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 function

apple_gemv

def apple_gemv[*, b_packed: Bool, transpose_b: Bool = False, elementwise_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None] = None](c: TileTensor[Storage=c.Storage, linear_idx_type=c.linear_idx_type], a: TileTensor[Storage=a.Storage, linear_idx_type=a.linear_idx_type], b: TileTensor[Storage=b.Storage, linear_idx_type=b.linear_idx_type], ctx: Optional[DeviceContext] = None)

Performs a parallelized and vectorized GEMV for the M=1 case on Apple CPUs.

Used as an alternative to apple_matmul when M=1 and the Apple Accelerate cblas_sgemm path exhibits suboptimal performance for single-row inputs. Iterates over columns of B using SIMD FMA and optional parallelism.

Parameters:

b_packed (Bool): Whether B is already in packed (transposed) layout.
transpose_b (Bool): Whether to treat B as transposed.
elementwise_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None]): Optional epilogue applied element-wise to output.

Args:

c (TileTensor[Storage=c.Storage, linear_idx_type=c.linear_idx_type]): Output matrix tile (M=1).
a (TileTensor[Storage=a.Storage, linear_idx_type=a.linear_idx_type]): Input A matrix tile (M=1).
b (TileTensor[Storage=b.Storage, linear_idx_type=b.linear_idx_type]): Input B matrix tile.
ctx (Optional[DeviceContext]): Optional device context for parallelism control.