【QUESTION】How to train an inline model with our own code?

mtrofin commented 10 months ago

See the examples under docs, while they are specific to Fuchsia or chrome, they go through corpus extraction as a fairly early step.

High-level, assuming your project is cmake based:

you add to your cmake command line -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
build your project
run compiler_opt/tools/extract_ir.py (see this for example - you may want to drop the --cmd_filter)
the rest of the steps (i.e. Collect Trace and Generate Vocab) are only dependent on the corpus and not on the specifics of your project.

Colibrow commented 10 months ago

A long time ago, I trained a model for my projects, here is the gist when training gist Hope it helps.

18liumin commented 10 months ago

why ninja toolchain-distribution ninja: error: unknown target 'toolchain-distribution DESTDIR=${LLVM_INSTALLDIR} ninja install-toolchain-distribution-stripped

Colibrow commented 9 months ago

why ninja toolchain-distribution ninja: error: unknown target 'toolchain-distribution DESTDIR=${LLVM_INSTALLDIR} ninja install-toolchain-distribution-stripped

maybe some of these dependencies were missed, it's been a long time.

google / ml-compiler-opt

【QUESTION】How to train an inline model with our own code? #314