Expose torch_xla.runtime.clear_computation_cache binding to clear computation cache

[jameszianxuTT]

Ticket: #9712

There is currently no way to clear the XLAGraphExecutor's computation cache.

This change exposes a cache clearing function torch_xla.runtime.clear_computation_cache to clear the computation cache. It also adds some simple tests to check that the in-memory and persistent disk cache are correctly cleared by this function.

[jameszianxuTT]

Hey @ysiraichi, apologies for the ping but would it be possible to either approve this PR to run CI, or otherwise comment on it?

We've been using it in our fork without issue so far for several weeks, and it's an opt-in feature.

[ysiraichi]

Could you add tests for memory and persistent cache?

[jameszianxuTT]

Added some simple tests. Some small instrumentation shows these results, as expected.

 python test/test_compilation_cache_utils.py TestClearCompilationCache.test_clear_persistent_cache
[...]
[ RUN      ] TestClearCompilationCache.test_clear_persistent_cache
2026-01-19 20:51:51.124016: W torch_xla/csrc/runtime/profiler.cpp:88] Profiler API not found for PJRT plugin
graph_cnt: 1
cache_files_before: ['339896306817392351095887498798814495545']

[run xr.clear_computation_cache()]

new_graph_cnt: 0
cache_files_after: []

python test/test_compilation_cache_utils.py TestClearCompilationCache.test_clear_memory_cache
[...]
[ RUN      ] TestClearCompilationCache.test_clear_memory_cache
2026-01-19 20:48:47.183612: W torch_xla/csrc/runtime/profiler.cpp:88] Profiler API not found for PJRT plugin
graph_cnt: 1

[run xr.clear_computation_cache()]

new_graph_cnt: 0
[       OK ] TestClearCompilationCache.test_clear_memory_cache