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_blockwise_fp8
Provides blockwise-scaled FP8 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[accum_type: DType] = TMemStorage[accum_type, MutAnyOrigin]
Parameters
- accum_type (
DType):
AccumTile
comptime AccumTile[accum_type: DType] = TileTensor[accum_type, Layout[TypeList[ComptimeInt[Int(128)], ComptimeInt[Int(512)]](), TypeList[ComptimeInt[Int(1)], ComptimeInt[Int(128)]]()], MutAnyOrigin, Engine=TMemEngine, linear_idx_type=.uint32]
Parameters
- accum_type (
DType):
FRAGMENT_LANES
comptime FRAGMENT_LANES = 16
Lanes of the block a tmem_16x256b view covers.
logger
comptime logger = Logger(stdout, prefix=String(""), source_location=False)
Functions
-
blackwell_gmm_tma_umma_warp_specialized_blockwise_fp8_kernel: -
grouped_matmul_dynamic_scaled_fp8: TileTensor primary implementation ofgrouped_matmul_dynamic_scaled_fp8. -
grouped_matmul_sm100_blockwise_scaled_fp8: Launches the basic (non-persistent) SM100 blockwise-scaled FP8 grouped GEMM kernel. -
grouped_matmul_sm100_blockwise_scaled_fp8_persistent: Launches the persistent warp-specialized SM100 blockwise-scaled FP8 grouped GEMM kernel. -
load_AB: Issues multicast TMA loads for A, B, and A scales into the producer stage of the load-MMA pipeline. -
load_AB_partial: Sibling toload_ABfor tiles the full-TMA path can't handle: fills A anda_scalesSMEM via a cooperative warp copy from gmem and issues TMA only for B. -
matmul_sm100_grouped_blockwise_scaled_fp8_1d2d_kernel: -
multi_stage_reg_epilogue: Writes the accumulated C fragments from registers to global memory through a staged shared-memory buffer. -
promote_accumulators: Loads MMA outputs from TMEM and accumulates them into register C fragments with blockwise FP8 scaling.