feat(gallery): add Qwen DFlash speculative-decoding models

[localai-bot]

What

Adds four ready-to-run DFlash speculative-decoding entries to the model gallery for the llama-cpp backend, now that upstream DFlash support (draft-dflash, ggml-org/llama.cpp#22105, merged 2026-06-28) is in the pinned llama.cpp (e3546c7).

Gallery name	Target (Q4_K_M)	DFlash drafter (Q8_0)
qwen3-4b-dflash	unsloth/Qwen3-4B-GGUF	lym00/Qwen3-4B-DFlash-GGUF-Test
qwen3.5-9b-dflash	unsloth/Qwen3.5-9B-GGUF	lym00/Qwen3.5-9B-DFlash-GGUF-Test
qwen3.6-27b-dflash	unsloth/Qwen3.6-27B-GGUF	Alittlehammmer/Qwen3.6-27B-DFlash-GGUF-llama.cpp
qwen3.6-35b-a3b-dflash	unsloth/Qwen3.6-35B-A3B-GGUF (UD-Q4_K_M)	Alittlehammmer/Qwen3.6-35B-A3B-DFlash-GGUF-llama.cpp

Each entry bundles a full target model plus its small z-lab block-diffusion drafter and configures:

flash_attention: "on"          # DFlash requires flash attention
draft_model: <drafter>.gguf
options:
  - use_jinja:true
  - spec_type:draft-dflash
  - spec_n_max:15

How DFlash works

The drafter emits a whole block of tokens in one forward pass (block diffusion) and consumes the target model's hidden states, so it stays tiny while drafting is GPU-friendly. The drafter is not a standalone chat model; it only runs paired with its target, which is why both are bundled per entry. A GPU is recommended.

Note on the 4B entry

It pairs the base Qwen3-4B target (not Qwen3.5-4B). The only upstream-compatible 4B drafter reports general.name = "Qwen3 4B DFlash" / general.basename = "Qwen3", i.e. it is trained for Qwen3-4B and is the canonical pairing documented upstream (z-lab/Qwen3-4B-DFlash + Qwen/Qwen3-4B). The Anbeeld / psychopenguin "Qwen3.5-4B-DFlash" GGUFs use the fork-only dflash-draft / DFlashDraftModel architecture and would fail to load in the upstream backend, so they were deliberately not used.

Verification

• Backend support: confirmed draft-dflash is registered in the pinned llama.cpp (common_speculative_type_from_name_map).
• Drafter load-compatibility: every drafter was downloaded and its GGUF general.architecture confirmed to be dflash (not dflash-draft / DFlashDraftModel).
• Integrity: every drafter SHA256 was verified against the downloaded bytes (they equal the HF x-linked-etag, which is distinct from x-xet-hash on these Xet repos, i.e. the true content hash). Target SHA256s use the HF content hash.
• go test ./core/config/meta/ and go test ./core/gallery/ pass (override keys valid, index parses).

Not verified: end-to-end acceptance/speedup on a GPU (config-and-integrity verification per scope).

🤖 Generated with Claude Code