For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo module
grouped_matmul_sm100
Provides warp-specialized persistent grouped GEMM kernels for SM100 (B200) GPUs.
comptime values
ACCUM_LAYOUT
comptime ACCUM_LAYOUT = Layout(Coord(), Coord())
The whole TMEM allocation as a lane-by-column grid.
AccumStorage
comptime AccumStorage = TMemStorage[dtype, MutAnyOrigin]
AccumTile
comptime AccumTile = TileTensor[dtype, Layout[TypeList[ComptimeInt[Int(128)], ComptimeInt[Int(512)]](), TypeList[ComptimeInt[Int(1)], ComptimeInt[Int(128)]]()], MutAnyOrigin, Engine=TMemEngine, linear_idx_type=.uint32]
FRAGMENT_LANES
comptime FRAGMENT_LANES = 16
Lanes of the block a tmem_16x256b view covers.
Structs
-
WarpRole: Enumerates the role each warp plays in the warp-specialized kernel.
Functions
-
blackwell_tma_umma_warp_specialized_kernel: Implements the warp-specialized persistent grouped GEMM kernel for SM100. -
consumer_main_loop: Performs the MMA consumer step for one pipeline stage. -
grouped_matmul_sm100_persistent: Launches the persistent grouped GEMM kernel for SM100 from host tensors. -
load_AB: Loads A and B tiles from global memory into shared memory via TMA multicast. -
load_AB_cuda_core: CUDA core fallback for load_AB when K*sizeof < 16 bytes. -
multi_stage_store_C: Drains accumulated results from tensor memory and stores them to global memory. -
stsm_helper: Stores a register fragment to shared memory using the stmatrix instruction. -
zero_output: Zero-fills an output tile for skipped or invalid expert assignments. -
zero_output_epilogue: Zero an inactive group's output THROUGH the elementwise epilogue.