This is an existing issue with the compilation cache tests.
The refactoring of the cache key generation part into a separate
file requires tagging the refactored tests also.
PiperOrigin-RevId: 551972670
This is in preparation for introducing a more robust key-generation
algorithm.
This refactoring does not introduce any change in behavior.
Testing: refactored unit tests and test workload.
PiperOrigin-RevId: 551744892
Test cases are still frequently skipped due to lack of CompileOptions
support, but the skip/run behavior does not seem to meaningfully change
compared to a clean checkout. This was verified by inserting an exception
in place of unittest.SkipTest.
PiperOrigin-RevId: 529437419
Metadata, in particular code location information is present in the HLO generated by JAX. The compilation cache uses the serialized HLO as a cache key, which begs the question: should code location information be part of that key? Simply changing the line number on which a function appears shouldn't necessarily cause a cache miss.
There are pros and cons: the main advantage of excluding metadata is that we will get more cache hits, and the main disadvantage is that debug information and profiling data in the HLO might become confusing, since it may refer to a different program entirely, or to a version of a program that does not correspond to the current state of the source tree. We argue that saving compilation time is the more important concern.
This change adds a tiny MLIR pass that strips Locations from a StableHLO module, and applies it in the compilation cache if metadata stripping is enabled.
PiperOrigin-RevId: 525534901
Use a Protocol instead of an abstract base class for the CacheInterface since it allows us to use one fewer file.
No functional change intended.
PiperOrigin-RevId: 524855263
Now that all functionality needed by frameworks is implemented, let's
remove the possibility of not noticing missing functionality due to
the bypass.
PiperOrigin-RevId: 519018438
Unfortunately we can't conditionally change the timeout, as size and timeout
are both non-configurable even if jax_test supported setting the size.
PiperOrigin-RevId: 514745247
The goal is to ensure that all shards fit into a medium timeout in sanitizer
configurations.
Running 256 entry vectors in spectral_dac is too slow, so let's replace that
with a smaller vector that isn't a power of 2. Avoiding a power of 2 requires
us to widen the tolerance a bit due to vectorization changes.
While here, specify deps a little more precisely as well.
PiperOrigin-RevId: 514440062
The spectral_dac tests are also shrunk because running the full suite on 256-entry vectors is too slow.
This allows them to run in ASAN in more situations.
While here, specify deps a little more precisely as well.
PiperOrigin-RevId: 511829646
These utils are currently shared with lax_vmap_test by importing lax_test as a
library, which is an odd thing to do.
The new package and the module within it are not built into the wheel, as these
are internal utilities for JAX's tests, not utilities for JAX users writing
their own tests.
Followup changes will add additional existing internal test utilities to this
package. This will allow removing sys.path manipulation from
deprecation_module_test and hopefully lazy_loader_test, as well as removing
the non-public test_util.py from _src to make it clearer that it should not be
used from outside JAX.
PiperOrigin-RevId: 510260230
