← Back to events
ActiveAIRelease

viable/strict/1790773504: Extend `native_layer_norm` param dtype check from CUDA to XPU (#198647)

What happened

_check_native_layer_norm_cuda_param_dtype early-returned on non-CUDA devices, so the decomposition skipped the weight/bias dtype validation on XPU and diverged from eager under aot_eager_decomp_partition . XPU has identical kernel semantics (including the num_rows == 0 carve-out), so gate on both devices and rename the helper accordingly. Unskips test_layer_norm_mixed_dtype_aot_eager_decomp_partition_errors for XPU.…

Summary assembled by rule from the sources below

Why it's spreading

Sources