b70-optimization-lab

TP4 ring readiness marker gate

Date: 2026-07-15

Outcome

The prerequisite for a resident MHC consumer passes. oneCCL commit 1edec4576f6424f645fcee8b18df6844c636423c adds the default-off CCL_SYCL_ALLREDUCE_LL=ring_markers route. The production ring arithmetic and forwarding remain unchanged. After each final local wire writeback, the ring release-publishes its current epoch in storage reserved from the unused tail of the LL scatter buffer.

All four ranks passed an exact smoke test and 24 changing epochs. Paired slowest-rank device times across 87 BF16 [4096] reductions were:

• ring control A: 5.165472 ms;
• ring plus readiness markers: 5.204264 ms;
• ring control B: 5.395494 ms.

The conservative comparison is against the faster control. Marker publication therefore adds 0.038792 ms / 87, or 0.446 us per boundary, below the hard 1 us gate. Against the two-control mean, the candidate is faster by 0.076219 ms; that is measurement noise, not a claimed speedup.

Why it matters

The earlier two-stream screen hid 0.612-0.642 ms only when independent MHC work was resident before communication. A real MHC consumer cannot begin from an ordinary post-ring dependency: it must already be resident and learn when individual reduced wire ranges are locally complete. This gate proves the ring can expose that information without consuming the overlap budget.

This is not yet a decode or collective speed result. The next experiment must have a second-queue consumer acquire-wait on the epochs, compute canonical MHC post work per ready wire, converge, then execute canonical MHC pre. It must preserve bitwise changed-state output and save at least 6 us per boundary at the slowest rank before any full-model integration.

Raw evidence:

• /mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/ring-markers-smoke-20260715T0352Z.json;
• /mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/ring-markers-control-a-20260715T0353Z.json;
• /mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/ring-markers-candidate-20260715T0353Z.json;
• /mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/ring-markers-control-b-20260715T0353Z.json.

The tested library SHA-256 is eb77b800754754c9f9ede04bff6ab0850a8e82a398c0c23baa2d38a10bbaef3f.

This site is open source. Improve this page.