer to different ccache artifacts
before versus after the build on Windows whereas the other bug can
sometimes cause the action to incorrectly infer that the cache is empty.
Second, this patch slightly alters the cache key, so that our old cache
artifacts, which have grown too big, are eventually discarded in favor
of the new, smaller cache artifacts. Along the way, this patch also
keeps the RollPyTorch's cache artifact separate from the regular build's
cache artifact so as to keep these artifacts small, and also because the
RollPyTorch action is off the critical path for most contributors.
Finally, this patch makes small changes to the CMake file so that on
Windows, the ccache binary is added as a prefix, as recommended on the
[ccache Wiki](https://github.com/ccache/ccache/wiki/MS-Visual-Studio).
This patch updates the build_linux_packages.sh script so that when
PyTorch needs to be built from source, it is built _before_ building
LLVM and before building Torch-MLIR. The rationale behind this change
is that previously, when the PyTorch build was triggered through the
Torch-MLIR build, the PyTorch compilation added more entries to the
ccache artifacts. However, since we cache the PyTorch _binary_ (i.e.
the WHL file), there is no need to add the PyTorch compilation to the
ccache artifacts. By removing the PyTorch compilation files, we keep
the ccache artifact size small, thus reducing the number of evictions
when we exceed GitHub's allowed limit.
This commit removes almost all of the valsem ops, since the value
semantics version of the ops now exist in PyTorch. The only op missing
is `aten.bernoulli_.float`. In addition, this commit also simplifies
the implementation of `aten.fill.Scalar` by moving it to the pattern
that converts elementwise ops.
Whether or not the PyTorch build is cached should not affect the success
of the Torch-MLIR build, but based on the existing code, a build may
fail if the `TM_PYTORCH_INSTALL_WITHOUT_REBUILD` variable was set but
the build cache doesn't exist.
Although that variable is set by CI upon a cache hit, nuances of
Github's caching behavior can create situations where the coupling
between `TM_PYTORCH_INSTALL_WITHOUT_REBUILD` and the cache lookup fails.
Specifically, a branch other than our default branch (`main`) may create
the cache entry, but because Github doesn't share this cache entry with
builds running on the `main` branch, the `main` branch build tries to
create it's own cache entry. However, since cache identifiers are
unique and because caches are immutable, the caching step running in the
`main` branch appears to create an invalid cache entry (of 233 bytes,
instead of the expected ~60 MB).
Consequently, subsequent builds observe a cache "hit", since caches
created by the `main` branch are shared with all other branches, but
because this cache entry is invalid (since it doesn't actually contain
the ~60 MB PyTorch WHL file), the builds fail.
One workaround would be to let only the `main` branch create caches, but
in doing so, we would also prevent other branches from _reading_ the
cache, making the builds in those branches terribly slow.
So this patch uses a different workaround, which is to check whether the
PyTorch WHL file exists, even if the build observed a cache hit. If the
file doesn't exist, even if it was a purported cache hit, the code
builds PyTorch from source, which is probably intuitive.
A longer term fix will follow, after a discussion with the wider team.
Without this patch, CI logs contained the line:
-- Linker detection: GNU ld
GNU ld is notoriously slow at linking large binaries, so this patch
swaps GNU ld with the LLVM linker.
Since the linker invocation is driven through the compiler, perhaps the
best way to use the LLVM linker is to tell the compiler which linker
binary to use. This patch adds the `-fuse-ld=lld` flag to all Linux
builds of Torch-MLIR in CI to make it use lld.
* ci: cache PyTorch source builds
This patch reduces the time spent in regular CI builds by caching
PyTorch source builds. Specifically, this patch:
1. Makes CI lookup the cache entry for the PyTorch commit hash in
pytorch-version.txt
2. If lookup was successful, CI fetches the previously-generated WHL
file into the build_tools/python/wheelhouse directory
3. CI sets the `TM_PYTORCH_INSTALL_WITHOUT_REBUILD` variable to `true`
4. The build_libtorch.sh script then uses the downloaded WHL file
instead of rebuilding PyTorch
* ci: warm up PyTorch source cache during daily RollPyTorch action
This patch makes the RollPyTorch action write the updated WHL file to
the cache, so that it can be later retrieved by CI that runs for each
PR. We deliberately add the caching step to the end of the action since
the RollPyTorch action never needs to read from the cache, although
executing this step earlier in the process should not cause problems
either.
We originally added these to help bring up more complex models with
heavier dependencies. However, over time it has become clear that these
models usually require more than just heavier dependencies -- they often
require a nontrivial amount of "one-off" code to extract the relevant
parts of the model and compile them. This is not a good fit for a
component in the core Torch-MLIR repo.
However, in the community, nod.ai has developed the ["Shark
Tank"](https://github.com/nod-ai/SHARK/tree/main/tank) which has all the
appropriate code to wrangle these models and organize them. We intend to
more heaviliy lean on that as a community and improve the symbiosis
there to serve the role that these heavydep tests were meant to play.
* build: disable LTC again so that we can bump PyTorch version
When built using PyTorch's master branch, the LTC code has been failing
to build for a few days. As a result, the PyTorch version referenced by
Torch-MLIR is stalled to the one from October 4th.
In an effort to advance to PyTorch version, this patch disables LTC, and
a subsequent patch will advance the PyTorch version.
* update PyTorch version to 1.14.0.dev20221010
Also disables the `UpSampleNearest2dDynamicFactor_basic` e2e test, since
the (PyTorch) oracle differs from the computed value for both the
refbackend and the eager_mode backends.
Instead of letting the auto-update script either fail because of script
errors or letting it commit bad versions, this patch makes the update
process manual, for now. Once the script stabilizes, I will its
re-enable periodic execution.
Updating the PyTorch version may break the Torch-MLIR build, as it did
recently, since the PyTorch update caused the shape library to change,
but the shape library was not updated in the commit for updating
PyTorch.
This patch introduces a new default-off environment variable to the
build_linux_packages.sh script called `TM_UPDATE_ODS_AND_SHAPE_LIB`
which instructs the script to run the update_torch_ods.sh and
update_shape_lib.sh scripts.
However, running these scripts requires an in-tree build and the tests
that run for an in-tree build of Torch-MLIR are more comprehensive than
those that run for an out-of-tree build, so this patch also swaps out
the out-of-tree build for an in-tree build.
Prior to this patch, the release process (`pip wheel`) retrieved
dependencies from the pyproject.toml file, which specified a version of
PyTorch that defaulted to the most recent nightly release. Instead, we
want the release process to use the same pinned PyTorch version as the
development build of PyTorch.
Since TOML files can't reference the pytorch-requirements.txt file, this
patch puts the dependencies from pyproject.toml into
whl-requirements.txt, which references pytorch-requirements.txt.
`git diff` does not work by default on untracked files. Since the
function `_check_file_not_changed_by` stores the new generated file in
an untracked file, `git diff` was not catching any modifications in
the new generated file. This commit adds the flag `--no-index` to make
`git diff` work with untracked files.
This patch fetches the most recent nightly (binary) build of PyTorch,
before pinning it in pytorch-requirements.txt, which is referenced in
the top-level requirements.txt file. This way, end users will continue
to be able to run `pip -r requirements.txt` without worrying whether
doing so will break their Torch-MLIR build.
This patch also fetches the git commit hash that corresponds to the
nightly release, and this hash is passed to the out-of-tree build so
that it can build PyTorch from source.
If we were to sort the torch versions as numbers (in the usual
descending order), then 1.9 appears before 1.13. To fix this problem,
we use the `--version-sort` flag (along with `--reverse` for specifying
a descending order). We also filter out lines that don't contain
version numbers by only considering lines that start with a digit.
As a matter of slight clarity, this patch renames the variable
`torch_from_src` to `torch_from_bin`, since that variable is initialized
to `TM_USE_PYTORCH_BINARY`.
Co-authored-by: powderluv <powderluv@users.noreply.github.com>
This adds a very long and obnoxious option to disable crashing tests.
The right fix here is to use the right multiprocessing techniques to
ensure that segfaulting tests can be XFAILed like normal tests, but we
currently don't know how to implement "catch a segfault" in Python
(patches or even just ideas welcome).
Motivated by #1361, where we ended up removing two tests from *all*
backends due to a failure in one backend, which is undesirable.
We added both ipc=host and explicit ulimits. This _may_ be causing slow downs on GHA. Remove the ulimit setting still passes all the CI tests locally. `--ipc=host` is still required.
The new logic has the following benefits:
1. It does not clobber the working tree state. We expect testing to not
change the work tree.
2. It correctly handles the case where a user has changes to the
generated files, but hasn't checked them in yet (this happens
frequently when adding new ops).
Gets both CI and Release builds integrated in one workflow.
Mount ccache and pip cache as required for fast iterative builds
Current Release docker builds still run with root perms, fix it
in the future to run as the same user.
There may be some corner cases left especially when switching
build types etc.
Docker build TEST plan:
tl;dr:
Build everythin: Releases (Python 3.8, 3.9, 3.10) and CIs.
TM_PACKAGES="torch-mlir out-of-tree in-tree"
2.57s user 2.49s system 0% cpu 30:33.11 total
Out of Tree + PyTorch binaries:
Fresh build (purged cache):
TM_PACKAGES="out-of-tree"
0.47s user 0.51s system 0% cpu 5:24.99 total
Incremental with ccache:
TM_PACKAGES="out-of-tree"
0.09s user 0.08s system 0% cpu 34.817 total
Out of Tree + PyTorch from source
Incremental
TM_PACKAGES="out-of-tree" TM_USE_PYTORCH_BINARY=OFF
1.58s user 1.81s system 2% cpu 1:59.61 total
In-Tree + PyTorch binaries:
Fresh build and tests: (purge ccache)
TM_PACKAGES="in-tree"
0.53s user 0.49s system 0% cpu 6:23.35 total
Fresh build/ but with prior ccache
TM_PACKAGES="in-tree"
0.45s user 0.66s system 0% cpu 3:57.47 total
Incremental in-tree with all tests and regression tests
TM_PACKAGES="in-tree"
0.16s user 0.09s system 0% cpu 2:18.52 total
In-Tree + PyTorch from source
Fresh build and tests: (purge ccache)
TM_PACKAGES="in-tree" TM_USE_PYTORCH_BINARY=OFF
2.03s user 2.28s system 0% cpu 11:11.86 total
Fresh build/ but with prior ccache
TM_PACKAGES="in-tree" TM_USE_PYTORCH_BINARY=OFF
1.58s user 1.88s system 1% cpu 4:53.15 total
Incremental in-tree with all tests and regression tests
TM_PACKAGES="in-tree" TM_USE_PYTORCH_BINARY=OFF
1.09s user 1.10s system 1% cpu 3:29.84 total
Incremental without tests
TM_PACKAGES="in-tree" TM_USE_PYTORCH_BINARY=OFF TM_SKIP_TESTS=ON
1.52s user 1.42s system 3% cpu 1:15.82 total
In-tree+out-of-tree + Pytorch Binaries
TM_PACKAGES="out-of-tree in-tree"
0.25s user 0.18s system 0% cpu 3:01.91 total
To clear all artifacts:
rm -rf build build_oot llvm-build libtorch docker_venv
externals/pytorch/build
We use it for more than TorchScript testing now. This is a purely
mechanical change to adjust some file paths to remove "torchscript".
The most perceptible change here is that now e2e tests are run with
```
./tools/e2e_test.sh
instead of:
./tools/torchscript_e2e_test.sh
```
Bumps the shape library:
- Updates the function signature for aten.arange.start_step
- upstream_shape_functions.mean_dim -> upstream_shape_functions.sum_mean_dim
* Replace CHECK_EQ with TORCH_CHECK_EQ
* Check value of TORCH_MLIR_USE_INSTALLED_PYTORCH during LTC build
* Update LTC XFAIL with NewZerosModule ops
* Explicitly blacklist _like ops
* Automatically blacklist new_/_like ops
* Prune away unused Python dependencies from LTC
* Add flag to disable LTC
* Autogen dummy _REFERENCE_LAZY_BACKEND library when LTC is disabled
* Implement compute_shape_var
* Removed Var tests from XFAIL Set
* XFAIL tests using _local_scalar_dense or index.Tensor
* Add StdDim tests to XFAIL set
* Autogen aten::cat
* Changed Example MLIR backend to Reference MLIR backend
* Moved reference_ltc_backend into csrc
* Merged sys_utils.h
* Renamed reference_ltc_backend to reference_lazy_backend
* Addressed review comments
* Update docs with new library name
* Removed _REFERENCE_LAZY_BACKEND from .gitignore
* Added reference_lazy_backend to the TorchMLIRPythonModules dependency list
Fixed typo in `ltc_examples.md`
Missed instance where `ltc_backend` was used instead of `lazy_backend`.
* Update native function definitions
* Add ops to support bert lowering
- Add empty_strided and as_strided
- Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy)
- Check for composite implicit ops and add device data IR
- Also fix codegen for functionalization
* Add autogen to CMakeList
* Remove PyTorch submodule
* Reduced BERT model size
* Print Mark Step status in Torch MLIR LTC debug string
* Apply fixes to work with latest upstream/main
- Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode
Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor
* Update shape inference functions
- Fixed compute_shape_native_batch_norm when mean and var are uninitialized
Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels.
- Implemented compute_shape_mul
- Fixed bug in reshape shape inference error message
* Get MLIR backend more consistent with TS backend
- Remove LazyNativeFunctions::_unsafe_view from autogen
- Blacklist ops to make JIT graph more like output of TS backend
- Print graph when SSA value has mismatch of types and results
- Remove normalize_index from LazyShapeInference
- Fix seeds for LTC example models
* Update and clean up shape inference functions
- Prune shape inference functions
- Add shape inference function for GenerateSlice
- Add shape inference function for GenerateCopy
Co-authored-by: Henry Tu <henry.tu@cerebras.net>
* Save InputOutputAliases to TorchMlirComputation
* Implement GetResultShape for TorchMlirLoweringContext
* Use optional return type for GetResultShape
* Remove support for aten::detach
With this op enabled, tensors were being copied, which resulted in incorrect aliasing.
* Add newline before printing I/O alias mapping
* Changed printout to use "Input param" as label instead of "Input"
* Remote shape inference function for aten::detach
* Moved implementation of SetUpAlias to MlirLoweringContext
As part of this change, TorchMlirComputation has been moved to the end of mlir_lowering_context.h so that it can access some new structs in TorchMlirLoweringContext
* Use updated PyTorch API
* Remove GetResultShape
Complements this upstream PyTorch PR: pytorch/pytorch#75828
This PR adds support for mapping input and output tensors which alias each other. (e.g. maps input weight tensor in parameter to the same tensor in output after a training iteration)
MLIR:
func @graph(%arg0: !torch.vtensor<[1,5],f32>, %arg1: !torch.vtensor<[1],si64>, ..., %arg6: !torch.vtensor<[10,5],f32>, %arg7: !torch.vtensor<[10],f32>, ...) {
...
return %arg0, %arg1, %17, %23, ... : !torch.vtensor<[1,5],f32>, !torch.vtensor<[1],si64>, !torch.vtensor<[10,5],f32>, !torch.vtensor<[10],f32>, ...
}
Input/Output Alias Mapping:
Output: 0 -> Input: 0
Output: 1 -> Input: 1
Output: 2 -> Input: 6
Output: 3 -> Input: 7
The aten::detach op has also been disabled in this PR to fix the issue of tensors not aliasing properly due to copying.
* Added JIT to MLIR lowering
Lowering to JIT is performed in a way similar to how it's done in the TS LTC backend. After a jit::Graph is constructed, it gets converted to a jit::Function, which is fed into the existing utility to generate an MlirModule in torch-mlir.
* Renamed `csrc/backend` to `csrc/base_lazy_backend`
This enables building Pytorch from source in the CI.
The build should mostly hit the ccache.
Release builds will follow once we have some runtime on the CI.
Remove all the libtorch downloads. If the user sets
-DTORCH_MLIR_USE_INSTALLED_PYTORCH=OFF then just build from src.
Doesn't change developer workflow since we still default to local
PyTorch versions.
TEST: Build and verify all tests (except one xfail quant) pass on linux
Found while trying to build torch-mlir on an AArch64 Linux VM, worth
a belts and braces to prevent such cases.
Change-Id: I89c6fccb62e666dbda0d9acac2d0ee43c2899e9b
On my local machine, `unzip` didn't exist (producing a "command not
found" error), but CMake ignored the error. Although the build did
succeed (because it found a previously-built version of libtorch), it
seems better to abort builds on such failures, so this patch checks the
return code of all external process invocations.
Along similar lines, this patch also updates the shell scripts in
`build_tools` to extensively use double-quoting to prevent unintentional
word splitting or globbing. Since some of the scripts execute `rm`
while using shell variables, this patch also adds the preamble `set -u`
to abort execution if an undefined variable is referenced, so that we
reduce the chances of executing `rm -rf /` if the path expression
happens to refer to an undefined variable.
Add an option to cache libtorch/ releases if you don't want to
download the latest. Add an option to enable source builds.
TESTS:
macOS: verify with / without cache downloads
verify source builds -- shared and static
Linux: Build Tests and Release builds
The MacOS builders are having linking trouble with the extension library.
Until it's fixed, all support for op extensions is disabled. It should be
easy to restore once the issue is resolved.
PyTorch allows new operators to be registered dynamically in modules.
Torch-mlir already makes it fairly straightforward to add support for
new operators, and this commit just extends that support to allow new
PyTorch ops to come from a external module.
This does *not* allow ops to be dynamically loaded into torch-mlir.
Torch-mlir must still be compiled with support built-in.
Add a `_torch_mlir_custom_op_example` subpackage to `torch_mlir` which
registers an demonstration op. It will not be imported by default when
importing torch_mlir. It's strictly for testing and documentation.
Adds an end-to-end test for the `torch_mlir_custom_op_example::identity` op.
With all these changes, we should now be actively testing PyTorch extension
support with all future patches.
1. With the help of `make_fx` we are able to get the full training graph
with weight updates.
2. NeuralNet_training passes. Bert_training passes after cherry-picking
https://github.com/llvm/torch-mlir/pull/844.
3. TODO: Remove the functorch's dependency after make_fx moves to
pytorch core.
powderluv
Ashay Rane
Gleb Kazantaev
Daniel Ellis
Ramiro Leal-Cavazos
Sean Silva
Tanyo Kwok
Jae Hoon (Antonio) Kim
Henry Tu
Quinn Dawkins
nithinsubbiah
Sambhav Jain
Antonio Kim
Ramana Radhakrishnan
Maksim Levental
Bob Adolf
Prashant Kumar
Vivek Khandelwal
