trunk/62146ea0c777ac1a6fcc637a647335b06fba3019: [xpu][fix]Relax fp16 grad tolerance for combinations on XPU (#195403)
- PyTorch: 350 events in the last 90 days
- PyTorch: 347th Release in the last 90 days
- Previous: earlier the same day · trunk/3c512baeddc8fdc688d905a33baf89d97ebb050e: Dispatch `index_put_` to `masked_fill_.Scalar` when possible (#188330)
What happened
Motivation #189102 bumped the gradient tolerances for test_comprehensive_combinations_cuda_float16 after #186595 / #189305 reimplemented torch.combinations on top of index_put , whose fp16 backward Inductor accumulates in fp16. That PR only added the override to inductor_override_kwargs["cuda"] ; inductor_override_kwargs["xpu"] is a separate dict and never got the equivalent entry, leaving XPU on the default fp16 gr…
Summary assembled by rule from the sources below