feat(llama.cpp): experimentelles Qwen3.8-27B DFlash2-Preset (200ctx, Q4_K_M Drafter)

docs/models/Qwen3.8-27B.md

@@ -43,3 +43,12 @@ The section now has a deliberate twin for real-scenario A/B testing instead of a
 **q4_1 KV is disqualified on CUDA despite winning the KLD ladder.** It measured KLD 0.0074 (between q8_0's 0.0039 and q4_0's 0.0099) but generation drops to 30.5 t/s short (q4_0: 43.2) and prompt processing collapses to ~30 t/s vs q4_0's 1989 t/s (~65x) - a 103k prompt was 22 % done after 13 minutes and hit the client timeout. VRAM also lands oddly lower (20618 vs 21241 MiB), consistent with a different, non-fused kernel path. Sharpened rule: **matching K and V is necessary but not sufficient on CUDA - only pairs with a fused FA path are usable, and on this build/model that is q8_0/q8_0 and q4_0/q4_0 (throughput-verified); every other pair needs its own probe before use.**
 
 **Renamed 2026-08-15 (same day):** the twins are now `[Qwen3.8-27B-small]` (ex `[Qwen3.8-27B]`) and `[Qwen3.8-27B-large]` (ex `[Qwen3.8-27B-maxctx]`) in both preset files and in the hermine OpenCode provider (models map + whitelist, live and backup). lieselotte keeps the plain `Qwen3.8-27B` ID - its GGUF and preset were never verified on that box, and renaming the OpenCode entry without touching that machine's models.ini would have broken the ID linkage. This doc file keeps the base-model name.
+
+## DFlash2 speculation section (added 2026-09-05, unmeasured)
+
+`[Qwen3.8-27B-UD-Q4_K_M-200ctx-q4_0-dflash2]` is an **experimental** twin of the 200ctx MTP section: same target GGUF, ctx 200000, q4_0/q4_0 KV, thinking sampling, mmproj on CPU — but `spec-type = draft-dflash` with the external block-diffusion drafter `incoai/Qwen3.8-27B-DFlash2-GGUF` (Q4_K_M, 1.1 GB, added to `models.list`) instead of the embedded MTP head. `spec-draft-n-max = 7` and the Q4_K_M drafter follow the drafter's HF card (incoai/Qwen3.8-27B-DFlash2-GGUF); draft KV q4_0/q4_0 follows the only other dflash section in this file (Muse-Glimmer-30B).
+
+Upstream status: dflash support landed in stock llama.cpp via PR #27816 (merged 2026-08-27), so the prebuilt bootstrap build picks it up; the "build from PR #27342" instructions on the HF card are stale.
+
+**Nothing on this box is measured yet:** whether 200k + target + drafter + draft KV fits in 24 GB (the MTP section alone already needed ~2.6 GiB of embedded-head overhead at 229376; the external drafter adds its own 1.1 GB weights + KV), and whether dflash beats the 1.85× MTP number on this model. The drafter card reports GSM8K acceptance length ~5.1–5.4 (Q4_K_M: 5.39) on the *ggml-org* Q4_K_M target at their sampling settings — a different target quant, different engine, not comparable to the hermine MTP runs. First load should go through `llama-fit-params` / real `nvidia-smi`, and an A/B short-probe against the 200ctx MTP section before this section is trusted or wired into any OpenCode provider.
+

llama.cpp/models.list

@@ -13,6 +13,7 @@
 unsloth/Qwen3.8-27B-GGUF              | Qwen3.8-27B-UD-Q4_K_M.gguf         | Qwen3.8-27B
 unsloth/Qwen3.8-27B-GGUF              | Qwen3.8-27B-UD-Q4_K_XL.gguf         | Qwen3.8-27B
 unsloth/Qwen3.8-27B-GGUF              | mmproj-F16.gguf                    | Qwen3.8-27B
+incoai/Qwen3.8-27B-DFlash2-GGUF       | Qwen3.8-27B-DFlash2-Q4_K_M.gguf    | Qwen3.8-27B
 unsloth/Qwen3.6-27B-MTP-GGUF          | Qwen3.6-27B-IQ4_XS.gguf            | Qwen3.6-27B
 unsloth/Qwen3.6-35B-A3B-MTP-GGUF      | Qwen3.6-35B-A3B-UD-IQ4_XS.gguf     | Qwen3.6-35B-A3B
 unsloth/gemma-4-26B-A4B-it-qat-GGUF   | gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf | gemma-4-26B-A4B

llama.cpp/presets/models.example.ini

@@ -54,6 +54,45 @@ cache-ram = 16384
 load-mode = dio
 kv-unified = true
 
+# Experimental twin of the 200ctx section above: same target GGUF, same 200k
+# / q4_0 KV / sampling, but speculation runs on the external DFlash2
+# block-diffusion drafter (incoai/Qwen3.8-27B-DFlash2-GGUF, Q4_K_M, ~1.1 GB)
+# instead of the embedded MTP head. Upstream llama.cpp support (spec-type
+# draft-dflash), n-max 7 per the drafter's HF card. NOT yet measured on
+# hermine: 200k load with drafter + draft KV, and speed delta vs draft-mtp,
+# are open. Draft cache q4_0/q4_0 follows the Muse-Glimmer dflash section.
+# Drafter download: models.list (incoai/Qwen3.8-27B-DFlash2-GGUF).
+[Qwen3.8-27B-UD-Q4_K_M-200ctx-q4_0-dflash2]
+alias = Qwen3.8-27B-UD-Q4_K_M-DFlash2
+model = C:/Users/Anwender/AppData/Local/llama.cpp/models/Qwen3.8-27B/Qwen3.8-27B-UD-Q4_K_M.gguf
+mmproj = C:/Users/Anwender/AppData/Local/llama.cpp/models/Qwen3.8-27B/mmproj-F16.gguf
+mmproj-offload = false
+model-draft = C:/Users/Anwender/AppData/Local/llama.cpp/models/Qwen3.8-27B/Qwen3.8-27B-DFlash2-Q4_K_M.gguf
+device = CUDA0
+ctx-size = 200000
+n-gpu-layers = -1
+cache-type-k = q4_0
+cache-type-v = q4_0
+cache-type-k-draft = q4_0
+cache-type-v-draft = q4_0
+flash-attn = on
+jinja = true
+reasoning-format = auto
+reasoning-preserve = true
+parallel = 1
+cache-reuse = 256
+fit = off
+spec-type = draft-dflash
+spec-draft-n-max = 7
+temp = 1.0
+top-p = 0.95
+top-k = 20
+min-p = 0.0
+presence-penalty = 0.0
+cache-ram = 16384
+load-mode = dio
+kv-unified = true
+
 [Qwen3.8-27B-UD-Q4_K_M-150ctx-q8_0]
 alias = Qwen3.8-27B-small
 model = C:/Users/Anwender/AppData/Local/llama.cpp/models/Qwen3.8-27B/Qwen3.8-27B-UD-Q4_K_M.gguf