Chi_Liu

@manishghop Find the error for torch-mlir "aten::_scaled_dot_product_flash_attention_for_cpu". It's an op rename from "aten::_scaled_dot_product_flash_attention" in pytorch. The decompose of `aten::_scaled_dot_product_flash_attention_for_cpu` is added in pytorch 2 days ago in https://github.com/pytorch/pytorch/pull/117390 and https://github.com/pytorch/pytorch/pull/117097...

I manually apply the https://github.com/pytorch/pytorch/pull/117390 and https://github.com/pytorch/pytorch/pull/117097 in my local shark.venv And run the export_chatglm.py but got **Bug**: ``` Traceback (most recent call last): File "/home/chi/src/SHARK/../test/chatglm/export_chatglm2.py", line 121, in ts_graph...

After the previous brevitas solution, I got another **bug**: ``` [DEBUG] Compiling torchscript graph Traceback (most recent call last): File "/home/chi/src/SHARK/../test/chatglm/export_chatglm2.py", line 131, in module = torch_mlir.compile( ^^^^^^^^^^^^^^^^^^^ File "/home/chi/src/torch-mlir/build/tools/torch-mlir/python_packages/torch_mlir/torch_mlir/__init__.py",...

Same numeric failure for conv op in convnext_nano.d1h_in1k model https://github.com/nod-ai/SHARK-TestSuite/pull/403 onnx.mlir ``` module { func.func @main(%arg0: !torch.vtensor, %arg1: !torch.vtensor, %arg2: !torch.vtensor) -> !torch.vtensor attributes {torch.onnx_meta.ir_version = 10 : si64, torch.onnx_meta.opset_version...

`torch-mlir-opt --convert-torch-to-stablehlo /nodclouddata/chi/src/models/t5/slicecopy/EmbeddingBagDenseBackwardsModule.mlir ` ``` #loc = loc(unknown) module attributes {torch.debug_module_name = "EmbeddingBagDenseBackwardModule"} { func.func @forward(%arg0: !torch.vtensor loc(unknown), %arg1: !torch.vtensor loc(unknown), %arg2: !torch.vtensor loc(unknown), %arg3: !torch.vtensor loc(unknown)) -> !torch.vtensor {...

https://github.com/iree-org/iree/pull/17692

``` %9 = torch.operator "onnx.Reshape"(%arg0, %8) : (!torch.vtensor, !torch.vtensor) -> !torch.vtensor ``` will be lower to ``` %42 = torch.aten.reshape %arg0, %41 : !torch.vtensor, !torch.list -> !torch.vtensor ``` @PhaneeshB Since...

will be fixed by https://github.com/llvm/torch-mlir/pull/3693