The investigation is largely recorded in
https://github.com/llvm/torch-mlir/pull/2881, but this change allows us
to capture non-persistent buffers that were lifted as tensor constants
(after https://github.com/pytorch/pytorch/pull/118969 landed in upstream
PyTorch), and propagate them to `Torch` dialect as "frozen"
`torch.vtensor.literal`. I believe this patch should work with both
nightly and stable PyTorch, but will let CI confirm the same. Thanks
@stellaraccident for the valuable pointers and guidance.
---------
Co-authored-by: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
Various improvements on sparsity metadata:
(1) define single data structure for all sparsity related metadata
(2) handle batched dense dimensions, as well as dense subtensor
dimensions
(3) refine sparsity propagation for deeper networks
This PR introduces a sparse_jit wrapper that can run simple models with
sparse tensor inputs end-to-end. The implementation shows all required
components on modifying sparse tensor types with a 1:N relation on the
call sites. Two tests shows that the JIT runs end-to-end while computing
the correct results.
More details to follow (generalizing to COO and different ranks, as well
as support for *output* sparse tensors), but the general concepts are
all here now.
**_Update: Thanks to Rob, bump to proper LLVM/MLIR hash is done!_**
_**NOTE that all parameter passing changes are nicely done "downstream"
in MLIR, so very little changes are required in torch-mlir code
proper**_
---------
Co-authored-by: Franz Haniel <77495327+frafranz@users.noreply.github.com>
Co-authored-by: Franz Haniel <franz.haniel@amd.com>
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>
Adds an escape hatch from creating a DenseResourceElementsAttr for
single value tensors into DenseElementsAttr.
For 0d or 1element, splats are better as DenseElementsAttr. Don't use
DenseResourceElementsAttr for it
Fix for https://github.com/llvm/torch-mlir/issues/2765
The onnx docs say that you can't do shape inference using the in-memory
API for models > 2 GB. This fix replaces that API with the file-based
API. Since the new API generates an intermediate file, also added a
--keep switch to keep that file, which I delete by default.
---------
Co-authored-by: Dave Liddell <dliddell@xilinx.com>
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.
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.
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.
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>
Changes made during upstreaming:
* Removed comments attributing some copied code back to torch-mlir
(since it is now repatriated).
* Re-organized imports.
* Inlined RefMapping/RefTracker and TypeSubclassMap from an external
utility module.
* Added FxImporter class comments.
* Updated stack trace extraction to be fail safe.
* Added an entry-point for `import_frozen_exported_program` which uses
the shiny new upstream `torch.export.export()` API (versus the
lower-level/older API that Turbine is presently using). This
necessitated a small FX rewrite to line external state management up
with current conventions.
* Adapted one of Turbine's importer tests to go with this initial
submission. Turbine unfortunately has a lot of more-integration-ey
tests, and I would like to extract those as more of unit tests of the
importer features and upstream them that way vs trying to copy directly.
For now, one overall test with the initial submission gets us moving.
I acknowledge that there are some code quality things that could be
improved in this submission: this was authored over the course of many
months (and often via some trial and error). I would like to keep it
relatively converged with the downstream for the next few steps while
getting the test suite upstreamed. And then it will be easier to take a
hygienic pass through the code.
Including co-authors for contributors in the git log of the original
repository.
Co-authored-by: Ean Garvey <87458719+monorimet@users.noreply.github.com>
Co-authored-by: Avinash Sharma <aviator1994@gmail.com>
Co-authored-by: Arham Khan <arhammkhan@gmail.com>
Co-authored-by: brucekimrokcmu <kwangkyk@alumni.cmu.edu>
Co-authored-by: saienduri <77521230+saienduri@users.noreply.github.com>
Simple Python console script to import an ONNX protobuf to the torch
dialect for additional processing.
For installed wheels, this can be used with something like:
```
torch-mlir-import-onnx test/python/onnx_importer/LeakyReLU.onnx
```
Or from a dev setup:
```
python -m torch_mlir.tools.import_onnx ...
```
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.
This is a first step towards the structure we discussed here:
https://gist.github.com/stellaraccident/931b068aaf7fa56f34069426740ebf20
There are two primary goals:
1. Separate the core project (C++ dialects and conversions) from the
hard PyTorch dependencies. We move all such things into projects/pt1 as
a starting point since they are presently entangled with PT1-era APIs.
Additional work can be done to disentangle components from that
(specifically LTC is identified as likely ultimately living in a
`projects/ltc`).
2. Create space for native PyTorch2 Dynamo-based infra to be upstreamed
without needing to co-exist with the original TorchScript path.
Very little changes in this path with respect to build layering or
options. These can be updated in a followup without commingling
directory structure changes.
This also takes steps toward a couple of other layering enhancements:
* Removes the llvm-external-projects/torch-mlir-dialects sub-project,
collapsing it into the main tree.
* Audits and fixes up the core C++ build to account for issues found
while moving things. This is just an opportunistic pass through but
roughly ~halves the number of build actions for the project from the
high 4000's to the low 2000's.
It deviates from the discussed plan by having a `projects/` tree instead
of `compat/`. As I was thinking about it, this will better accommodate
the follow-on code movement.
Once things are roughly in place and the CI passing, followups will
focus on more in-situ fixes and cleanups.
NonValueSemantic Ops like Add_, div_, etc. expect result DType to be the
same as the first input. However, current implementation would result in
wrong result type for case like:
```python
a = torch.randn(3, 3).half() # float16
b = torch.randn(3, 3) # float32
a += b # i.e. torch.ops.aten.add_(a, b)
```
torch expects `a` to be float16, but dtype refinement would infer
float32 type, since it's replaced by `aten.add`.
Attempt to solve https://github.com/llvm/torch-mlir/issues/2490
Changes for Non Value Semantic Ops having the
`IsTrailingUnderscoreInplaceVariant` trait :
- AnyTorchTensorType -> Torch_NonValueTensorType
- AnyTorchOptionalTensorType -> AnyTorchOptionalNonValueTensorType
- AnyTorchListOfOptionalTensorType ->
AnyTorchListOfOptionalNonValueTensorType
- AnyTorchListOfTensorType -> AnyTorchListOfNonValueTensorType
Created three new tensor types for optional and list non value tensors.
Add aten.isclose op
Add its torch-to-tosa lowering
Update the TorchToTosa/basic.mlir tests
To test e2e tosa lowering:
`python -m e2e_testing.main -v -c=tosa`
---------
Co-authored-by: Ze Zhang <ze.zhang@getcruise.com>
Add aten.unflatten.int op
Add its torch-to-tosa lowering
Update the TorchToTosa/basic.mlir tests
To test e2e tosa lowering:
`python -m e2e_testing.main -v -c=tosa`
---------
Co-authored-by: Ze Zhang <ze.zhang@getcruise.com>
Add linspace/cumprod/roll ops to ODS and add shape inference functions
to make it work with LTC.
Also, add some tensor utils to LTC library for searching for non-detach
copy nodes.
Set PyTorch and TorchVision version to nightly release 2023-09-28.
aten.baddbmm changes done because upstream PyTorch has now added
support for fp16 gemm on CPU.
Refer: 9399e0b1ff
This commit adds to the lowering of `aten.view` handling for the
following cases:
- `(..., a.size(i))` -> `(..., a.size(i), 1, ..., 1)`
- `(..., a.size(i), 1, ..., 1)` -> `(..., a.size(i))`
Fixes: https://github.com/llvm/torch-mlir/issues/2448
Set PyTorch and TorchVision version to nightly release 2023-09-26.
aten._convolution.deprecated changes done because upstream PyTorch has
now added support for fp16 native convolution on CPU.
Refer: 7c9052165a
Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
Making the same PR with #2457, as I accidentally thought the review was already made and merged it (reverted).
Add decompose empty_strided op.
Referring to #1776, this decomposition op only supports default stride values, because accessing the tensor or indexing over that, the indices are determined by the strides.
In MLIR, this is not implicitly supported but assumes that the strides are default while iterating over the tensor.
We just have to do this: I ran into an issue today where I needed to make a one line patch to stablehlo to work around a compiler issue, and it is completely unapparent how to do so given that the mlir-hlo repo is a read-only export and is at the tail end of a multi-week integration chain from the open-source stablehlo repo.
We've discussed this often enough and gotten +1 from everyone that they are ok with taking the e2e testing hit if it becomes necessary: It is necessary as the current situation is unmanageable.
Looking at it, I expect it wouldn't actually be very difficult to build a little runner binary out of the stablehlo interpreter and subprocess call that in order to get the testing coverage back. I leave that as an exercise to the users of this part of the stack and recommend following the breadcrumbs from the deleted python/torch_mlir_e2e_test/stablehlo_backends/linalg_on_tensors.py file and the main.py changes.
Note that I am pointing us at a stablehlo fork for the moment until it is apparent that we don't need to carry any local patches to it. We can update this in a few days if everything is clear.
Sambhav Jain
Aart Bik
saienduri
Daniel Garvey
Dave Liddell
Rob Suderman
Stella Laurenzo
Yuanqiang Liu
Zhekun(Josh) Zhang
xiaolou86
JianzheXiao
Sarthak Gupta
Ze Zhang
Ramiro Leal-Cavazos
Vivek Khandelwal
Jae Hoon (Antonio) Kim
Bruce Kim
Gleb Kazantaev
David Gens
Arham Khan
Jiawei Wu