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 function

group_norm_gpu_multi_block_stats

def group_norm_gpu_multi_block_stats[StatsLayoutType: TensorLayout, stats_origin: MutOrigin, //, dtype: DType, simd_width: Int, InputFnType: def[width: Int](row: Int, col: Int) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable](stats: TileTensor[get_accum_type[dtype](), StatsLayoutType, stats_origin], num_splits: Int32, group_size: Int32, input_fn: InputFnType)

Multi-block stats kernel: computes partial Welford statistics per split.

Grid: num_rows * _num_splits blocks. Each block handles one split of one group and writes partial (mean, m2, count) to the stats buffer. Stats layout: stats[block_idx * 3 + {0,1,2}] = {mean, m2, count}.

input_fn is a trailing host-layout argument so enqueue does not DevicePassable-encode that capturing closure.