docs(dflash): plan for gpt-oss-120b DFlash speculative decoding support

[drunkcoding]

What

Adds a phased, de-risked design doc (RFC) for supporting DFlash speculative decoding on openai/gpt-oss-120b in MoE-Infinity, using the z-lab/gpt-oss-120b-DFlash 0.8B block-diffusion drafter (arXiv 2602.06036).

Doc: docs/plans/gpt-oss-dflash-support.md (doc-only PR — no code changes).

Why

gpt-oss-120b is already a supported target, but MoE-Infinity has no speculative-decoding scaffolding today. DFlash is a lightweight, distribution-lossless way to speed up decode. This RFC captures the verified integration contract and a staged, low-risk path before any code lands.

Approach (summary)

• Phase 0 — validate the SGLang reference, fully resident (go/no-go gate): confirm greedy losslessness vs plain gpt-oss-120b + record acceptance length and decode tok/s. If the resident-mode win isn't clearly positive, stop.
• Then build natively into the synchronous engine/generation_loop.py path first — structurally the closest fit (HF past_key_values rollback), greedy first.
• Resident-only for v1. A 9-token verify block would thrash the current next-layer/top-2 expert prefetcher, so expert-offload + spec-decode is explicitly deferred. Async continuous-batching serving is a later phase.

Contents

• Verified DFlash runtime contract (drafter dims, block_size=10 → 9 draft tokens, mask_token_id=200000, target layers [1,9,17,25,33], target-hidden-state KV injection, accept rule).
• MoE-Infinity integration anchors (file:line) for both decode paths.
• Risk table (SGLang-on-Blackwell build, hidden-state extraction, KV rollback, offload amplification, losslessness).
• Executable per-phase QA gates (pytest invocations + pass criteria), including a golden greedy-parity test.

Status

RFC for review. Phase 0 reference validation is being run separately; results will inform the go/no-go on native integration.