adaptive decay as an overridable param (+1 squashed commits)

Squashed commits:

[d94df7843] adaptive decay as an overridable param

expose.h

@@ -124,6 +124,7 @@ struct generation_inputs
     const float smoothing_factor = 0.0f;
     const float smoothing_curve = 1.0f;
     const float adaptive_target = -1.0f;
+    const float adaptive_decay = 0.9f;
     const float dry_multiplier = 0.0f;
     const float dry_base = 0.0f;
     const int dry_allowed_length = 0;

gpttype_adapter.cpp

@@ -1303,16 +1303,13 @@ llama_token_data_array * cur_p)
 
     //update EMA history AFTER sampling, update_adaptive_p_history(original_prob[idx])
 }
-inline void adaptive_p_update_history(float selected_token_prob, float & weighted_sum, float & total_weight) {
+inline void adaptive_p_update_history(float selected_token_prob, float & weighted_sum, float & total_weight, float adaptive_decay) {
     // decay controls how quickly history influence fades (0.0 to 0.99)
     // lower values = faster adaptation, more reactive to recent tokens
     // higher values = slower adaptation, more stable over time
-    // effective history length ≈ 1/(1-decay) tokens
-    // example: decay=0.5 --> ~2 tokens; decay=0.9 --> ~10 tokens; decay=0.95 --> ~20 tokens
     // keep <= 0.99 to prevent unbounded accumulation
-    const float adaptive_p_decay = 0.90f;
-    weighted_sum = selected_token_prob + adaptive_p_decay * weighted_sum;
-    total_weight = 1.0f + adaptive_p_decay * total_weight;
+    weighted_sum = selected_token_prob + adaptive_decay * weighted_sum;
+    total_weight = 1.0f + adaptive_decay * total_weight;
 }
 
 
@@ -3656,6 +3653,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
     kcpp_data->smoothing_factor = inputs.smoothing_factor;
     kcpp_data->smoothing_curve = inputs.smoothing_curve;
     kcpp_data->adaptive_target = inputs.adaptive_target;
+    kcpp_data->adaptive_decay = inputs.adaptive_decay;
 
     // Parse dry sequence breakers / restart sequences
     kcpp_data->dry_sequence_breakers.clear();
@@ -4485,6 +4483,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
             const float smoothing_factor = kcpp_data->smoothing_factor;
             const float smoothing_curve = kcpp_data->smoothing_curve;
             const float adaptive_target = kcpp_data->adaptive_target;
+            const float adaptive_decay = kcpp_data->adaptive_decay;
 
             if (!startedsampling)
             {
@@ -4622,7 +4621,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
 
                 if (adaptive_target > 0.0f) {
                     float original_prob = original_candidates[id].p;
-                    adaptive_p_update_history(original_prob, adaptive_p_weighted_sum, adaptive_p_total_weight);
+                    adaptive_p_update_history(original_prob, adaptive_p_weighted_sum, adaptive_p_total_weight, adaptive_decay);
                 }
 
                 if(draft_used)

koboldcpp.py

@@ -266,6 +266,7 @@ class generation_inputs(ctypes.Structure):
                 ("smoothing_factor", ctypes.c_float),
                 ("smoothing_curve", ctypes.c_float),
                 ("adaptive_target", ctypes.c_float),
+                ("adaptive_decay", ctypes.c_float),
                 ("dry_multiplier", ctypes.c_float),
                 ("dry_base", ctypes.c_float),
                 ("dry_allowed_length", ctypes.c_int),
@@ -1605,8 +1606,9 @@ def generate(genparams, stream_flag=False):
     smoothing_factor = tryparsefloat(genparams.get('smoothing_factor', 0.0),0.0)
     smoothing_curve = tryparsefloat(genparams.get('smoothing_curve', 1.0),1.0)
     adaptive_target = tryparsefloat(genparams.get('adaptive_target', -1.0),-1.0)
+    adaptive_decay = tryparsefloat(genparams.get('adaptive_decay', 0.9),0.9)
     if adaptive_target>0 and min_p<=0 and top_p>=1.0: #adaptive p sampler requires a truncation sampler first, force a tiny min-p
-        min_p = 0.01
+        min_p = 0.002
     logit_biases = genparams.get('logit_bias', {})
     render_special = genparams.get('render_special', False)
     banned_strings = genparams.get('banned_strings', []) # SillyTavern uses that name
@@ -1671,6 +1673,7 @@ def generate(genparams, stream_flag=False):
     inputs.smoothing_factor = smoothing_factor
     inputs.smoothing_curve = smoothing_curve
     inputs.adaptive_target = adaptive_target
+    inputs.adaptive_decay = adaptive_decay
     inputs.grammar = grammar.encode("UTF-8")
     inputs.grammar_retain_state = grammar_retain_state
     inputs.allow_eos_token = not ban_eos_token

otherarch/otherarch.h

@@ -51,6 +51,7 @@ struct kcpp_params {
     float   dynatemp_range     = 0.0f;  // enables DynaTemp if neq 0. dynatemp_min = temperature - dt_range, dynatemp_max = temperature + dt_range
     float   dynatemp_exponent  = 1.0f;
     float adaptive_target     = -1.0f; // 0.0 - 1.0, <=0.0 is disabled
+    float adaptive_decay      = 0.9f;
 
     std::string model_filename       = ""; // model path
     std::string prompt               = "";