# 32 GiB VRAM (dedicated AI GPU, single user) # /home/hans/ai/models/llm_presets/32gb.ini [*] flash-attn = on gpu-layers = all parallel = 1 batch-size = 4096 ubatch-size = 4096 load-mode = none spec-type = draft-mtp spec-draft-n-max = 3 image-min-tokens = 1024 image-max-tokens = 2048 # KV cache VRAM # Qwen3.8-27B: 16 attn layers × 4 KV heads × 256 dim × 2 (K+V) # q8_0/q8_0 : 1.0625 B/elem → 34 KiB/token → GiB ≈ ctx / 30840.47 # https://datakeskus.io/files/qwen38-27b-kv-cache.txt cache-type-k = q8_0 cache-type-v = q8_0 #ctx-size = 2048 # 0.07 GiB #ctx-size = 4096 # 0.13 GiB #ctx-size = 8192 # 0.27 GiB #ctx-size = 16384 # 0.53 GiB #ctx-size = 32768 # 1.06 GiB #ctx-size = 49152 # 1.59 GiB #ctx-size = 65536 # 2.12 GiB #ctx-size = 87040 # 2.82 GiB #ctx-size = 98304 # 3.19 GiB #ctx-size = 131072 # 4.25 GiB #ctx-size = 163840 # 5.31 GiB #ctx-size = 196608 # 6.38 GiB #ctx-size = 229376 # 7.44 GiB #ctx-size = 262144 # 8.50 GiB # Model VRAM # Qwen3.8-27B-Uncensored-Q6_K.gguf 20.9 GiB # mmproj-Qwen3.8-27B-Uncensored-f16.gguf 0.89 GiB # https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-GGUF m = /home/hans/ai/models/llm/Qwen3.8-27B-Uncensored-Q6_K.gguf ; [Instruct] ; ctx-size = 262144 ; temperature = 0.7 ; top-p = 0.8 ; top-k = 20 ; min-p = 0.0 ; presence-penalty = 1.5 ; repeat-penalty = 1.0 ; reasoning = off ; chat-template-kwargs = {"enable_thinking": false} # [Instruct Vision CPU] # ctx-size = 98304 # mmproj = /home/hans/ai/models/llm_aux/mmproj-Qwen3.8-27B-Uncensored-f16.gguf # # saves VRAM; slower image/PDF/video analysis # no-mmproj-offload = on # temperature = 0.7 # top-p = 0.8 # top-k = 20 # min-p = 0.0 # presence-penalty = 1.5 # repeat-penalty = 1.0 # reasoning = off # chat-template-kwargs = {"enable_thinking": false} # [Instruct Vision GPU] # ctx-size = 98304 # mmproj = /home/hans/ai/models/llm_aux/mmproj-Qwen3.8-27B-Uncensored-f16.gguf # temperature = 0.7 # top-p = 0.8 # top-k = 20 # min-p = 0.0 # presence-penalty = 1.5 # repeat-penalty = 1.0 # reasoning = off # chat-template-kwargs = {"enable_thinking": false} ; [Thinking] ; ctx-size = 98304 ; temperature = 1.0 ; top-p = 0.95 ; top-k = 20 ; min-p = 0.0 ; presence-penalty = 0.0 ; repeat-penalty = 1.0 ; reasoning-preserve = on ; reasoning-effort = medium ; chat-template-kwargs = {"reasoning_effort": "medium"} # [Thinking Vision CPU] # ctx-size = 98304 # mmproj = /home/hans/ai/models/llm_aux/mmproj-Qwen3.8-27B-Uncensored-f16.gguf # # saves VRAM; slower image/PDF/video analysis # no-mmproj-offload = on # temperature = 1.0 # top-p = 0.95 # top-k = 20 # min-p = 0.0 # presence-penalty = 0.0 # repeat-penalty = 1.0 # reasoning-preserve = on # reasoning-effort = medium # chat-template-kwargs = {"reasoning_effort": "medium"} [Thinking Vision GPU] ctx-size = 98304 mmproj = /home/hans/ai/models/llm_aux/mmproj-Qwen3.8-27B-Uncensored-f16.gguf temperature = 1.0 top-p = 0.95 top-k = 20 min-p = 0.0 presence-penalty = 0.0 repeat-penalty = 1.0 reasoning-preserve = on reasoning-effort = xhigh chat-template-kwargs = {"reasoning_effort": "xhigh"}