Overview

Attributes are written above a kernel, one or more per line, each starting with @. They set the kernel’s launch shape, declare its tuning space, make promises about the data, and opt into hardware paths:

@autotune(TILE_M in [32, 256], TILE_N in [32, 256], TILE_K in [4, 32])
@launch(256)
@pipeline
@aligned(M = TILE_M, N = TILE_N, K = TILE_K)
kernel gemm(A: tensor<f32>[M, K], B: tensor<f32>[K, N], C: tensor<f32>[M, N], alpha: f32, beta: f32) {

An attribute’s arguments take three forms: a search dimension (TILE_M in [64, 128]), a keyword argument (M = TILE_M), or a positional value or flag (256, wmma). Which forms an attribute accepts is on its page.

The normative table is in the spec under Attributes.

The attributes

Attribute	Effect
@autotune	Declares tile constants and the values a tuner searches.
@launch	Sets the CTA size, and optionally an occupancy hint and a register cap.
@aligned	Promises a dimension is a whole number of tiles, which removes its masks.
@pipeline	Asserts that the kernel’s loops are double-buffered, and enables it for the GEMM.
@padstage	Pads staging tiles whose rows would collide on shared-memory banks.
@dynshared	Uses dynamic shared memory, sized at launch, instead of the static 48 KB.
@registers	Keeps named tiles in registers instead of shared memory.
@tensorcore	Runs dot and dot_t on tensor cores.
@persistent	Declares that the grid must be co-resident, as grid_barrier needs.
@cluster	Experimental. Super-tile dimensions for cluster tuning.

Unknown attributes are errors

An attribute the compiler does not read is a compile-time error, not a silent no-op. A misspelled @pipline would otherwise compile into a kernel that quietly skips the promise it was meant to check.

@unroll
@autotune(BLOCK in [1024])
kernel copy(x: tensor<f32>[N], y: tensor<f32>[N]) {
  let base = ctx.block.x * BLOCK
  y[base :+ BLOCK] = x[base :+ BLOCK]
}

The message names the kernel, the attribute it did not recognize, and every attribute it does read. Unrolling is not an attribute: a loop asks for it itself, as for<unroll n>.

Planned attributes

The spec lists three attributes as planned: @readonly, @fast_math and @assert_coalesced. None of them is implemented yet, and each is a compile-time error until it is, the same error as for @unroll above.