qwen35: DSpark speculative decoding (Qwen3.8-27B drafters)

[davide221]

Adds DSpark drafter support to the qwen35 family spec-decode loop, targeting the published Qwen3.8-27B DSpark checkpoints (DFlash backbone + low-rank Markov bigram head + confidence head, e.g. RadixArk/Qwen3.8-27B-DSpark).

What's in here

Spec loop (qwen35_backend.cpp)

• When the drafter GGUF ships DSpark heads, the chain draft goes through the markov-corrected greedy chain (fused single-graph variant, non-fused fallback) instead of plain argmax projection. DDTree candidate top-k also gets the markov bias.
• Env gates, all default-on: DFLASH_QWEN35_DSPARK, DFLASH_QWEN35_FUSED_DSPARK, DFLASH_QWEN35_DSPARK_TREE.
• Target capture layers now follow the drafter GGUF's dflash.target_layer_ids instead of the evenly-spaced derivation. The Qwen3.8 drafter is trained on layers 4/16/28/40/52; capturing the derived 1/16/31/46/61 silently destroys acceptance.

Draft loader (draft_gguf_loader.cpp)

• dflash.mask_token_id from the drafter GGUF wins over the family default (this drafter trains with MASK 248077; the default was 248070 — another silent acceptance killer).
• Optional YaRN rope-scaling keys (rope.scaling.*) parsed into DraftWeights.

Draft graph (draft_graph.cpp)

• All rope calls go through one helper that honors the drafter's YaRN config; previously plain NEOX rope was hardcoded.

Converter (convert_dflash_to_gguf.py)

• Single-file DSpark releases (markov/confidence heads inline in model.safetensors, no dflash_aux_heads.pt sidecar) now convert directly.
• Reads transformers>=5 nested rope_parameters (theta + YaRN) and dflash_config.mask_token_id; emits YaRN scaling metadata.

Fused head path: Qwen35DFlashTarget exposes lm_head_tensor() (the fused chain needs the target lm_head; layer-split target falls back gracefully).

Validation (R9700 / gfx1201, Qwen3.8-27B Q4_K_M, greedy, 300-token gens)

	code	prose	mixed
AR baseline	22.9	~22	~22
DSpark spec decode	38.4	23.0	30.2

Serving config for the numbers: q4-mix requantized drafter, --fa-window 2048, q8_0 KV, DFLASH_SINGLE_CHAIN_CHECKPOINT_F32=1 DFLASH_FAST_ROLLBACK_THRESHOLD=1. Acceptance length matches the checkpoint's published SGLang reference (~3.4 tokens/step incl. bonus).

Known limitations / follow-ups

• Confidence-gate adaptive block length is not wired (per-request step buffers are sized by q_len); the chain runs with the gate off.
• YaRN on the drafter measured slightly worse acceptance than plain rope at short context on this checkpoint, so the canonical drafter GGUF is converted without it; the plumbing is data-driven either way.
• DDTree with markov candidates loses to chain verify on prose with this drafter; chain remains the default.
• Verify/AR forward-cost optimization (AMD-native requant formats) is separate follow-up work.