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

P3PeerSendConfig

struct P3PeerSendConfig

Runtime state for the in-epilogue EP-combine peer scatter-send.

Bundles everything needed to resolve and issue a per-row peer send directly from the L2 epilogue's SMEM. One struct instead of ~8 separate new params: write_absolute_with_bounds_check has 3 callers spanning 2 kernel families unrelated to EP-comm (blockwise_fp8_1d2d, blockwise_fp8_output_writer) that must stay byte-identical, so this adds exactly one new defaulted param to their call sites, not eight.

recv_count_layout((e, rk)) = e*n_ranks + rk and recv_buf_layout((src_idx, src_topk, byte_off)) = (src_idx*top_k + src_topk)*msg_bytes + byte_off are the row-major linear-index formulas EPCombineKernel computes for those coordinate tuples -- copied here as plain arithmetic rather than threading the full EPCombineKernel[...] comptime specialization (num_threads/n_sms/n_experts/max_tokens_per_rank/p2p_world_size), none of which either formula depends on. Same precedent as this file's own _stage2_recv_offset_in_bounds, a documented copy of EPCombineKernel._recv_offset_in_bounds for the identical reason.

disabled() is the default: .unsafe_dangling() pointers (never dereferenced -- gated behind p3_control >= 0) and 0-valued scalars.

Fields

• ​atomic_counter (Pointer[Int32, MutUntrackedOrigin]):
• ​src_info_ptr (Pointer[Int32, ImmUntrackedOrigin]):
• ​row_base_ptr (Pointer[UInt64, MutUntrackedOrigin]):
• ​recv_buf_ptrs (StaticTuple[Pointer[UInt8, MutUntrackedOrigin], Int(8)]):
• ​n_ranks (Int):
• ​p2p_world_size (Int):
• ​top_k (Int):
• ​msg_bytes (Int):
• ​max_tokens_per_rank (Int):
• ​my_rank (Int):
• ​send_mbx (P5SendMailboxPtr):
• ​send_stamp_ptr (Pointer[UInt64, MutUntrackedOrigin]):
• ​send_stamp_on (Int):
• ​trace_ptr (Pointer[UInt64, MutUntrackedOrigin]):
• ​trace_on (Int):
• ​send_trace_slot (Int):

Implemented traits

AnyType, Copyable, Deinitable, ImplicitlyCopyable, Movable

Methods

disabled

static def disabled() -> Self