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

topk_wrapper_no_shmem

def topk_wrapper_no_shmem[input_type: DType, index_type: DType, *, is_top_p: Bool, block_size: Int, largest: Bool = True, _test_sort: Bool = False](K: Int, num_elements: Int, num_blocks_per_input: Int, in_buffer: UnsafePointer[Scalar[input_type], ImmutUntrackedOrigin], local_topk_vals: UnsafePointer[Scalar[input_type], MutUntrackedOrigin], local_topk_idxs: UnsafePointer[Scalar[index_type], MutUntrackedOrigin], p_threshold: UnsafePointer[Scalar[input_type], MutUntrackedOrigin], skip_sort: UnsafePointer[Scalar[DType.bool], MutUntrackedOrigin])

Shared-memory-free variant of topk_wrapper for Apple GPUs.

Uses warp-level reduction and register-based invalidation instead of shared memory. Only correct when block_size <= WARP_SIZE.

Was this page helpful?

Thank you! We'll create more content like this.

Thank you for helping us improve!

😔 What went wrong?

Some code doesn’t work

It includes inaccurate information

It's missing information I need

It was difficult to understand

Other

Submit