Sampling: Add adaptive-P params

backends/exllamav3/model.py

@@ -918,6 +918,10 @@ class ExllamaV3Container(BaseModelContainer):
         if params.temperature_last:
             sampler_builder.temperature(params.temperature)
 
+        # Apply adaptive-P
+        if params.adaptive_target < 1.0:
+            sampler_builder.adaptive_p(params.adaptive_target, params.adaptive_decay)
+
         # Build the sampler
         # Set greedy if temperature is 0
         sampler = sampler_builder.build(params.temperature == 0)

backends/exllamav3/sampler.py

@@ -11,6 +11,7 @@ from exllamav3.generator.sampler import (
     SS_TopP,
     SS_Sample,
     SS_Base,
+    SS_AdaptiveP
 )
 
 
@@ -43,9 +44,16 @@ class ExllamaV3SamplerBuilder:
     def greedy(self):
         self.stack.append(SS_Argmax())
 
+    def adaptive_p(self, adaptive_target, adaptive_decay):
+        self.stack.append(SS_AdaptiveP(adaptive_target, adaptive_decay))
+
     def build(self, greedy):
         """Builds the final sampler from stack."""
 
+        # Adaptive-P does categorical sampling already
+        if len(self.stack) and isinstance(self.stack[-1], SS_AdaptiveP):
+            return CustomSampler(self.stack)
+
         # Use greedy if temp is 0
         if greedy:
             return CustomSampler([SS_Argmax()])

common/sampling.py

@@ -275,6 +275,15 @@ class BaseSamplerRequest(BaseModel):
         ge=0,
     )
 
+    adaptive_target: Optional[float] = Field(
+        default_factory=lambda: get_default_sampler_value("adaptive_target", 1.0)
+    )
+
+    adaptive_decay: Optional[float] = Field(
+        default_factory=lambda: get_default_sampler_value("adaptive_decay", 0.9)
+    )
+
+
     @field_validator("top_k", mode="before")
     def convert_top_k(cls, v):
         """Fixes instance if Top-K is -1."""

sampler_overrides/sample_preset.yml

@@ -156,3 +156,11 @@ cfg_scale:
 negative_prompt:
   override:
   force: false
+
+# MARK: Adaptive-P
+adaptive_target:
+  override: 1.0
+  force: false
+adaptive_decay:
+  override: 0.9
+  force: false