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 trait
CommTuningConfig
Tuning-table entry for a multi-GPU communication collective.
Extends TuningConfig with the four dimensions that drive kernel-launch
selection: SM version, GPU count, total data size, and thread-block count.
Implement this trait to supply custom tuning tables to
dispatch_select_comm_config.
Implemented traits
AnyType,
Copyable,
Deinitable,
ImplicitlyCopyable,
Movable,
RegisterPassable,
TrivialRegisterPassable,
TuningConfig,
Writable
Required methods
get_num_blocks
def get_num_blocks(self) -> Int
Returns the number of thread blocks to launch for this configuration.
Returns:
Int: The thread-block count, which must not exceed 512
(MAX_NUM_BLOCKS_UPPER_BOUND).
get_num_bytes
def get_num_bytes(self) -> Int
Returns the maximum input size in bytes covered by this entry.
dispatch_select_comm_config selects the first entry whose
get_num_bytes() is at least the actual transfer size.
Returns:
Int: The upper-bound byte count for this tuning entry, or -1 for the
default (catch-all) entry.
get_sm_version
def get_sm_version(self) -> StaticString
Returns the SM version string this entry targets.
Returns:
StaticString: A string such as "sm_90a" or "sm_100a", or "default" for
the architecture-agnostic fallback entry.
get_ngpus
def get_ngpus(self) -> Int
Returns the GPU count this entry targets.
Returns:
Int: The number of participating GPUs, or -1 for the default
(catch-all) entry.