building from source fails due to mlir BUILD referring to MlirHloModule.cc instead of MlirHloModule.cpp
randyau opened this issue · 3 comments
building jax from source with --enable_cuda flag,
freshly cloned from the repo
on an WSL2 ubuntu install
gets this error in building:
ERROR: /mnt/e/linux/ubuntu/jax/jaxlib/mlir/_mlir_libs/BUILD.bazel:90:13: no such target '@org_tensorflow//tensorflow/compiler/mlir/hlo:python/MlirHloModule.cc': target 'python/MlirHloModule.cc' not declared in package 'tensorflow/compiler/mlir/hlo' (did you mean 'python/MlirHloModule.cpp'?) defined by /home/agi/.cache/bazel/_bazel_agi/e71f5b8290a4fc88939262d32d7b9edb/external/org_tensorflow/tensorflow/compiler/mlir/hlo/BUILD and referenced by '//jaxlib/mlir/_mlir_libs:_mlirHlo.so'
Issue can be fixed by going to
/mnt/e/linux/ubuntu/jax/jaxlib/mlir/_mlir_libs/BUILD.bazel:90
and changing the file extension from .cc to .cpp
Looks like the line was recently patched from .cpp to .cc to "Fix broken builds"
not sure why there's conflicting build experiences
See the linked pull-request I just opened - the real change happened inside of the Tensorflow repository, where the MlirHLO compiler lies, and where cpp files were renamed to cc files. When bumping the version of the Tensorflow repo in the WORKSPACE file to a more current one, this issue is fixed.
Try again now from main - it should work now :)
This is now fixed at head due to #11073 (Thanks for the PR @ludgerpaehler !)