← Back to events
ActiveAIRelease

viable/strict/1788218631: [MPS] Fix conv backward to preserve input memory format (#174241)

Photo: PyTorch Releases

What happened

Summary Fixes convolution_backward on MPS to preserve the original input tensor's memory format, rather than deriving it from grad_output . When convolution_backward is called with a channels_last input but contiguous grad_output , the grad_input should preserve the input's memory format ( channels_last ), not adopt grad_output 's format (contiguous). Problem Previously, mps_convolution_backward_input determined the…

Summary assembled by rule from the sources below

Why it's spreading

Sources