[Stacked][Feature] Support NVFP4 Gemm on Blackwell arch (SM100,110,120) #2324

Open

Hale423 wants to merge 26 commits into

tile-ai:mainfrom

Hale423:feat/gemm-nvfp4

Commits on Mar 12, 2026

feat: add float4_e2m1fn (NV FP4) support for SM100 TCGEN5 MMA
Hale423
committed
fix: python pipeline (LayoutInference + LowerTileOp) passes
Hale423
committed
fix: SM120 FP4 CUDA compilation – type bridge, codegen naming, MMA dispatch
Hale423
committed

Commits on Mar 13, 2026

fix(buffer overflow, fp4 address misalignment): ldmatrix writes in by raw bytes, get_ldmatrix_offset add fp4 special layout, add SM120_FP4_FP4_F32_TN MmaDispatcher specialization + fp4 << 2 bit-sh…
Hale423
committed

Commits on Mar 14, 2026

Commits on May 8, 2026

Merge origin/main into feat/gemm-fp4

Hale423
and
cursoragent
committed

Commits on May 11, 2026

Commits on May 15, 2026

Commits on Jun 3, 2026

feat(sm120): add T.nvfp4_gemm as frontend, lower to mma_blockscaled, fix ptx_mma_blockscaled meta param must be of StringImm/IntImm, add and verified gemm_nvfp4 & fused moe examples
Hale423
committed

Commits on Jun 5, 2026

Commits on Jun 15, 2026

Commits on Jun 22, 2026