Make logit bias work correctly(?) when prob is -inf

samplers'll do that to you

though now i am curious: what kind of effect would running the bias
before the samplers have? maybe a future option

modeling/inference_models/hf_torch.py

@@ -225,9 +225,6 @@ class HFTorchInferenceModel(HFInferenceModel):
         )
 
         class KoboldLogitsWarperList(LogitsProcessorList):
-            def __init__(self):
-                pass
-
             def __call__(
                 lw_self,
                 input_ids: torch.LongTensor,
@@ -244,16 +241,10 @@ class HFTorchInferenceModel(HFInferenceModel):
                     ), f"Scores are None; processor '{processor}' is to blame"
                 return scores
 
-        def new_get_logits_warper(
-            beams: int = 1,
-        ) -> LogitsProcessorList:
-            return KoboldLogitsWarperList()
-
         def new_sample(self, *args, **kwargs):
             assert kwargs.pop("logits_warper", None) is not None
-            kwargs["logits_warper"] = new_get_logits_warper(
+            kwargs["logits_warper"] = KoboldLogitsWarperList()
-                beams=1,
+
-            )
             if utils.koboldai_vars.newlinemode in ["s", "ns"]:
                 kwargs["eos_token_id"] = -1
                 kwargs.setdefault("pad_token_id", 2)

modeling/logits_processors.py

@@ -267,6 +267,14 @@ class PhraseBiasLogitsProcessor:
 
         for batch in range(scores_shape[0]):
             for token, bias in self._get_biased_tokens(input_ids[batch]).items():
-                scores[batch][token] += bias
+                if bias > 0 and bool(scores[batch][token].isneginf()):
+                    # Adding bias to -inf will do NOTHING!!! So just set it for
+                    # now. There may be more mathishly correct way to do this
+                    # but it'll work. Also, make sure the bias is actually
+                    # positive. Don't give a -inf token more chance by setting
+                    # it to -0.5!
+                    scores[batch][token] = bias
+                else:
+                    scores[batch][token] += bias
 
         return scores