[BUG] EpilogueTileAuto doesn't work when tile shape is (128, 112, 64)

[henrylhtsang]

Describe the bug
I am using gemm configs from generator.py. op.configuration name is cutlass3x_sm90_tensorop_s64x56x16gemm_f16_f16_f32_void_f16_128x112x64_4x1x1_0_ttn_align8_warpspecialized_pingpong_epi_tma

This is the code https://gist.github.com/henrylhtsang/aa797c7d280b792eef574317111ac215

Question:

1. Should I consider not use EpilogueTileAuto? Or is this fixable from this logic?

   cutlass/include/cutlass/epilogue/collective/builders/sm90_builder.inl

   Lines 115 to 121 in eefa171

   	if constexpr (cute::is_same_v<EpilogueTileType, EpilogueTileAuto>) {
   	auto epi_tile = [&] () {
   	if constexpr (detail::sm90_is_cooperative_v<Schedule>) {
   	auto tile_m = cute::min(_128{}, size<0>(TileShape_MNK{}));
   	auto tile_n = cute::min(_32{}, size<1>(TileShape_MNK{}));
   	return make_shape(tile_m, tile_n);
   	}

2. What epilogue tile shape should I use?

Steps/Code to reproduce bug
compile https://gist.github.com/henrylhtsang/aa797c7d280b792eef574317111ac215

     static_assert(size<1>(CtaTileMNK{}) % size<1>(shape(EpilogueTile{})) == 0, "EPI_TILE_N must divide CTA_N");

Expected behavior
compilation should be successful

Environment details (please complete the following information):

• Environment location: [Bare-metal, Docker, Cloud(specify cloud provider)]

Additional context
Add any other context about the problem here.

[tridao]

You'd want

auto tile_n = cute::gcd(cute::min(_32{}, size<1>(TileShape_MNK{})), size<1>(TileShape_MNK{}));