Re-enable draft-mtp,ngram-mod on Qwen3.6 24GB tiers (RS-rollback fix #23269)

Combined-mode RS-rollback regression (#23184) was fixed upstream in #23269, which is present in the pinned llama.cpp. Re-add ngram-mod fallback to both Qwen3.6 24GB entries, re-enable mmproj-offload, and drop the now-stale 'combined mode broken' warning comment.

Author: countzero

CHANGELOG.md

@@ -8,9 +8,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [Unreleased]
 
 ### Changed
-- [Presets] Switch Qwen3.6-35B-A3B 24 GB tier from ngram-mod to draft-mtp
-- [Presets] Drop ngram-mod from Qwen3.6-27B 24 GB tier (combined mode broken, #23184)
-- [Presets] Disable mmproj-offload on Qwen3.6-27B 24 GB tier
+- [Presets] Combine draft-mtp with ngram-mod on Qwen3.6 24 GB tiers (RS-rollback regression fixed upstream, #23269)
+- [Presets] Re-enable mmproj-offload on Qwen3.6 24 GB tiers
 - [Presets] Raise Qwen3.6-27B 24 GB ctx-size from 172032 to 184320
 
 

presets/models_24GB_VRAM.ini

@@ -25,18 +25,11 @@ spec-ngram-mod-n-match = 24
 spec-ngram-mod-n-min = 48
 spec-ngram-mod-n-max = 64
 
-# WARNING: spec-type draft-mtp,ngram-mod is broken on hybrid Qwen 3.6
-#
-# Speculators run independently (https://github.com/ggml-org/llama.cpp/issues/23184)
-# and any non-MTP speculator unconditionally disables MTP partial rollback on hybrid
-# models (https://github.com/ggml-org/llama.cpp/blob/255582687/common/common.cpp#L1258).
-# This causes ~30% TG regression vs draft-mtp alone. Use either solo, never combined!
-
 [Qwen3.6-27B.IQ4_XS.gguf]
 alias = Qwen3.6-27B.IQ4_XS.gguf
 model = D:\AI\LLM\gguf\Qwen3.6-27B\Qwen3.6-27B.IQ4_XS.gguf
 mmproj = D:\AI\LLM\gguf\Qwen3.6-27B\mmproj.Qwen3.6-27B.BF16.gguf
-no-mmproj-offload = true
+mmproj-offload = true
 ctx-size = 262144
 threads = 24
 parallel = 2
@@ -57,15 +50,18 @@ top-p = 0.95
 top-k = 20
 temp = 0.6
 presence-penalty = 1.5
-spec-type = draft-mtp
+spec-type = draft-mtp,ngram-mod
 spec-draft-n-max = 4
+spec-ngram-mod-n-match = 24
+spec-ngram-mod-n-min = 48
+spec-ngram-mod-n-max = 64
 fit = off
 
 [Qwen3.6-35B-A3B.IQ4_XS.gguf]
 alias = Qwen3.6-35B-A3B.IQ4_XS.gguf
 model = D:\AI\LLM\gguf\Qwen3.6-35B-A3B\Qwen3.6-35B-A3B.IQ4_XS.gguf
 mmproj = D:\AI\LLM\gguf\Qwen3.6-35B-A3B\mmproj.Qwen3.6-35B-A3B.BF16.gguf
-no-mmproj-offload = true
+mmproj-offload = true
 ctx-size = 262144
 threads = 24
 parallel = 2
@@ -86,6 +82,9 @@ top-p = 0.95
 top-k = 20
 temp = 0.6
 presence-penalty = 1.5
-spec-type = draft-mtp
+spec-type = draft-mtp,ngram-mod
 spec-draft-n-max = 3
+spec-ngram-mod-n-match = 24
+spec-ngram-mod-n-min = 48
+spec-ngram-mod-n-max = 64
 fit = off