distributed_tensor_test#

Attributes#

test_data

Functions#

`_inductor_cpu_available`()	Probe whether torch.compile's Inductor CPU backend can actually build a kernel.
`make_test_data`()
`test_dstandard_tensor_scaler`()
`test_dminmax_tensor_scaler`()
`test_dquantile_tensor_scaler`()
`test_dquantile_tensor_vmap_edge_cases`()
`test_dquantile_tensor_refit_rebuilds_stacked_centroids`()
`test_tensor_scaler_with_attribute`()
`test_dquantile_tensor_fast_transform`()
`test_dquantile_tensor_compile_matches_eager`()

Module Contents#

distributed_tensor_test._inductor_cpu_available()#

Probe whether torch.compile’s Inductor CPU backend can actually build a kernel.

On some HPC module environments the g++ on PATH is a wrapper (e.g. NCAR’s ncarcompilers) that injects link flags, which breaks Inductor’s header-only precompile step and raises a CppCompileError. That is a toolchain issue, not a bridgescaler bug, so tests that require compilation skip rather than fail. Set CXX to a plain g++ to make this probe (and the compile path) succeed.

distributed_tensor_test.make_test_data()#

distributed_tensor_test.test_data#

distributed_tensor_test.test_dstandard_tensor_scaler()#

distributed_tensor_test.test_dminmax_tensor_scaler()#

distributed_tensor_test.test_dquantile_tensor_scaler()#

distributed_tensor_test.test_dquantile_tensor_vmap_edge_cases()#

distributed_tensor_test.test_dquantile_tensor_refit_rebuilds_stacked_centroids()#

distributed_tensor_test.test_tensor_scaler_with_attribute()#

distributed_tensor_test.test_dquantile_tensor_fast_transform()#

distributed_tensor_test.test_dquantile_tensor_compile_matches_eager()#