Samplers: Expose skew sampling

Skew is an extra unused sampler in ExllamaV2. Add it in for coverage.

Signed-off-by: kingbri <bdashore3@proton.me>

backends/exllamav2/model.py

@@ -709,6 +709,7 @@ class ExllamaV2Container:
         gen_settings.tfs = unwrap(kwargs.get("tfs"), 1.0)
         gen_settings.typical = unwrap(kwargs.get("typical"), 1.0)
         gen_settings.mirostat = unwrap(kwargs.get("mirostat"), False)
+        gen_settings.skew = unwrap(kwargs.get("skew"), 0)
 
         # DynaTemp settings
         max_temp = unwrap(kwargs.get("max_temp"), 1.0)

common/sampling.py

@@ -75,6 +75,11 @@ class BaseSamplerRequest(BaseModel):
         examples=[1.0],
     )
 
+    skew: Optional[float] = Field(
+        default_factory=lambda: get_default_sampler_value("skew", 0.0),
+        examples=[0.0],
+    )
+
     frequency_penalty: Optional[float] = Field(
         default_factory=lambda: get_default_sampler_value("frequency_penalty", 0.0)
     )
@@ -295,6 +300,7 @@ class BaseSamplerRequest(BaseModel):
             "typical": self.typical,
             "min_p": self.min_p,
             "tfs": self.tfs,
+            "skew": self.skew,
             "frequency_penalty": self.frequency_penalty,
             "presence_penalty": self.presence_penalty,
             "repetition_penalty": self.repetition_penalty,

sampler_overrides/sample_preset.yml

@@ -73,6 +73,9 @@ tfs:
 typical:
   override: 1.0
   force: false
+skew:
+  override: 0.0
+  force: false
 
 # MARK: Penalty settings
 frequency_penalty: