Merge pull request #414 from one-some/submit-ctx-menu

Submit context menu

modeling/inference_models/hf_torch.py

@@ -34,6 +34,7 @@ from modeling.stoppers import Stoppers
 from modeling.post_token_hooks import PostTokenHooks
 from modeling.inference_models.hf import HFInferenceModel
 from modeling.inference_model import (
+    GenerationMode,
     GenerationResult,
     GenerationSettings,
     ModelCapabilities,
@@ -253,7 +254,10 @@ class HFTorchInferenceModel(HFInferenceModel):
             assert kwargs.pop("logits_warper", None) is not None
             kwargs["logits_warper"] = KoboldLogitsWarperList()
 
-            if utils.koboldai_vars.newlinemode in ["s", "ns"]:
+            if (
+                utils.koboldai_vars.newlinemode in ["s", "ns"]
+                and not m_self.gen_state["allow_eos"]
+            ):
                 kwargs["eos_token_id"] = -1
                 kwargs.setdefault("pad_token_id", 2)
             return new_sample.old_sample(self, *args, **kwargs)
@@ -604,3 +608,9 @@ class HFTorchInferenceModel(HFInferenceModel):
             self.breakmodel = False
             self.usegpu = False
             return
+
+    def get_supported_gen_modes(self) -> List[GenerationMode]:
+        # This changes a torch patch to disallow eos as a bad word.
+        return super().get_supported_gen_modes() + [
+            GenerationMode.UNTIL_EOS
+        ]