#2754 sycl/A770: MTP shared-expert branch off by default on Alchemist (a non-linear draft graph resets ~50 events a replay)

本文

## sycl/A770: MTP shared-expert branch off by default on Alchemist (a non-linear draft graph resets ~50 events a replay)

The shared expert of a draft step was forked onto a side queue and joined before the combine. On Arc A-series every
kernel runs on the one compute engine in submission order, so the branch runs nothing in parallel; but the fork / join
makes each captured draft graph non-linear, and the SYCL graph then replays as a command buffer with an event per node,
reset after every run: ~50 zeCommandListAppendEventReset + 2 barriers per draft graph, 0.45-0.7 ms on the GPU before
the next one (unitrace chrome timeline, A770, Level Zero v2 adapter). Linear, it replays as an in-order command buffer
with none. Same kernels on one stream: the same bits.

A770 (xe, oneAPI 2026.1, UR_LOADER_USE_LEVEL_ZERO_V2=1), Coder IQ1_M, 8K code prompt, 3 pairs with
STRATA_MTP_SHARED_BRANCH=1 vs 0: MTP drafting 7.91 -> 7.27 ms a round (-0.64, every pair), decode within noise,
reference tokens. STRATA_MTP_SHARED_BRANCH=1 forces the branch back on; other devices are unchanged.

🤖 Generated with [Claude Code](https://claude.com/claude-code)