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

short_conv_ring_fwd

def short_conv_ring_fwd[dtype: DType, ring_dtype: DType, //, target: StringSpan[ImmStaticOrigin]](x: TileTensor[dtype, Engine=x.Engine, address_space=x.address_space, linear_idx_type=x.linear_idx_type], weight: TileTensor[dtype, Engine=weight.Engine, address_space=weight.address_space, linear_idx_type=weight.linear_idx_type], ring: TileTensor[ring_dtype, Engine=ring.Engine, address_space=ring.address_space, linear_idx_type=ring.linear_idx_type], input_row_offsets: TileTensor[.uint32, Engine=input_row_offsets.Engine, address_space=input_row_offsets.address_space, linear_idx_type=input_row_offsets.linear_idx_type], positions: TileTensor[.uint32, Engine=positions.Engine, address_space=positions.address_space, linear_idx_type=positions.linear_idx_type], conv_row: TileTensor[.uint32, Engine=conv_row.Engine, address_space=conv_row.address_space, linear_idx_type=conv_row.linear_idx_type], output: TileTensor[dtype, Engine=output.Engine, address_space=output.address_space, linear_idx_type=output.linear_idx_type], context: DeviceContext)

x + conv(x) over a ragged batch; reads the ring, writes nothing.

x, output: [total_seq_len, channels]. weight: [channels, width]. ring: [slots, R, channels]. positions is per token, conv_row per sequence.