Fix default-build for SM120/121 block-scaled GEMM: keep largest tile first

[krauqen]

Problem

A default CUTLASS library / profiler build for SM120-class GPUs — including SM121 (DGX Spark), using the flag set suggested in the Quick Start Guide —

cmake .. -DCUTLASS_NVCC_ARCHS=121a -DCUTLASS_ENABLE_TESTS=OFF -DCUTLASS_UNITY_BUILD_ENABLED=ON
make cutlass_profiler -j

fails with static asserts in every SM120 block-scaled (bstensorop) GEMM shard whose data types touch FP4/FP6 or block-scale-factor output. Representative failures (CUDA 13.0, sm_121a):

include/cutlass/gemm/collective/builders/sm90_common.inl(321): error: static assertion failed with
  "BLK_K0 must be a multiple of size<1>(GMMA::Layout_K_INTER_Atom<ElementType>{})"
  [cutlass3x_sm120_bstensorop_gemm_..._void_e5m2_128x8x128_..._cooperative_q]

include/cute/atom/copy_atom.hpp(206): error: static assertion failed with
  "TiledCopy uses too few vals for selected CopyAtom"
  [cutlass3x_sm120_bstensorop_gemm_ue8m0xe2m1_ue8m0xe2m3_f32_void_f32_128x8x128_..._cooperative_q]

include/cutlass/epilogue/fusion/sm120_visitor_store_tma_warpspecialized.hpp(72): error: static assertion failed with
  "EpilogueTileN should be divisible by SFVecSize"
  [cutlass3x_sm120_bstensorop_gemm_ue4m3xe2m1_ue4m3xe2m1_f32_f16_ue8m0xe2m1_128x8x128_..._epiVs16t]

include/cutlass/epilogue/collective/builders/sm120_builder.inl(158): error: static assertion failed with
  "CTA tile for FP6 ElementD must have a contiguous extent that is a multiple of 128."

The consequence is that no NVFP4/MXFP4/FP6 block-scaled GEMM is buildable through the library on SM120/SM121 without hand-maintained CUTLASS_LIBRARY_EXCLUDE_KERNELS patterns, even though the underlying kernels are fully functional at larger tiles.

Root cause

With an empty CUTLASS_LIBRARY_KERNELS, CreateGemmUniversal3xOperator prunes each family's tile list to tile_descriptions[0]:

# by default, only generate the largest tile and largest alignment
if manifest.kernel_filter == '':
    tile_descriptions = [tile_descriptions[0]]

This relies on tile lists being ordered largest-first. The SM120 block-scaled generators (GenerateSM120_TensorOp_mixed_8bits_UMMA_gemm_with_block_scaled, GenerateSM120_TensorOp_fp4_UMMA_gemm_with_block_scaled) list their tiles smallest-first, so the default set instantiates the family exclusively at 128x8x128 (cooperative) / 128x16x128 (pingpong), crossed with every output variant the generators emit. Several of those combinations cannot compile at small tile N:

• narrow-byte ElementD (e5m2, e2m1) — no valid epilogue smem atom at N=8 (and N=16 pingpong for e2m1+SFD)
• FP6 ElementD — requires a contiguous CTA extent that is a multiple of 128
• block-scale-factor output — epilogue tile N (= min(CTA_N, 32)) must be divisible by SFVecSize
• FP6 B operand — SM100_SU6_DU8x16_x4_LDSM_N needs per-MMA-atom N ≥ 16 (fails at N=8 cooperative and N=16 pingpong)

The same combinations compile at 128x128x128, which was tile_descriptions[0] before the small-N tiles were added.

Fix

Reorder tile_sizes_cooperative / tile_sizes_pingpong in the two SM120 block-scaled generator functions to largest-first, restoring the tile_descriptions[0] convention. The default set then selects 128x128x128, which compiles for every generated dtype/output combination. No tiles are removed: builds with a non-empty CUTLASS_LIBRARY_KERNELS still instantiate the full lists, including the small-N tiles for skinny-N problems.

Add a regression test, test/python/cutlass_library/test_sm120_blockscaled_default_tiles.py (pure Python, no GPU required):

• default-set path (empty kernel filter) must select TileN ≥ 128 for both generators and both dense/grouped kinds — fails on current main (TileN=8 selected), passes with this change;
• non-empty-filter path must still contain the N=8/16 tiles, guarding against fixing this by deletion.

PYTHONPATH=python python3 test/python/cutlass_library/test_sm120_blockscaled_default_tiles.py

Validation

Tested on NVIDIA DGX Spark (GB10, SM121) with CUDA 13.0.88.

• Generator run for 121a, empty kernel filter, no excludes: 500 bstensorop kernels emitted, all 128x128x128 (previously: all 128x8x128/128x16x128, of which large subsets fail to compile).
• Full default cutlass_profiler build for 121a with no CUTLASS_LIBRARY_EXCLUDE_KERNELS: completes with 0 errors (previously fails in every FP4/FP6-touching shard).
• Runtime sweep: all 484 block-scaled kernels in the resulting profiler binary executed at m=n=k=1024, one process per kernel: 484/484 passed, including NVFP4×NVFP4 with FP4-requantization output (LinCombBlockScaleFactor) and grouped variants.
• New regression test: passes with this change; fails on current main (4 failures, default set selects TileN=8).
• git diff --check passes.