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

rope_q_proj

def rope_q_proj[dtype: DType, freq_dtype: DType, rank: Int, width: SIMDLength, output_dtype: DType, //, *, interleaved: Bool, has_nope_prefix: Bool = False, rope_dim: Int = Int(0), alignment: Int = Int((get_alignof SIMD[dtype, width], _current_target()))](q_proj: TileTensor[dtype, Storage=q_proj.Storage, address_space=q_proj.address_space, linear_idx_type=q_proj.linear_idx_type], output: TileTensor[output_dtype, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type], idx: IndexList[rank], freq_val: SIMD[freq_dtype, width], head_size: Int)

Applies RoPE to a query projection tile and stores the result.

Loads the query projection values at the given coordinate, applies the rotary position embedding via rope_value, and writes the transformed result to the output tile. Supports both interleaved and split (real and imaginary stored separately) layouts, and optionally leaves a nope prefix region unrotated.

Parameters:

dtype (DType): Element type of the q_proj tile tensor (inferred).
freq_dtype (DType): Element type of the freq_val frequency coefficients (inferred).
rank (Int): Rank of the index list and tile tensors (inferred).
width (SIMDLength): Number of elements per SIMD vector (inferred).
output_dtype (DType): Element type of the output tile tensor (inferred).
interleaved (Bool): Whether the RoPE weights use interleaved real and imaginary layout.
has_nope_prefix (Bool): Whether a leading prefix of the head dimension is left unrotated (defaults to False).
rope_dim (Int): Number of trailing head dimensions that undergo RoPE when has_nope_prefix is set (defaults to 0).
alignment (Int): Memory alignment in bytes for tile loads and stores (defaults to the natural alignment of SIMD[dtype, width]).

Args:

q_proj (TileTensor[dtype, Storage=q_proj.Storage, address_space=q_proj.address_space, linear_idx_type=q_proj.linear_idx_type]): The query projection tile tensor to read from.
output (TileTensor[output_dtype, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type]): The mutable output tile tensor to write the rotated result.
idx (IndexList[rank]): The index list identifying the load and store coordinate.
freq_val (SIMD[freq_dtype, width]): The RoPE frequency coefficients for this position.
head_size (Int): The size of each attention head dimension.