Jiawei Wu issues

Results 8 issues of


                                            Jiawei Wu

[MHLO] Add configurable op decomposition rules

See RFC #999 We register some options in `TorchLoweringPipeline` and `DecomposeComplexOpsPass` to make the decomposition rules configurable. Co-authored-by: Bairen Yi Co-authored-by: Jiawei Wu Co-authored-by: Tianyou Guo Co-authored-by: Xu Yan Co-authored-by:...

`graph.HeteroGraph` Error happened when running example

Recently, I run the example provided with the following command: ``` dglke_train ``` By default, it will use FB15k knowledge graph. However, it seems to construct a graph.HeteroGraph which is...

[FxImporter]: support parsing nodes which return list

As described in title. The context is that I'm using torch.export + torch_mlir.fx_importer to trace a Mixtral MoE layer. It has aten.unbind.int op, which returns a list of tensor as...

[Torch Dialect] emit aten.nonzero op, recompose aten.index.Tensor with aten.nonzero_numpy to aten.masked_select

In detail, this PR includes: 1. emit `aten.nonzero` and `aten.nonzero_numpy` op; 2. add missing shape infer function for `aten.masked_select` op 3. add recomposition rule for `aten.index.Tensor` op. When it's indices...

[torch-frontend] add stablehlo IRs for Mixtral model.

In this PR, we provide stablehlo IR of a single Mixtral decoder layer using ByteIR stack. The IR is elided by `--mlir-elide-resource-strings-if-larger=1000` option, so not all dialect resources storing the...

enhancement

[Torch Dialect] emit aten.index_add op and decompose it to scatter.add op

[stablehlo] support dynamic-shaped index in stablehlo conversion for aten.index-like ops

[torch-frontend] fx export FakeTensor weight as SplatElementsAttr.

In the previous PR, we export FakeTensor as a DenseElementsAttr with random values. We change to create a splat DenseElementsAttr in this PR, as FakeTensor doesn't care about tensor values...