docs(design): PD-dflash-MoE serving design doc

[drunkcoding]

docs(design): DFlash-on-MoE-Infinity serving design doc

Design/architecture proposal — docs-only. Adds
docs/design/pd-dflash-moe-serving.md.

What "DFlash" is

DFlash = Chen, Liang, Liu, "DFlash: Block Diffusion for Flash Speculative
Decoding" (Z Lab, z-lab.ai, arXiv:2602.06036):
a lightweight (~0.8 B) block-diffusion draft model proposes a whole block of B
tokens in one parallel forward; the large target verifies them in parallel
(lossless). z-lab ships ready draft/target pairs on HuggingFace.

Thesis (why DFlash × expert-offloading is synergistic)

For an offloaded-MoE target, the bottleneck is fetching the target's experts for
each verification step. A DFlash draft proposes the entire candidate block before
the target verifies it, so we can run the target router on the draft tokens,
obtain the exact expert set the width-B verify will read, and prefetch those
experts during the draft forward — a route-ahead prefetch (near-exact, not a
speculative guess). DFlash turns MoE-Infinity's hardest problem (hiding target expert
transfer) into a scheduled, near-exact prefetch.

Extends the existing spec_decode/dflash.py (DFlashSpeculator) seed — a
re-composition, not a rewrite.

Scope / target

• Hardware: single sm_120 / RTX PRO 6000 (Blackwell), local; target MoE
  experts FP4-offloaded to host RAM (moe_infinity._v4_fp4); DFlash draft resident.
• Generalizes to (must-cover): Qwen and GPT-OSS MoE targets that
  MoE-Infinity supports and have z-lab DFlash drafts — Qwen3-Coder-30B-A3B,
  Qwen3.5-35B-A3B, gpt-oss-20b, gpt-oss-120b (DeepSeek-V4-Flash when its draft
  lands). Route-ahead prefetch is architecture-agnostic at the MoE level.

The doc covers

Verification kernels (width-B target forward), draft/verify PD scheduling (2-D
deficit over {tokens, expert-bytes}), route-ahead expert prefetch (the
differentiator), an analytical cost model (route-ahead hiding inequality), and an
evaluation plan on the RTX PRO 6000. 3 Mermaid diagrams; 19-row symbol table.

Honesty notes (preserved)

• Gating experiment: measure route-ahead prefetch coverage and rejected-token
  waste vs. acceptance a; if coverage is high / waste low, the thesis holds.
• The advantage is regime-scoped: it pays in the memory-constrained (offloaded,
  cold-cache) regime; honest elsewhere. Cost-model numbers are labeled illustrative.

Not in scope

No implementation, no training, no measured benchmark numbers. Draft; not for merge
until reviewed.

[drunkcoding]

Ready for review. Full design doc landed in ed8a07d (11 sections + TL;DR, 3 Mermaid diagrams, analytical cost model, evaluation plan, 19-row symbol table). F1 QA green; changes are docs-only (docs/design/pd-dflash-moe-serving.md + an ARCHITECTURE.md link). Note: this is a design/architecture proposal — the thesis is gated on measuring per-step expert churn q (stability study first), and the AR-vs-diffusion advantage is stated as regime-scoped. Not auto-merged.