Sayan Saha
Sayan Saha
> @sahas3 can you rebase this PR again to solve the conflicts, please? Hi @leandron , looks like https://github.com/tensorflow/tensorflow/pull/61892 which is recently merged implemented the same change. So I am...
> @sahas3 Please rebase your PR, so that the CI can be run. Rebased PR and CI is clean. Thanks!
I gave this a shot by reusing [https://github.com/llvm/torch-mlir/blob/main/build_tools/python_deploy/build_windows_ci.sh](https://github.com/llvm/torch-mlir/blob/main/build_tools/python_deploy/build_windows_ci.sh) which is used for building the windows wheels in [torch-mlir-release](https://github.com/llvm/torch-mlir-release) repo. The build worked fine but [testing fails](https://github.com/sahas3/torch-mlir/actions/runs/13413554894/job/37468876174#step:10:75) because `signal.SIGALRM` (used for...
> I’ll update this PR to produce an error if any tosa.const is encountered while running --tosa-convert-tfl-uint8, guiding the user to run it before legalization. Based on the discussion here...
Also added the unit test to lock down the fix submitted in https://github.com/llvm/llvm-project/pull/168564 here since the change was simple. If I should create a separate PR, please let me know....
Result of `bazel test -c opt tensorflow/compiler/mlir/tosa/tests:all` ``` //tensorflow/compiler/mlir/tosa/tests:convert-tfl-uint8.mlir.test (cached) PASSED in 1.6s //tensorflow/compiler/mlir/tosa/tests:convert_metadata.mlir.test (cached) PASSED in 1.7s //tensorflow/compiler/mlir/tosa/tests:fuse-bias-tf.mlir.test (cached) PASSED in 2.1s //tensorflow/compiler/mlir/tosa/tests:lower-complex-types.mlir.test (cached) PASSED in 1.5s //tensorflow/compiler/mlir/tosa/tests:multi_add.mlir.test (cached)...
> I noticed that we currently run the `--tosa-convert-tfl-uint8` in the tfl-to-tosa pipeline to convert quantized uint8 to i8. I was wondering if you thought about doing something similar for...
Hi @vivekkhandelwal1, do you have any thoughts here?
Hi @vivekkhandelwal1, just verified that this is still an issue. Torch IR from the above example: ``` func.func @main(%arg0: !torch.vtensor) -> !torch.vtensor attributes {torch.assume_strict_symbolic_shapes} { %none = torch.constant.none %true =...
Another alternative to option 3 is to add a canonicalization pattern for the AvgPool op to extend the kernel size and other such attributes to 1x2 from 1x1 shape. This...