piano_123

[FEA] LinearCombinationSilu epilogue

14

I modified the epilogue function in Example 17 from LinearCombinationRelu to LinearCombinationSilu, like this: using EpilogueOp = cutlass::epilogue::thread::LinearCombinationSilu< ElementOutput, // Data type of output matrix. 128 / cutlass::sizeof_bits::value, // The...

INT8 accuray issue for Multiscale deformable attention

3

hello, I'm curious about the accuray of FUNCTION:ms_deformable_im2col_cuda_int8() channels /= 4; const int value_step = num_heads * spatial_size * channels; const int output_step = num_heads * num_query * channels; const...