Closed NullSenseStudio closed 2 months ago
Hi @NullSenseStudio , thanks for your feedback. We'll be including a fix to address non-contiguous inputs in layer_norm in our upcoming torch-directml build releasing soon.
@NullSenseStudio We just released our new build that addresses the layer_norm
issue. Please pip install torch-directml --upgrade
to update to torch-directml 0.2.2.dev240614
torch-directml: 0.2.1.dev240521 python: 3.11.7
0.6974620223045349 3.6198389530181885
The result isn't close to what it is on CPU or other devices.7.381072464340832e-08 2.384185791015625e-07
But it will work as expected as long as it is made contiguous.This non-contiguous input will cause an error instead.