diff --git a/src/main/java/cz/kamma/llamarunner/Main.java b/src/main/java/cz/kamma/llamarunner/Main.java index c61f452..c8bc7ad 100644 --- a/src/main/java/cz/kamma/llamarunner/Main.java +++ b/src/main/java/cz/kamma/llamarunner/Main.java @@ -814,8 +814,8 @@ public class Main extends JFrame { gbc.gridx = 1; gbc.weightx = 1.0; - cacheTypeKComboBox = new JComboBox<>(new String[] { "bf16", "f16", "f32", "f8", "q8_0", "turbo3", "turbo4" }); - cacheTypeKComboBox.setSelectedIndex(0); + cacheTypeKComboBox = new JComboBox<>(new String[] { "bf16", "f16", "f32", "f8", "q8_0", "q5_0", "turbo3", "turbo4" }); + cacheTypeKComboBox.setSelectedIndex(4); increaseComboBoxHeight(cacheTypeKComboBox); cacheTypeKComboBox.addActionListener(e -> updateCommandPreview()); panel.add(cacheTypeKComboBox, gbc); @@ -827,8 +827,8 @@ public class Main extends JFrame { gbc.gridx = 1; gbc.weightx = 1.0; - cacheTypeVComboBox = new JComboBox<>(new String[] { "bf16", "f16", "f32", "f8", "q8_0", "turbo3", "turbo4" }); - cacheTypeVComboBox.setSelectedIndex(0); + cacheTypeVComboBox = new JComboBox<>(new String[] { "bf16", "f16", "f32", "f8", "q8_0", "q5_0", "turbo3", "turbo4" }); + cacheTypeVComboBox.setSelectedIndex(4); increaseComboBoxHeight(cacheTypeVComboBox); cacheTypeVComboBox.addActionListener(e -> updateCommandPreview()); panel.add(cacheTypeVComboBox, gbc); diff --git a/src/main/java/cz/kamma/llamarunner/ModelConfig.java b/src/main/java/cz/kamma/llamarunner/ModelConfig.java index f2d613c..f74b974 100644 --- a/src/main/java/cz/kamma/llamarunner/ModelConfig.java +++ b/src/main/java/cz/kamma/llamarunner/ModelConfig.java @@ -39,8 +39,8 @@ public class ModelConfig implements Serializable { this.threads = 99; this.flashAttention = true; this.kvUnified = true; - this.cacheTypeK = "bf16"; - this.cacheTypeV = "bf16"; + this.cacheTypeK = "q8_0"; + this.cacheTypeV = "q8_0"; this.temperature = 0.6; this.topP = 0.95; this.topK = 20;