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 module

concat

Implements tensor concatenation along a specified axis for CPU and GPU targets.

comptime values

elementwise_epilogue_type

comptime elementwise_epilogue_type = def[c_type: DType, rank: Int, width: SIMDLength = 1, *, alignment: Int = Int(1)](IndexList[rank], SIMD[c_type, width]) capturing thin -> None

ElementwiseEpilogueFn

comptime ElementwiseEpilogueFn = def[c_type: DType, rank: Int, width: SIMDLength, *, alignment: Int](IndexList[rank], SIMD[c_type, width]) -> None & RegisterPassable & ImplicitlyCopyable

Value-taking counterpart of elementwise_epilogue_type.

A unified closure passed as a runtime value carries the origins of its captures, so the output it writes stays alive until the kernel that calls it has launched. width and alignment have no defaults; pass both where the legacy type relied on its defaults.

Functions

• ​concat: Concatenates inputs along axis into output for the given target.
• ​concat_shape: Compute the output shape of a pad operation, and assert the inputs are compatible.
• ​fused_concat: Concatenates inputs produced by input_fn along axis into output, applying output_0_fn to each element.
• ​memcpy_or_fuse: Copies n bytes from src_data into dest_data at out_byte_offset, applying epilogue_fn elementwise when has_epilogue is set.
• ​preferred_simd_width: SIMD scalar count for fused GPU concat vectorization.