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 struct
Struct_ep_combine_async_mx
struct Struct_ep_combine_async_mx
Registers the ep.combine_async.mx graph op with the graph compiler.
Implemented traits
Methods
execute
static def execute[combine_dtype: DType, hidden_size: Int, top_k: Int, n_experts: Int, max_token_per_rank: Int, n_gpus_per_node: Int, n_nodes: Int, //, *, mx_min_tokens: Int](atomic_counters: T, input_tokens: T, src_info: T, send_ptrs: T, recv_ptrs: T, recv_count_ptrs: T, total_tokens: UInt32, context: DeviceContext)
ep.combine_async that sends MXFP8 messages for large steps.
When total_tokens reaches mx_min_tokens, each expert output goes
back to its rank as MXFP8 (E4M3 values with an E8M0 scale per 32
elements, about half the bytes), which is lossy; smaller steps send it
exactly. total_tokens must be the same on every rank, and the
matching ep.combine_wait.mx must get the same value and threshold.
Parameters:
- combine_dtype (
DType): Expert output dtype (inferred). - hidden_size (
Int): Model hidden size (inferred). - top_k (
Int): Experts each token is routed to (inferred). - n_experts (
Int): Experts across all GPUs (inferred). - max_token_per_rank (
Int): Most tokens per GPU (inferred). - n_gpus_per_node (
Int): GPUs per node (inferred). - n_nodes (
Int): Physical nodes (inferred). - mx_min_tokens (
Int): Smallesttotal_tokenssent as MXFP8.
Args:
- atomic_counters (
T): Atomic counters coordinating the combine. - input_tokens (
T): Expert outputs. Shape[num_tokens, hidden_size]. - src_info (
T): Each row's originating rank and token from the dispatch. - send_ptrs (
T): Send buffer pointers for the combine phase. - recv_ptrs (
T): Receive buffer pointers for the combine phase. - recv_count_ptrs (
T): Receive count buffer pointers. - total_tokens (
UInt32): Tokens of the step across all ranks. - context (
DeviceContext): GPU device context for the current device.