webscout 2.2b0__py3-none-any.whl → 2.3__py3-none-any.whl

webscout/Local/samplers.py

@@ -0,0 +1,161 @@
+
+from ._version import __version__, __llama_cpp_version__
+
+"""Submodule containing SamplerSettings class and some preset samplers"""
+
+from sys import maxsize
+
+
+MAX_TEMP = float(maxsize)
+
+class SamplerSettings:
+    """
+    A SamplerSettings object specifies the sampling parameters that will be
+    used to control text generation
+    """
+
+    ParamTypes: dict[str, type] = {
+        'max_len_tokens':    int,
+        'temp':              float,
+        'top_p':             float,
+        'min_p':             float,
+        'frequency_penalty': float,
+        'presence_penalty':  float,
+        'repeat_penalty':    float,
+        'top_k':             int
+    }
+
+    def __init__(
+        self,
+        max_len_tokens:    int   = -1,
+        temp:              float = 0.8,
+        top_p:             float = 0.95,
+        min_p:             float = 0.05,
+        frequency_penalty: float = 0.0,
+        presence_penalty:  float = 0.0,
+        repeat_penalty:    float = 1.0,
+        top_k:             int   = 40
+    ):
+        """
+        Construct a new SamplerSettings instance
+        """
+
+        self.max_len_tokens    = max_len_tokens
+        self.temp              = temp
+        self.top_p             = top_p
+        self.min_p             = min_p
+        self.frequency_penalty = frequency_penalty
+        self.presence_penalty  = presence_penalty
+        self.repeat_penalty    = repeat_penalty
+        self.top_k             = top_k
+
+        for sampler_param in SamplerSettings.ParamTypes:
+            expected_type = SamplerSettings.ParamTypes[sampler_param]
+            actual_type = type(getattr(self, sampler_param))
+            if actual_type != expected_type:
+                raise TypeError(
+                    f"wrong type for SamplerSettings parameter '{sampler_param}'"
+                    f" - expected {expected_type}, got {actual_type}"
+                )
+    
+    def __repr__(self) -> str:
+        repr_str = 'SamplerSettings('
+        repr_str += f'max_len_tokens={self.max_len_tokens}, '
+        repr_str += f'temp={self.temp}, '
+        repr_str += f'top_p={self.top_p}, '
+        repr_str += f'min_p={self.min_p}, '
+        repr_str += f'frequency_penalty={self.frequency_penalty}, '
+        repr_str += f'presence_penalty={self.presence_penalty}, '
+        repr_str += f'repeat_penalty={self.repeat_penalty}, '
+        repr_str += f'top_k={self.top_k})'
+        return repr_str
+
+# most likely token is always chosen
+GreedyDecoding = SamplerSettings(
+    temp = 0.0,
+)
+
+# reflects llama.cpp
+DefaultSampling = SamplerSettings()
+
+# unmodified probability distribution (i.e. what the model actually thinks)
+SimpleSampling = SamplerSettings(
+    temp = 1.0,
+    top_p = 1.0,
+    min_p = 0.0,
+    top_k = -1
+)
+
+# reflects old llama.cpp defaults
+ClassicSampling = SamplerSettings(
+    min_p=0.0,
+    repeat_penalty = 1.1
+)
+
+# halfway between DefaultSampling and SimpleSampling
+SemiSampling = SamplerSettings(
+    temp=0.9,
+    top_p=0.975,
+    min_p=0.025,
+    top_k=80
+)
+
+# for models with large vocabulary, which tend to run hot
+TikTokenSampling = SamplerSettings(
+    temp=0.6,
+    repeat_penalty=1.1
+)
+
+# use min_p as the only active sampler (more permissive)
+LowMinPSampling = SamplerSettings(
+    temp = 1.0,
+    top_p = 1.0,
+    min_p = 0.05,
+    top_k = -1
+)
+
+# use min_p as the only active sampler (moderate)
+MinPSampling = SamplerSettings(
+    temp = 1.0,
+    top_p = 1.0,
+    min_p = 0.1,
+    top_k = -1
+)
+
+# use min_p as the only active sampler (more restrictive)
+StrictMinPSampling = SamplerSettings(
+    temp = 1.0,
+    top_p = 1.0,
+    min_p = 0.2,
+    top_k = -1
+)
+
+# https://arxiv.org/abs/2210.14140
+ContrastiveSearch = SamplerSettings(
+    temp = 0.0,
+    presence_penalty = 0.4
+)
+
+# https://arxiv.org/abs/2210.14140
+WarmContrastiveSearch = SamplerSettings(
+    temp = 0.0,
+    presence_penalty = 0.8
+)
+
+# outputs completely random tokens from vocab (useless)
+RandomSampling = SamplerSettings(
+    temp = MAX_TEMP,
+    top_p = 1.0,
+    min_p = 0.0,
+    top_k = -1
+)
+
+# default sampling with reduced temperature
+LowTempSampling = SamplerSettings(
+    temp = 0.4
+)
+
+# default sampling with increased temperature
+HighTempSampling = SamplerSettings(
+    temp = 1.2
+)