llama.cpp-utils/models.ini at main · ronoaldo/llama.cpp-utils · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
; =================================================================
; Global Defaults (100k context, q8_0 KV cache, desktop safe)
; =================================================================
[*]
ctx-size = 102400
ubatch-size = 1024
batch-size = 1024
cache-type-k = q8_0
cache-type-v = q8_0
reasoning-budget = 2048
n-predict = 8192

; =================================================================
; Google Models
; - V cache mantido em f16 (Gemma e sensivel a quantizacao do V cache)
; - Perfis performance usam ngram-mod (GGUF oficial nao possui MTP)
; =================================================================

; --- Gemma 4 26B A4B (MoE) ---
; Default: 16 experts na CPU (n-cpu-moe) para caber em 16GB VRAM com folga desktop
[google/gemma-4-26B-A4B]
hf = google/gemma-4-26B-A4B-it-qat-q4_0-gguf
alias = google-gemma-4-26b-a4b
ubatch-size = 256
batch-size = 256
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
n-gpu-layers = 99
n-cpu-moe = 16
threads = 8

; Performance: auto-fit com margem de 2GB de VRAM e aceleracao n-gram
[google/gemma-4-26B-A4B-performance]
hf = google/gemma-4-26B-A4B-it-qat-q4_0-gguf
alias = google-gemma-4-26b-a4b-perf
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
fit-target = 2048
threads = 12
spec-type = ngram-mod
spec-draft-n-max = 3

; --- Gemma 4 12B (Dense) ---
; Default: balanceado para uso geral com cores livres para desktop
[google/gemma-4-12B]
hf = google/gemma-4-12B-it-qat-q4_0-gguf
alias = google-gemma-4-12b
ubatch-size = 256
batch-size = 256
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
n-gpu-layers = 99
threads = 8

; Performance: maximo offload com margem de 2GB de VRAM e aceleracao n-gram
[google/gemma-4-12B-performance]
hf = google/gemma-4-12B-it-qat-q4_0-gguf
alias = google-gemma-4-12b-perf
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
fit-target = 2048
threads = 12
spec-type = ngram-mod
spec-draft-n-max = 3

; --- Gemma 4 E4B ---
; Default: modelo leve balanceado com todas camadas na GPU
[google/gemma-4-E4B]
hf = google/gemma-4-E4B-it-qat-q4_0-gguf
alias = google-gemma-4-e4b
ubatch-size = 256
batch-size = 256
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
n-gpu-layers = 99
threads = 8

; Performance: maxima utilizacao de CPU/GPU com aceleracao n-gram
[google/gemma-4-E4B-performance]
hf = google/gemma-4-E4B-it-qat-q4_0-gguf
alias = google-gemma-4-e4b-perf
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
fit-target = 2048
threads = 12
spec-type = ngram-mod
spec-draft-n-max = 3

; --- Gemma 4 E2B ---
; Default: modelo compacto para tarefas rapidas e baixo consumo
[google/gemma-4-E2B]
hf = google/gemma-4-E2B-it-qat-q4_0-gguf
alias = google-gemma-4-e2b
ubatch-size = 256
batch-size = 256
n-gpu-layers = 99
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
threads = 8

; Performance: maxima utilizacao com margem de 2GB de VRAM e aceleracao n-gram
[google/gemma-4-E2B-performance]
hf = google/gemma-4-E2B-it-qat-q4_0-gguf
alias = google-gemma-4-e2b-perf
fit-target = 2048
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
threads = 12
spec-type = ngram-mod
spec-draft-n-max = 3

; =================================================================
; Unsloth Models
; - Perfis performance usam draft-mtp (GGUF Unsloth inclui cabecote MTP)
; =================================================================

; --- Gemma 4 26B A4B (Unsloth MoE) ---
; Default: 16 experts na CPU (n-cpu-moe) para caber em 16GB VRAM com folga desktop
[unsloth/gemma-4-26B]
hf = unsloth/gemma-4-26B-A4B-it-qat-GGUF:UD-Q4_K_XL
alias = unsloth-gemma-4-26b
ubatch-size = 256
batch-size = 256
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
n-gpu-layers = 99
n-cpu-moe = 16
threads = 8

; Performance: auto-fit na VRAM e decodificacao especulativa via cabecote MTP nativo
[unsloth/gemma-4-26B-performance]
hf = unsloth/gemma-4-26B-A4B-it-qat-GGUF:UD-Q4_K_XL
alias = unsloth-gemma-4-26b-perf
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
fit-target = 2048
threads = 12
spec-type = draft-mtp
spec-draft-n-max = 2

; --- Gemma 4 12B (Unsloth Dense) ---
; Default: balanceado para uso geral com V cache f16
[unsloth/gemma-4-12B]
hf = unsloth/gemma-4-12B-it-qat-GGUF:UD-Q4_K_XL
alias = unsloth-gemma-4-12b
ubatch-size = 256
batch-size = 256
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
n-gpu-layers = 99
threads = 8

; Performance: auto-fit na VRAM e decodificacao especulativa via MTP nativo
[unsloth/gemma-4-12B-performance]
hf = unsloth/gemma-4-12B-it-qat-GGUF:UD-Q4_K_XL
alias = unsloth-gemma-4-12b-perf
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
fit-target = 2048
threads = 12
spec-type = draft-mtp
spec-draft-n-max = 2

; --- Gemma 4 E2B (Unsloth) ---
; Default: versao ultraleve para tarefas rapidas
[unsloth/gemma-4-E2B]
hf = unsloth/gemma-4-E2B-it-qat-GGUF:UD-Q4_K_XL
alias = unsloth-gemma-4-e2b
ubatch-size = 256
batch-size = 256
n-gpu-layers = 99
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
threads = 8

; Performance: auto-fit na VRAM e decodificacao especulativa via MTP nativo
[unsloth/gemma-4-E2B-performance]
hf = unsloth/gemma-4-E2B-it-qat-GGUF:UD-Q4_K_XL
alias = unsloth-gemma-4-e2b-perf
fit-target = 2048
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 64
threads = 12
spec-type = draft-mtp
spec-draft-n-max = 2

; --- Qwen 3.8 27B (Dense) ---
; Q3 com KV cache q4_0 e 64k ctx para caber em 16GB VRAM; MTP nativo para ~45 tok/s
[unsloth/qwen-3.8-27B]
hf = unsloth/Qwen3.8-27B-GGUF:UD-Q3_K_XL
alias = qwen-3.8-27b
ctx-size = 65536
ubatch-size = 256
batch-size = 256
flash-attn = on
temp = 1.0
top-p = 0.95
top-k = 20
min-p = 0.00
presence-penalty = 0
n-gpu-layers = 99
threads = 12
cache-type-k = q4_0
cache-type-v = q4_0
spec-type = draft-mtp
spec-draft-n-max = 2

; --- Qwen 3.6 35B (MoE) ---
; Default: 32 experts na CPU (n-cpu-moe) para manter desktop fluido
[unsloth/qwen-3.6-35B]
hf = unsloth/Qwen3.6-35B-A3B-MTP-GGUF:UD-Q4_K_XL
alias = qwen-3.6-35b
ubatch-size = 256
batch-size = 256
flash-attn = on
temp = 0.6
top-p = 0.95
top-k = 20
min-p = 0.00
n-gpu-layers = 99
n-cpu-moe = 32
threads = 8

; Performance: 24 experts na CPU (mais VRAM usada) com MTP nativo e 12 threads
[unsloth/qwen-3.6-35B-performance]
hf = unsloth/Qwen3.6-35B-A3B-MTP-GGUF:UD-Q4_K_XL
alias = qwen-3.6-35b-perf
flash-attn = on
temp = 0.6
top-p = 0.95
top-k = 20
min-p = 0.00
n-gpu-layers = 99
n-cpu-moe = 24
threads = 12
spec-type = draft-mtp
spec-draft-n-max = 2