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 struct
TileScheduler
struct TileScheduler[tile_shape: UInt32, num_heads: UInt32, /, num_ctas: UInt32 = SIMD(GPUInfo.from_family(AcceleratorArchitectureFamily(Int(32), Int(2048), Int(233472), Int(65536), Int(1024)), StringSlice("H100"), Vendor(Int8(2)), StringSlice("cuda"), StringSlice("hopper"), SIMD(9), StringSlice("sm_90a"), Int(132)).sm_count), schedule: MHASchedule = MHASchedule.DEFAULT]
Implements a persistent tile scheduler that cycles CTAs through work tiles.
Each CTA begins at its block_idx and strides by num_ctas on every
advance, reusing the same SMs across multiple work tiles to amortize launch
overhead and improve cache locality.
Parameters
- tile_shape (
UInt32): Size of each query tile along the sequence dimension, in tokens. - num_heads (
UInt32): Number of attention heads mapped across the grid'sydimension. - num_ctas (
UInt32): Number of CTAs to launch for the persistent kernel, where each CTA strides by this count between tiles (defaults to the H100 SM count). - schedule (
MHASchedule): Strategy for mapping work tiles to thread blocks (defaults toMHASchedule.DEFAULT).
Implemented traits
AnyType,
Copyable,
Defaultable,
DevicePassable,
ImplicitlyCopyable,
ImplicitlyDeletable,
MHATileScheduler,
Movable,
RegisterPassable,
TrivialRegisterPassable
comptime members
device_type
comptime device_type = TileScheduler[tile_shape, num_heads, num_ctas, schedule]
may_advance
comptime may_advance = True
mha_schedule
comptime mha_schedule = schedule
Methods
__init__
def __init__() -> Self
get_type_name
get_current_work_info
def get_current_work_info[ValidLengthType: OptionalPointer, //](self, ts: MHATileSummary[ValidLengthType], state: MHATileState) -> WorkInfo
Returns:
fetch_next_work
def fetch_next_work(self, ts: MHATileSummary, mut state: MHATileState) -> WorkInfo
Returns:
advance
def advance[ValidLengthType: OptionalPointer, //, producer: Bool, sync: MHASchedulerSynchronization = MHASchedulerSynchronization.DEFAULT](self, ts: MHATileSummary[ValidLengthType], mut state: MHATileState, pipeline_idx: UInt32) -> OptionalReg[SeqInfo]
Returns:
grid_dim
static def grid_dim(batch_size: UInt32, max_num_prompt_tiles: UInt32) -> Tuple[Int, Int, Int]
Returns:
initial_state
def initial_state[ValidLengthType: OptionalPointer, //](self, ptr: Pointer[UInt32, MutAnyOrigin, address_space=AddressSpace.SHARED, _safe=False], tile_summary: MHATileSummary[ValidLengthType]) -> MHATileState
Returns:
unsafe_seq_info
def unsafe_seq_info[ValidLengthType: OptionalPointer, //](self, ts: MHATileSummary[ValidLengthType], state: MHATileState) -> SeqInfo
Returns:
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!