Comments (11)
Right! I'll keep an eye on the progress :)
from torch-mlir.
Hey Maxi, you didn't miss any steps.
For the first error, the short answer is that aten.cos.out
is not supported yet. In order to support this, we need to update torch_ods_gen.py and run build_tools/update_torch_ods.sh
to generate the ods and lower this op in TorchToLinalg.
For the second error, I can reproduce on my end but I am not sure what's the problem either. @silvasean is more experienced with this project but he's on vacation at the moment. Sry!
from torch-mlir.
@cathyzhyi
Hello Cathy,
Thanks for the reply! Is there any code submission similar to this kind of operator support that I can take a look at? Just want to get to know the whole lowering pass :)
from torch-mlir.
Hey Maxi, I can't find a good example of one single PR containing all the supports for one operator but you could use aten::relu
as an example. Running build_tools/update_torch_ods.sh
generates a file called JITOperatorRegistryDump.txt
which contains all the operators with schema. You can find aten::relu
and aten::cos.out
in there. After that, you could try to follow example of aten::relu
and AtenReluOp
in the codebase. Please let me know if you have any issues.
from torch-mlir.
@cathyzhyi
Hello Cathy,
I have been reading code and running some test...
I got some questions wanna ask:
- What is your current status and plan for supporting the common deep learning models? For example, Resnet 50. I see that there is a Resnet18 test case (torchscript e2e test) but it fails during the lowering phase (I am using IREE backend). And I see that silva is working on supporting the Resnet model here: #212
- Are you considering merging into the IREE project as one of its frontends in the future? It would be great because currently IREE already has Tensorflow frontend.
Thanks!
from torch-mlir.
Sean and Cathy can comment on their roadmap for supporting more models -- it is being actively worked on.
Regarding (2) - I am working this month on the python deployment machinery so that both npcomp and iree can be easily installed and the intent is that this will allow npcomp to use iree as a backend directly by regular users. It may take some iterations to get right as these are complicated dependencies but I expect this to start to become usable for folks within the next few months. IREE is hoping that the npcomp project can use it out of the box as an out of tree backend without the iree team needing to maintain an in tree PyTorch frontend itself.
from torch-mlir.
Great to know that, looking forward to it becomes more usable!
from torch-mlir.
What is your current status and plan for supporting the common deep learning models? For example, Resnet 50. I see that there is a Resnet18 test case (torchscript e2e test) but it fails during the lowering phase (I am using IREE backend). And I see that silva is working on supporting the Resnet model here: Progress on lowering ResNet #212
yea, we are working on making Resnet18 work which is one of the targets in the Q3 roadmap. Most of the recent PRs are relevant to this item.
from torch-mlir.
FWIW, the .out variants are likely best handled in ReduceOpVariants. They can all be systematically handled there.
from torch-mlir.
@hunbssfy btw resnet_18 to iree now works, just edit the frontends/pytorch/examples/ script and set iree as backend.
from torch-mlir.
@hunbssfy btw resnet_18 to iree now works, just edit the frontends/pytorch/examples/ script and set iree as backend.
Yeah, I did the same modification and it works :)
from torch-mlir.
Related Issues (20)
- Out of tree build in conda environment / overdependency problem HOT 5
- FX Importer error: `Could not deduce type from value info: tensor_meta=None, val=s0, sparsity=None` HOT 4
- LTC build error after PT bump: multiple definition of `torch::lazy::compute_shape_arange` HOT 1
- [ONNX] Dynamic size support for the onnx.GreaterOrEqual operator. HOT 1
- [MLIR][TORCH] Onnx.MaxPool fails on dynamic sizes
- [ONNX] onnx.Loop HOT 5
- [ONNX] MatMulInteger Conversion Fails For Arg Rank != 2 HOT 2
- [TORCH] `AtenFloorDivideOp` truncates instead of flooring which doesn't match `torch.floor_divide()` . HOT 1
- [TORCH] add support for torch.aten.floor_divide.Scalar
- buildAndTest.yml lacking trigger causes a bunch of notifications HOT 1
- [BUG][MLIR][BERT] examples/torchscript_stablehlo_backend_tinybert.py
- Slice Folder Fails Result Type Match Assert
- [onnx][torch][linalg] Support of corner align mode for gridsampler
- [e2e testing] make e2e_testing/main.py print ir / more informative error messages
- Build failed on x86 MacOS
- `make_simple_dynamo_backend` does not work for single argument models. HOT 2
- Compilation of Quantized Model Failed HOT 1
- Missing dtype support for OnnxToTorch HOT 3
- Strange issue related to `pyproject.toml` when building release wheels HOT 4
- iree-compiler failed to compile gpt2
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from torch-mlir.