Perf: A/B DSpark speculative depth k=7 vs k=5 for DeepSeek cycle (AICAD runs k=7)

[Saolence]

Summary

A/B the DSpark speculative depth for the DeepSeek cycle profiles: k=7 vs the current k=5. Speculative decoding is the only mechanism that amortises the ~87 per-step allreduce latency on this topology (more draft tokens per forward pass = fewer collectives per output token), so a higher k that keeps acceptance high directly raises decode throughput. The AICAD TP4 production deployment runs DSpark at k=7 static (no ladder) with a smoothly decaying acceptance curve (0.94/0.86/0.77/0.69/0.61/0.49/0.35); SparkRing ships k=5.

Evidence

• AICAD production config (same 0731 DSpark checkpoint, 4× DGX Spark): dspark k=7 静态（无 ladder）, acceptance 0.94…0.35 smooth decay — see config/production-nccl-env.md in https://github.com/luxingcom/aicad-nccl-optimization.
• Their analysis: decode is latency-bound on ~87 small allreduces/step; wider speculation is the "already working" mechanism that enlarges each collective and dilutes per-step latency.
• SparkRing's current Mean acceptance length on the deployed cycle is ~3.23 at k=5 (measured 2026-09-03). If k=7 holds a comparable per-position acceptance curve, decode could gain roughly +10–30% (7 vs 5 draft slots × preserved acceptance).

Current state in sparkring

• DeepSeek env templates: NUM_SPECULATIVE_TOKENS=5.
• --speculative-config '{"method":"dspark","num_speculative_tokens":5,"moe_backend":"b12x"}'.

Proposed change

Pure configuration A/B — no code change:

1. Set NUM_SPECULATIVE_TOKENS=7 in all four cycle rank envs and relaunch.
2. First verify the checkpoint/image accepts k=7: the 0731 DSpark checkpoint's dspark_block_size is 5 (k<5 silently truncates; the constraint above 5 must be checked on the sparkring 0.27 image — the AICAD lane ran k=7 on vLLM 0.26, so it is feasible on this checkpoint).
3. Measure Mean acceptance length + decode c1/c4/c8 before/after (nonce-fronted, same benchmark).

If k=7 is rejected at startup, fall back to testing a k-ladder (e.g. 5,7) or k=6.

Reference

• AICAD production snapshot: https://github.com/luxingcom/aicad-nccl-optimization config/production-nccl-env.md (dspark k=7, acceptance curve).
• SparkRing quickstart records the DSpark contract at k=5 (docs/DEEPSEEK_V4_FLASH_QUICKSTART.md).

Risks / notes

• Zero code risk; one relaunch per arm.
• Watch startup logs for a dspark block-size rejection before benchmarking.
• Keep decode benchmarks nonce-fronted so prefix caching cannot inflate numbers.

Verification

• Startup succeeds at k=7; log shows speculative depth 7.
• Mean acceptance length and decode c1/c4/c8 aggregate at k=7 vs k=5 (same benchmark script, ≥3 reps each).
• If acceptance per position does not decay meaningfully at k=7, promote; otherwise revert to k=5.