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 function

cp_async_bulk_tensor_shared_cluster_global_elect

def cp_async_bulk_tensor_shared_cluster_global_elect[dst_type: AnyType, mbr_type: AnyType, rank: Int, /, *, cta_group: Int = Int(1), eviction_policy: CacheEviction = CacheEviction.EVICT_NORMAL](dst_mem: Pointer[dst_type, address_space=AddressSpace.SHARED], tma_descriptor: Pointer[NoneType], mem_bar: Pointer[mbr_type, address_space=AddressSpace.SHARED], coords: IndexList[rank], elect: Int32)

Elect-predicated variant of cp_async_bulk_tensor_shared_cluster_global.

Behaves exactly like cp_async_bulk_tensor_shared_cluster_global except that the TMA instruction is guarded by a PTX predicate derived from elect: the instruction is issued only when elect != 0. All lanes follow the same PTX control flow, so there is no Mojo-level branch and no warp-divergent if elect != 0: wrapper at the call site.

Parameters:

• ​dst_type (AnyType): The data type of the destination memory.
• ​mbr_type (AnyType): The data type of the memory barrier.
• ​rank (Int): The dimensionality of the tensor (1, 2, 3, 4, or 5).
• ​cta_group (Int): The CTA group to use for the copy operation. Must be 1 or 2.
• ​eviction_policy (CacheEviction): Optional cache eviction policy. Defaults to EVICT_NORMAL.

Args:

• ​dst_mem (Pointer[dst_type, address_space=AddressSpace.SHARED]): Pointer to the destination in shared memory.
• ​tma_descriptor (Pointer[NoneType]): Pointer to the TMA descriptor.
• ​mem_bar (Pointer[mbr_type, address_space=AddressSpace.SHARED]): Pointer to the cluster-shared memory barrier.
• ​coords (IndexList[rank]): Coordinates specifying which tile to copy.
• ​elect (Int32): 0 on non-elected lanes (skip the TMA), non-zero on the single elected lane (issue the TMA). Typically the Int32 returned by elect() from nn.attention.gpu.nvidia.sm100.attention_utils.