For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo module

output_writer_trait

Output-writer trait surface for the SM100 structured matmul epilogue.

This module defines the pluggable boundary that lets BlackwellMatmulSM100Kernel emit its accumulator tiles through an injected writer policy, instead of hard-referencing a concrete writer struct. This breaks the source dependency from matmul_kernels.mojo onto rs_output_writer.mojo, so the reduce-scatter writer can later move to a closed-source tree.

OutputWriter is the policy the kernel takes as a comptime parameter (output_writer_type). It carries:

needs_sync: Bool: whether the epilogue must wrap the write in a cross-GPU barrier (reduce-scatter) or not (local store).
num_peers: Int: number of C TMA descriptors the kernel must supply in c_tma_ops (1 for the local store, one per peer for reduce-scatter).
write_batched[...]( c_tma_ops, c_tiles, stage, tile_coord, shape, alpha ): a static method that constructs the concrete writer and writes one batched output tile. Construction lives inside the policy, where the concrete writer type is named and non-erased; the kernel only passes the c_tma_ops array pointer plus the shared-typed SMEM tiles / stage.

Why a static-method policy rather than injecting the concrete writer type or a trait-typed writer value:

A Mojo struct parameter default cannot reference the kernel's Self.-derived aliases (opc, SmemType.Output*, ...), and the kernel's callers cannot reconstruct that config-derivation to spell the full writer type at the injection site. So the concrete writer must be resolved inside the kernel.
A trait associated type / value erases to its declared bound, so a writer obtained as a trait-typed value cannot have write_batched called on it (its associated CTileArray / Stage arg types are opaque, and matching a by-value-Tuple requirement against an @always_inline impl trips a parameter-origin mismatch). Keeping the construct-and-write step inside the policy static method dodges erasure entirely: there the writer is concrete, and c_tiles / stage are the shared SMemTileArray2DRowMajor[...] / OutputStage[opc] types both writers already use.

Note (interim design): folding construction + write into one policy static method is a temporary approach. It is expected to fold into the upstream TileConsumer/TileOperation traits once those land a non-erasing pluggable-writer shape.

Target hardware: SM100 (B200).

Traits

OutputWriter: Injected output-writer policy for BlackwellMatmulSM100Kernel.

Traits​

Traits