(WIP) perf(comm): add two-stage producer-direct reduction

[panditsa]

Summary

• keep one-stage Iris reduction for small producer-direct payloads
• switch world-size 4/8 payloads above the dtype-aware threshold to reduce-scatter plus all-gather
• retain BF16, FP16, and FP32 packing behind the existing symmetric-output API

Performance

Exact parent 2ebd8f81 → head 60babeb9, world-size 8 BF16 producer-direct reduction under CUDA-graph replay, critical-rank p50:

Payload	M	Parent	Head	Reduction
AttnRes [M,7168]	2	9.26 µs	7.96 µs	14.0%
AttnRes [M,7168]	4	10.68 µs	8.70 µs	18.6%
AttnRes [M,7168]	8	14.52 µs	10.25 µs	29.4%
AttnRes [M,7168]	16	21.04 µs	10.47 µs	50.2%
Joint MoE [M,3584]+[M,7168]	2	9.31 µs	8.30 µs	10.8%
Joint MoE [M,3584]+[M,7168]	4	11.93 µs	9.68 µs	18.9%
Joint MoE [M,3584]+[M,7168]	8	16.88 µs	10.41 µs	38.3%
Joint MoE [M,3584]+[M,7168]	16	26.99 µs	10.75 µs	60.2%

M=2/4 remain on the one-stage path; M=8/16 cross the world-size-8 BF16 two-stage threshold.

Tests

• threshold, world-size 4/8, noncontiguous subgroup, large-payload, dtype, and repeated graph-replay coverage: 7 passed
• benchmark validates exact reduced values before timing each eager/captured operation

Stack

• depends on (WIP) perf(kimi-k3): tune small-M MoE decode #1140

[jhinpan]

@panditsa I am validating the #1132 M=8 path on current main (9196fa27) and want to coordinate rather than duplicate this stack.

Exact 131072/1024 TP8/EP8 forced-acceptance=1 now measures 36.80 ms/round versus 17.06 ms ordinary decode (2.157x); forced DSpark is token-identical to ordinary greedy on both a 4K/1K repeated prompt and an exact 131K/128 natural prompt. An all-rank first-64 trace reproduces 5,888 MoE reductions and 11,968 AttnRes kernels. The important caveat for this PR is producer skew: summed attention/MoE Iris durations range from 0.666/0.496 s on rank 5 to 2.624/1.382 s on rank 7, so I will judge #1141 on unprofiled wall time, not isolated reduction latency.

Unless you prefer a different arrangement, I plan to benchmark these existing changes independently on current main: #1141, the M=8 add3 portion of #1135, #1142 plus the communication-only gate from #1146, and #1140's router-warp line. I will not use #1145 or open overlapping kernel PRs; I will post exact results back to the owning PRs first.

[jhinpan]

Exact #1132 result on 9196fa27: this commit applied standalone did not reach a benchmarkable serving state. The engine became ready, but 1-token health requests took about 4.5–5.5 s and SMG's TokenSpeed gRPC probe repeatedly timed out, so the worker never registered. I stopped the run rather than publish an isolated-kernel extrapolation.

The same #1141 code inside #1141 + #1142 + #1146 does register and completes; that full stack improves forced 131K/1K TPOT from 36.80 to 36.25 ms (-1.49%, three runs). This suggests the downstream state/orchestration changes are required for an E2E-valid use of the two-stage path.

[panditsa]

@panditsa I am validating the #1132 M=8 path on current main (9196fa27) and want to coordinate rather than duplicate this stack.

Exact 131072/1024 TP8/EP8 forced-acceptance=1 now measures 36.80 ms/round versus 17.06 ms ordinary decode (2.157x); forced DSpark is token-identical to ordinary greedy on both a 4K/1K repeated prompt and an exact 131K/128 natural prompt. An all-rank first-64 trace reproduces 5,888 MoE reductions and 11,968 AttnRes kernels. The important caveat for this PR is producer skew: summed attention/MoE Iris durations range from 0.666/0.496 s on rank 5 to 2.624/1.382 s on rank 7, so I will judge #1141 on unprofiled wall time, not isolated reduction latency.

Unless you prefer a different arrangement, I plan to benchmark these existing changes independently on current main: #1141, the M=8 add3 portion of #1135, #1142 plus the communication-only gate from #1146, and #1140's router-warp line. I will not use #1145 or open overlapping kernel PRs; I will post exact results back to the owning PRs first.

These optimizations help with 4K/1K scenarios but I have not been using/looking at dspark and 131072/1024 performance. Feel free to continue with your investigations. I am glad these PRs are at least getting us to serving state :)