diff --git a/config.properties b/config.properties index 7f623ff..c199559 100644 --- a/config.properties +++ b/config.properties @@ -1,5 +1,5 @@ #Llama Runner Configuration -#Sun Mar 29 17:54:05 CEST 2026 +#Thu Apr 02 10:20:00 CEST 2026 windowHeight=1189 windowWidth=711 windowX=1849 diff --git a/profiles.json b/profiles.json index 95dc982..c6d8e87 100644 --- a/profiles.json +++ b/profiles.json @@ -3,6 +3,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -24,6 +25,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -45,6 +47,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -66,6 +69,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -87,6 +91,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -108,6 +113,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -129,6 +135,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -150,6 +157,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -171,6 +179,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -192,6 +201,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -213,6 +223,7 @@ "host": "0.0.0.0", "port": 3080, "parallel": 1, + "batchSize": 2048, "threads": 99, "flashAttention": true, "kvUnified": true, @@ -229,5 +240,27 @@ "ngl": -1, "fit": true, "reasoning": true + }, + "Qwen35-35B-A3B-Q6K-turbo4-256k": { + "host": "0.0.0.0", + "port": 3080, + "parallel": 1, + "batchSize": 2048, + "threads": 99, + "flashAttention": true, + "kvUnified": true, + "cacheTypeK": "turbo4", + "cacheTypeV": "turbo4", + "temperature": 0.6, + "topP": 0.95, + "topK": 20, + "minP": 0.0, + "ctxSize": 256000, + "enableThinking": false, + "modelPath": "/home/kamma/models/Qwen3.5-35B-A3B-Q6_K.gguf", + "chatTemplateKwargs": "{\"enable_thinking\": true}", + "ngl": 999, + "fit": true, + "reasoning": true } } \ No newline at end of file