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

launch

def launch[Body: RowBody, //, axis: Int, simd_width: Int, target: StringSpan[ImmStaticOrigin], num_phases: Int, computationally_expensive: Bool = False, associative: Bool = False, dtype_size: Int = Int(0)](body: Body, shape: Coord, ctx: Optional[DeviceContext] = None)

Top-level scaffolder. num_phases picks reduce-shaped (num_phases == 1, output collapses the reduced axis, terminal emit) vs. normalize-shaped (num_phases > 1, output keeps the input shape, terminal elementwise) — see the module note above for the full picture, including the GPU tier-dispatch decision tree.

Parameters:

• ​axis (Int): Axis being reduced.
• ​simd_width (Int): SIMD width for tile dispatch. Bodies compute this via rowwise.pick_simd_width[...] and pass it directly.
• ​target (StringSpan[ImmStaticOrigin]): "cpu" or "gpu" (anything non-CPU routes through the GPU backend).
• ​num_phases (Int): Total phase count — 1 for reduce-shaped bodies (the reduce's own result is the final output); > 1 for normalize-shaped bodies (num_phases - 1 dependent row.reduce phases before a final per-element write). Asserted > 0.
• ​computationally_expensive (Bool): GPU author hint — flips the block tier's SIMD-full threshold for compute-heavy bodies. CPU ignores.
• ​associative (Bool): Reduce-shaped (num_phases == 1) only, CPU-only. Set True to widen the accumulator on long rows so the running total doesn't serialize into one add-per-element chain (recovers memory bandwidth). Only correct when reordering the accumulation doesn't change the result — additive reductions like reduce_sum / reduce_mean. Leave False for anything where element order matters (product's narrower-dtype rounding, tie-breaking in arg_max / arg_min, ...). Ignored on GPU and meaningless for normalize-shaped bodies.
• ​dtype_size (Int): Normalize-shaped (num_phases > 1) only, GPU only. Byte size of the body's primary dtype (size_of[dtype]()), used only by the phase-aware split-K byte gate; 0 (the default) leaves that tier's runtime condition permanently false, so the extra codegen it introduces is unreachable (softmax / log-softmax pass their dtype's size to actually engage it). Meaningless for reduce-shaped bodies.

Args:

• ​body (Body): The per-row computation. Receives a Context and the row's Coord; uses reduce / pjoin / once / simd to compose the algorithm.
• ​shape (Coord): Tensor shape.
• ​ctx (Optional[DeviceContext]): Optional DeviceContext; required for target="gpu", unused on CPU.

Raises:

On launch or worker failure.