update modes for v3-5-f

Author: Originalimoc

exllamav2/exllamav2_ext/ext_quant.cpp

@@ -174,9 +174,9 @@ std::tuple<std::vector<std::tuple<uint64_t, float>>, std::vector<int>, float, ui
 )
 {
     // --- Mode-Specific Parameters ---
-    enum Mode { MODE_BALANCED, MODE_UNIFORM, MODE_AGGRESSIVE, MODE_3_5_2, MODE_3_5_6, MODE_CUSTOM };
+    enum Mode { MODE_RELAXED, MODE_BALANCED, MODE_UNIFORM, MODE_AGGRESSIVE, MODE_3_5_2, MODE_3_5_6, MODE_CUSTOM };
     // --- Mode Selection ---
-    Mode mode = MODE_3_5_2;  // Default mode, Can be changed into other mode or MODE_CUSTOM
+    Mode mode = MODE_UNIFORM;  // Default mode, Can be changed into other mode or MODE_CUSTOM
 
     // Define a struct to hold parameters for different modes
     struct ModeParams {
@@ -198,23 +198,20 @@ std::tuple<std::vector<std::tuple<uint64_t, float>>, std::vector<int>, float, ui
 
     // Define the parameter sets for each mode
     const std::vector<ModeParams> mode_params = {
+        // MODE_RELAXED: Minize error first
+        {0.1f, 3.0f, 0.05f, 0.0f, 1.2f, 5.0f, 0.1f, 0.0009f, 25, 15000, 3, 5, 0.5f, 3.3f},
+
         // MODE_BALANCED: Balanced trade-off between BPW uniformity and error
-        {0.6f, 3.2f, 0.1f, 0.0001f, 1.3f, 5.5f, 1.5f, 0.001f, 60, 30000, 8, 8, 0.75f, 3.5f},
+        {0.6f, 3.3f, 0.12f, 0.0005f, 1.5f, 5.5f, 1.8f, 0.002f, 50, 30000, 8, 8, 0.75f, 3.6f},
 
         // MODE_UNIFORM: Strong emphasis on BPW uniformity
-        {0.8f, 3.5f, 0.12f, 0.0005f, 1.5f, 6.0f, 3.0f, 0.001f, 80, 40000, 8, 10, 0.8f, 3.7f},
+        {0.8f, 3.5f, 0.12f, 0.0005f, 1.6f, 6.0f, 3.0f, 0.001f, 80, 40000, 8, 10, 0.8f, 3.7f},
 
         // MODE_AGGRESSIVE: Aggressively avoids low BPW, potentially higher error
-        {1.0f, 3.8f, 0.15f, 0.001f, 1.6f, 6.5f, 4.0f, 0.001f, 100, 50000, 8, 12, 0.9f, 3.9f},
-
-        // MODE_3_5_2: Approximates the behavior of Version 3-5-2
-        {0.1f, 3.0f, 0.05f, 0.0f, 1.2f, 5.0f, 0.1f, 0.0009f, 25, 15000, 3, 5, 0.5f, 3.3f},
-
-        // MODE_3_5_6: Replicates the behavior of Version 3-5-6
-        {0.6f, 3.3f, 0.12f, 0.0005f, 1.5f, 5.5f, 1.8f, 0.002f, 50, 30000, 8, 8, 0.75f, 3.6f},
+        {1.0f, 3.8f, 0.15f, 0.001f, 1.7f, 6.5f, 4.0f, 0.001f, 100, 50000, 8, 12, 0.9f, 3.9f},
 
         // MODE_CUSTOM: User-defined parameters, will be overwritten if using custom mode
-        {0.6f, 3.2f, 0.1f, 0.0001f, 1.3f, 5.5f, 1.5f, 0.001f, 60, 30000, 8, 8, 0.75f, 3.5f} 
+        {0.8f, 5.0f, 0.12f, 0.0005f, 1.5f, 6.0f, 3.0f, 0.001f, 80, 40000, 8, 10, 0.8f, 5.5f},
     };
 
     ModeParams params;