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 trait
TokenFormat
Specifies the wire format for a single MoE token in EP dispatch/combine.
Implementors encode how a token's hidden-state vector is packed for cross-GPU transfer (quantization, scale placement, alignment) and how the received bytes are unpacked into the output tensor. The graph compiler selects a concrete implementation based on the model's quantization config.
All size and alignment properties are compile-time constants so the dispatch kernel can allocate receive buffers and issue vectorized copies without runtime branching.
Implemented traits
AnyType,
DevicePassable,
ImplicitlyDeletable
comptime members
alignment
comptime alignment
device_type
comptime device_type
Indicate the type being used on accelerator devices.
dispatch_smem_size
comptime dispatch_smem_size
dispatch_wait_tile_shape
comptime dispatch_wait_tile_shape
hid_dim
comptime hid_dim
top_k
comptime top_k
Required methods
token_size
copy_token_to_send_buf
static def copy_token_to_send_buf[src_type: DType, block_size: Int, buf_addr_space: AddressSpace = AddressSpace.GENERIC](buf_p: Pointer[UInt8, address_space=buf_addr_space, _safe=False], src_p: Pointer[Scalar[src_type], address_space=src_p.address_space, _safe=False], input_scale: Float32)
Copy the token to the send buffer. This function needs to be called by all threads in the block.
copy_msg_to_output_tensor
def copy_msg_to_output_tensor[buf_addr_space: AddressSpace = AddressSpace.GENERIC](self, buf_p: Pointer[UInt8, address_space=buf_addr_space, _safe=False], token_index: Int, expert_slot: Int = Int(0), expert_start: Int = Int(0))
Copy the message to the output tensor. This function needs to be called by all threads in a warp.
expert_slot (= expert_id + shared_expert_offset) and expert_start
(the expert's first output row) are supplied by the tile loop and used
only by formats that fold the grouped-matmul scale preshuffle into this
copy (MXFP4 KS224); other formats ignore them.
get_type_name
static def get_type_name() -> String
Gets the name of the host type (the one implementing this trait). For example, Int would return "Int", DeviceBuffer[DType.float32] would return "DeviceBuffer[DType.float32]". This is used for error messages when passing types to the device. TODO: This method will be retired soon when better kernel call error messages arrive.
Returns:
String: The host type's name.
Provided methods
src_info_size
static def src_info_size() -> Int
Returns the size of the source info in bytes. Currently, source info is a single int32 that stores a token's index in the original rank.
Returns:
topk_info_size
static def topk_info_size() -> Int
Returns the size of the top-k info in bytes. Currently, top-k info is an array of uint16 that stores a token's top-k expert IDs.
Returns:
msg_size
src_info_offset
static def src_info_offset() -> Int
Returns the offset of the source info in the message.
Returns:
topk_info_offset
static def topk_info_offset() -> Int
Returns the offset of the top-k info in the message.
Returns:
pad_expert_offsets
def pad_expert_offsets[n_groups: Int](self, row_offsets: Pointer[UInt32, address_space=row_offsets.address_space, _safe=False])
Pad the offsets to satisfy the grouped matmul alignment requirement.
init_smem_resources
def init_smem_resources(self)
Initialize the shared memory resources for the token format.
copy_msg_tile_to_output_tensor
def copy_msg_tile_to_output_tensor[extract_topk_info_func: def(Pointer[UInt8, MutUntrackedOrigin, _safe=False], Int) -> None, recv_buf_ptr_func: def(Int) -> Pointer[UInt8, MutUntrackedOrigin, _safe=False], //, n_warps: Int, shared_expert_offset: Int = Int(0)](self, expert_id: Int, expert_start_pos: Int, tile_id: Int, tile_end: Int, extract_topk_info_functor: extract_topk_info_func, recv_buf_ptr_functor: recv_buf_ptr_func)
Copy a tile of tokens from the receive buffer to the output tensor.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!