"...composable_kernel_onnxruntime.git" did not exist on "e823d518cb46ad61ddb3c70eac8529e0a58af1f8"
Limit topK
This changes output/runtime of topk to run significantly faster since we're not not performing a sort and using a large heap to do said sorting. Originally in resnext we were performing a sort on the entire input vector instead of a subset.
Showing
Please register or sign in to comment