Release v1.36.0

Author: countzero

CHANGELOG.md

@@ -7,13 +7,22 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+
+## [1.36.0] - 2026-06-08
+
+### Added
+- [Presets] Add gemma-4-26B-A4B and gemma-4-12B entries to 24 GB tier
+
 ### Changed
 - [Build] Default build parallelism to 80% of cores on non-SMT CPUs, physical cores on SMT CPUs
 - [Presets] Add direct-io, no-mmap, and fit=off to Qwen3.6 and gemma 16/24 GB tiers
 - [Presets] Add draft KV cache types (cache-type-k-draft/cache-type-v-draft) to spec-decode entries
 - [Presets] Combine draft-mtp with ngram-mod on Qwen3.6 16 GB tiers
 - [Presets] Rework gemma-4-31B 16 GB entry (q4_0 KV cache, no-mmproj-offload)
-- [Vendor] Bump llama.cpp submodule to d4c8e2c
+- [Presets] Rework gemma 24 GB tier to QAT q4_0 unquantized models with draft-mtp,ngram-mod speculative decoding
+- [Presets] Set parallel=1 on Qwen3.6 24 GB entries
+- [Presets] Lower Qwen3.6-27B 24 GB spec-draft-n-max to 3
+- [Vendor] Bump llama.cpp submodule to 42a0afd
 
 ### Removed
 - [Presets] Drop gemma-4-26B-A4B and Qwen3-Coder-Next from 16 GB tier

presets/models_24GB_VRAM.ini

@@ -1,11 +1,79 @@
-[gemma-4-31B-it.IQ4_XS.gguf]
-alias = gemma-4-31B-it.IQ4_XS.gguf
-model = D:\AI\LLM\gguf\gemma-4-31B-it\gemma-4-31B-it.IQ4_XS.gguf
-mmproj = D:\AI\LLM\gguf\gemma-4-31B-it\mmproj.gemma-4-31B-it.BF16.gguf
+[gemma-4-31B-it-qat-q4_0.IQ4_XS.gguf]
+alias = gemma-4-31B-it-qat-q4_0.IQ4_XS.gguf
+model = D:\AI\LLM\gguf\gemma-4-31B-it-qat-q4_0-unquantized\gemma-4-31B-it-qat-q4_0-unquantized.IQ4_XS.gguf
+mmproj = D:\AI\LLM\gguf\gemma-4-31B-it-qat-q4_0-unquantized\mmproj.gemma-4-31B-it-qat-q4_0-unquantized.BF16.gguf
+spec-draft-model = D:\AI\LLM\gguf\gemma-4-31B-it-qat-q4_0-unquantized-assistant\mtp-gemma-4-31B-it-qat-q4_0-unquantized-assistant.Q8_0.gguf
+no-mmproj-offload = true
+ctx-size = 156928
+threads = 24
+parallel = 1
+n-gpu-layers = -1
+cache-type-k = q4_0
+cache-type-v = q4_0
+cache-type-k-draft = q8_0
+cache-type-v-draft = q8_0
+flash-attn = true
+reasoning = on
+jinja = true
+kv-unified = true
+ctx-checkpoints = 32
+cache-ram = 51200
+min-p = 0.0
+top-p = 0.95
+top-k = 64
+temp = 1.0
+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
+direct-io = true
+no-mmap = true
+fit = off
+
+[gemma-4-26B-A4B-it-qat-q4_0.IQ4_XS.gguf]
+alias = gemma-4-26B-A4B-it-qat-q4_0.IQ4_XS.gguf
+model = D:\AI\LLM\gguf\gemma-4-26B-A4B-it-qat-q4_0-unquantized\gemma-4-26B-A4B-it-qat-q4_0-unquantized.IQ4_XS.gguf
+mmproj = D:\AI\LLM\gguf\gemma-4-26B-A4B-it-qat-q4_0-unquantized\mmproj.gemma-4-26B-A4B-it-qat-q4_0-unquantized.BF16.gguf
+spec-draft-model = D:\AI\LLM\gguf\gemma-4-26B-A4B-it-qat-q4_0-unquantized-assistant\mtp-gemma-4-26B-A4B-it-qat-q4_0-unquantized-assistant.Q8_0.gguf
 mmproj-offload = true
 ctx-size = 262144
 threads = 24
+parallel = 1
+n-gpu-layers = -1
+cache-type-k = q8_0
+cache-type-v = q8_0
+cache-type-k-draft = q8_0
+cache-type-v-draft = q8_0
+flash-attn = true
+reasoning = on
+jinja = true
+kv-unified = true
+ctx-checkpoints = 32
+cache-ram = 51200
+min-p = 0.0
+top-p = 0.95
+top-k = 64
+temp = 1.0
+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
+direct-io = true
+no-mmap = true
+fit = off
+
+[gemma-4-12B-it-qat-q4_0.IQ4_XS.gguf]
+alias = gemma-4-12B-it-qat-q4_0.IQ4_XS.gguf
+model = D:\AI\LLM\gguf\gemma-4-12B-it-qat-q4_0-unquantized\gemma-4-12B-it-qat-q4_0-unquantized.IQ4_XS.gguf
+mmproj = D:\AI\LLM\gguf\gemma-4-12B-it-qat-q4_0-unquantized\mmproj.gemma-4-12B-it-qat-q4_0-unquantized.BF16.gguf
+spec-draft-model = D:\AI\LLM\gguf\gemma-4-12B-it-qat-q4_0-unquantized-assistant\mtp-gemma-4-12B-it-qat-q4_0-unquantized-assistant.Q8_0.gguf
+mmproj-offload = true
+ctx-size = 1048576
+threads = 24
 parallel = 4
+n-gpu-layers = -1
 cache-type-k = q8_0
 cache-type-v = q8_0
 cache-type-k-draft = q8_0
@@ -16,17 +84,18 @@ jinja = true
 kv-unified = true
 ctx-checkpoints = 32
 cache-ram = 51200
-min-p = 0.01
+min-p = 0.0
 top-p = 0.95
 top-k = 64
 temp = 1.0
-spec-type = ngram-mod
+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
 direct-io = true
 no-mmap = true
-fit-target = 1024
+fit = off
 
 [Qwen3.6-27B.IQ4_XS.gguf]
 alias = Qwen3.6-27B.IQ4_XS.gguf
@@ -35,7 +104,7 @@ mmproj = D:\AI\LLM\gguf\Qwen3.6-27B\mmproj.Qwen3.6-27B.BF16.gguf
 mmproj-offload = true
 ctx-size = 262144
 threads = 24
-parallel = 2
+parallel = 1
 n-gpu-layers = -1
 cache-type-k = q4_0
 cache-type-v = q4_0
@@ -54,7 +123,7 @@ top-k = 20
 temp = 0.6
 presence-penalty = 0
 spec-type = draft-mtp,ngram-mod
-spec-draft-n-max = 4
+spec-draft-n-max = 3
 spec-ngram-mod-n-match = 24
 spec-ngram-mod-n-min = 48
 spec-ngram-mod-n-max = 64
@@ -69,7 +138,7 @@ mmproj = D:\AI\LLM\gguf\Qwen3.6-35B-A3B\mmproj.Qwen3.6-35B-A3B.BF16.gguf
 mmproj-offload = true
 ctx-size = 262144
 threads = 24
-parallel = 2
+parallel = 1
 n-gpu-layers = -1
 cache-type-k = q4_0
 cache-type-v = q4_0