[AzeezIsh]

Summary

• Main's default MTP ladder (speculative/ladder.rs) started with (4, usize::MAX). qwen4_exp: serve NVIDIA's Qwen3.8-Flash-Next FP4-Local on its native formats (one GB10) #195 added that rung so qwen4_exp's MTP depth 4..7 reaches the verify at n≤4.
• On 2×GB10 TP=EP=2, at n=4 that means every sequence drafts all 7 positions. Each verify step gets wider than the extra accepted tokens pay back: C4 aggregate decode is about 11–19% lower than with 3 drafts.
• This PR keeps the passthrough for a lone sequence and restores 3 drafts for n=2..4, the value pre-qwen4_exp: serve NVIDIA's Qwen3.8-Flash-Next FP4-Local on its native formats (one GB10) #195 and SparkQwen 0.4.0 use: [(1, usize::MAX), (4, 3), (8, 3), (16, 1), (32, 1)]. ATLAS_MTP_K_LADDER still overrides it.
• With the deep path on (ATLAS_MTP_DYNAMIC_DEPTH=1, ATLAS_MTP_DEEP_MAX_SEQS=2, as in both profiles below), n≤2 doesn't use the ladder's value, so C1/C2 are unaffected.

How it was found (#220)

• At C4, SparkQwen 0.4.0's engine was 11% ahead of main under the same serving config.
• Bisect by switch: the gap goes away with ATLAS_MTP_MAX_SEQS=1 and stays with the RDMA path off.
• With ATLAS_MTP_ACCEPT_DEBUG, 0.4.0 drafts 3–4 positions at C4 where main drafts 7.
• ATLAS_MTP_K_LADDER='4:3,8:3,16:1,32:1' on main closes it: C4 141–144 vs 0.4.0's 140–144, C1 floor unchanged.

Test plan

• cargo test -p spark-model --lib ladder: 7 passed. Asserts updated: (1,5)=5, (2,5)=3, (4,5)=3, (8,5)=3, (16,5)=1, (1,7)=7.

• clippy and fmt clean.

• Pair check of this branch's code on reiner + spark2 with the rc16 profile. Built fresh from 346ffb6fc; receipts in reiner's ~/code/night1027/out/ladpr-346ffb6fc/.

  	this PR	main (default ladder, sqab9)
  sq_bench conc C4 aggregate	122.99 / 122.34	101.71 / 103.60
  conc C1 / C2	54.02, 53.47 / 86.01, 85.30	53.73, 54.37 / 84.26, 85.06
  decode-floor (1500-token stream)	84.19 / 84.09 / 84.24	84.14–84.60 (n=6)
  decode C1 code / JSON / prose	76.5 / 69.2 / 56.5	76.9 / 69.4 / 56.7

  • C1/C8 greedy text is identical to current main. The one row that differs from qwen4_exp: serve FP4-Local at TP=EP=2 with the n-gram table resident, native MXFP8 sharding, BF16 routing #198's references comes from serving: MODEL.toml [behavior] keys to serve like the reference server; qwen3.8-flash-next adopts them #271's intended thinking-budget change; current main shows the same diff.
  • Tools 21/21.

• On SparkQwen 0.4.0's spark config, via this ladder as ATLAS_MTP_K_LADDER: C4 142.3 vs 125.6, C1/C2/decode-floor unchanged (sqab9).

• Not measured: single-GB10 FP4-Local (qwen4_exp: serve NVIDIA's Qwen3.8-Flash-Next FP4-Local on its native formats (one GB10) #195's target) at n=2..4 without the deep path. There, n=2..4 returns to the pre-qwen4_exp: serve NVIDIA's Qwen3.8-Flash-Next FP4-Local on its native formats (one GB10) #195 value of 3.

Generated with Devin

[AzeezIsh]

Follow-up check: is anything left past C4? Main plus this PR's ladder (ATLAS_MTP_K_LADDER='1:7,4:3,8:3,16:1,32:1', which equals the new default) against SparkQwen 0.4.0's engine, on 0.4.0's config. sq_bench conc; spark rows are 2 boots per arm, strict is one boot. Receipts are in reiner's ~/code/night1027/out/sqab10/.

	main + this ladder	0.4.0 engine
spark C4	140.0 / 140.7 / 144.9 / 145.2	133.2 / 137.9 / 142.9 / 143.9
spark C8 (= max-num-seqs)	208.8 / 208.8 / 209.2 / 210.6	207.8 / 209.1 / 209.9 / 211.9
strict C4	127.8 / 137.5	125.5 / 137.3

With this PR, main matches 0.4.0 at C4 and C8, in both modes.