feat(llama.cpp): experimentelles Qwen3.8-27B DFlash2-Preset (200ctx, Q4_K_M Drafter) - #9
Merged
Merged
Conversation
Add [Qwen3.8-27B-UD-Q4_K_M-200ctx-q4_0-dflash2] as an experimental twin of the 200ctx MTP section: same target GGUF, 200k ctx, q4_0/q4_0 KV and thinking sampling, but speculation 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. spec-draft-n-max 7 per the drafter card; draft KV q4_0 follows the Muse-Glimmer dflash section. Drafter added to models.list. Experimental: 200k load with drafter and speed delta vs draft-mtp are not measured on hermine yet (no GPU in this environment). DFlash2 support is in stock llama.cpp (PR #27816, merged 2026-08-27; present in v0.4.0). Rationale and open questions: docs/models/Qwen3.8-27B.md. Co-Authored-By: Tim Köster <40566351+tilloh-dev@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Neues experimentelles Preset für Qwen3.8-27B mit DFlash2-Spekulation, als Twin zur bestehenden 200ctx/MTP-Variante:
[Qwen3.8-27B-UD-Q4_K_M-200ctx-q4_0-dflash2]: gleiche Ziel-GGUF, 200k ctx, q4_0/q4_0 KV, Thinking-Sampling, mmproj auf CPU — aberspec-type = draft-dflashmit externem Block-Diffusion-Drafter (incoai/Qwen3.8-27B-DFlash2-GGUF, Q4_K_M, 1.1 GB) statt eingebautem MTP-Head.spec-draft-n-max = 7und Quant Q4_K_M nach der Drafter-Model-Card (Acceptance Length ~5.4 auf deren Target).q4_0/q4_0folgt dem einzigen anderen dflash-Section im Repo (Muse-Glimmer-30B).models.list(landet neben der Ziel-GGUF inQwen3.8-27B/).docs/models/Qwen3.8-27B.md.Experimentell: 200k-Load mit Drafter (VRAM) und Geschwindigkeits-Delta vs.
draft-mtpsind auf hermine nicht gemessen (keine GPU in dieser Umgebung). Erster Load sollte überllama-fit-params/nvidia-smi+ kurzer A/B-Probe gegen die MTP-Section gehen. Kein OpenCode-Provider-Wechsel in diesem PR.Test Plan
llama-server --help(lokal verifiziert)./download-model.sh --allholt den Drafter (manifest-only, nicht hier getestet)llama-fit-params/ echter Load @200k mit Drafter → VRAM-Marge prüfen