97 lines
2.4 KiB
JSON
97 lines
2.4 KiB
JSON
{
|
|
"Nemotron Cascade Q8 160k": {
|
|
"host": "0.0.0.0",
|
|
"port": 3080,
|
|
"parallel": 1,
|
|
"threads": 99,
|
|
"flashAttention": true,
|
|
"kvUnified": true,
|
|
"cacheTypeK": "q8_0",
|
|
"cacheTypeV": "q8_0",
|
|
"temperature": 0.6,
|
|
"topP": 0.95,
|
|
"topK": 20,
|
|
"minP": 0.0,
|
|
"ctxSize": 160000,
|
|
"enableThinking": false,
|
|
"modelPath": "/home/kamma/models/Nemotron-Cascade-2-30B-A3B.Q8_0.gguf",
|
|
"chatTemplateKwargs": "{\"enable_thinking\": false}",
|
|
"ngl": -1
|
|
},
|
|
"Qwen3.5-q6k-180k": {
|
|
"host": "0.0.0.0",
|
|
"port": 3080,
|
|
"parallel": 1,
|
|
"threads": 99,
|
|
"flashAttention": true,
|
|
"kvUnified": true,
|
|
"cacheTypeK": "q8_0",
|
|
"cacheTypeV": "q8_0",
|
|
"temperature": 0.6,
|
|
"topP": 0.95,
|
|
"topK": 20,
|
|
"minP": 0.0,
|
|
"ctxSize": 180000,
|
|
"enableThinking": false,
|
|
"modelPath": "/home/kamma/models/Qwen3.5-35B-A3B-Q6_K.gguf",
|
|
"chatTemplateKwargs": "{\"enable_thinking\": false}",
|
|
"ngl": 999
|
|
},
|
|
"QwenCoderNext-160k": {
|
|
"host": "0.0.0.0",
|
|
"port": 3080,
|
|
"parallel": 1,
|
|
"threads": 99,
|
|
"flashAttention": true,
|
|
"kvUnified": true,
|
|
"cacheTypeK": "q8_0",
|
|
"cacheTypeV": "q8_0",
|
|
"temperature": 0.6,
|
|
"topP": 0.95,
|
|
"topK": 20,
|
|
"minP": 0.0,
|
|
"ctxSize": 160000,
|
|
"enableThinking": false,
|
|
"modelPath": "/home/kamma/models/Qwen3-Coder-Next-UD-Q2_K_XL.gguf",
|
|
"chatTemplateKwargs": "{\"enable_thinking\": false}",
|
|
"ngl": 999
|
|
},
|
|
"Nemotron Cascade 180k": {
|
|
"host": "0.0.0.0",
|
|
"port": 3080,
|
|
"parallel": 1,
|
|
"threads": 99,
|
|
"flashAttention": true,
|
|
"kvUnified": true,
|
|
"cacheTypeK": "q8_0",
|
|
"cacheTypeV": "q8_0",
|
|
"temperature": 0.6,
|
|
"topP": 0.95,
|
|
"topK": 20,
|
|
"minP": 0.0,
|
|
"ctxSize": 180000,
|
|
"enableThinking": false,
|
|
"modelPath": "/home/kamma/models/Nemotron-Cascade-2-30B-A3B.Q5_K_M.gguf",
|
|
"chatTemplateKwargs": "{\"enable_thinking\": false}",
|
|
"ngl": 999
|
|
},
|
|
"Qwen3.5 q6xl 160k": {
|
|
"host": "0.0.0.0",
|
|
"port": 3080,
|
|
"parallel": 1,
|
|
"threads": 99,
|
|
"flashAttention": true,
|
|
"kvUnified": true,
|
|
"cacheTypeK": "q8_0",
|
|
"cacheTypeV": "q8_0",
|
|
"temperature": 0.6,
|
|
"topP": 0.95,
|
|
"topK": 20,
|
|
"minP": 0.0,
|
|
"ctxSize": 160000,
|
|
"enableThinking": false,
|
|
"modelPath": "/home/kamma/models/Qwen3.5-35B-A3B-UD-Q6_K_XL.gguf",
|
|
"chatTemplateKwargs": "{\"enable_thinking\": false}",
|
|
"ngl": -1
|
|
}
|
|
} |