mtp_ep_propose_enabled

spark_model::speculative

Function mtp_ep_propose_enabled 

Source 
pub fn mtp_ep_propose_enabled() -> bool
Expand description

Run the drafter on EVERY rank with the communicator, instead of rank-0-only with comm: None. DEFAULT ON since 2026-08-29; kill switch ATLAS_NO_MTP_EP_PROPOSE=1 restores the rank-0-only path.

Both halves move together and neither is safe alone:

  • the head broadcasts EP_CMD_MTP_PROPOSE before every propose, so the worker runs the SAME drafter forward and issues the SAME collectives in the same stream order;
  • DraftProposer::needs_comm then hands the block a comm.

Only a proposer that returns true from DraftProposer::needs_comm is affected, and today that is GLM-5.3 alone — the Qwen and DeepSeek-V4 MTP modules load every expert on every rank and must keep comm: None.

Measured on 2 x GB10 (t67, six-probe gate byte-identical on every arm): open512 17.53 -> 19.11 tok/s, p1 0.747 -> 0.875.

🪤 ATLAS_MTP_EP_PROPOSE=1 (the opt-in name it shipped behind for one day) still reads as ON, so a launch script carrying it keeps working.